<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Eric Green on Medium]]></title>
        <description><![CDATA[Stories by Eric Green on Medium]]></description>
        <link>https://medium.com/@ericpgreen?source=rss-c3dc9bdc0078------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*TNiIPBNfNh2ReMp1ECMYWA.png</url>
            <title>Stories by Eric Green on Medium</title>
            <link>https://medium.com/@ericpgreen?source=rss-c3dc9bdc0078------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 21:23:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ericpgreen/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What do users want?]]></title>
            <link>https://medium.com/nivi-inc/what-do-users-want-35cc9fec12a7?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/35cc9fec12a7</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[womens-health]]></category>
            <category><![CDATA[sexual-health]]></category>
            <category><![CDATA[family-planning]]></category>
            <category><![CDATA[chatbots]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Mon, 12 Nov 2018 08:52:19 GMT</pubDate>
            <atom:updated>2018-11-12T13:01:07.801Z</atom:updated>
            <content:encoded><![CDATA[<h4>(hint: more than we expected)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*77O23dGx63vTfNZIzWPqZw.jpeg" /></figure><p>In our health chatbot <a href="https://www.nivi.io/">askNivi</a>, the first step in the user journey is <strong>intent classification</strong>—what does she want to know? In a <a href="https://towardsdatascience.com/building-your-own-mturk-style-app-in-r-using-shiny-flexdashboard-3beacba50e09">previous post</a>, we discussed our approach to answering this question.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CEqFAhMr32jUPdgodqHomw.png" /></figure><p>At the time we wrote that piece, the challenge was that our algorithm did not know how to handle a substantial chunk of SMS messages. Word was spreading about askNivi, and we started getting a wider range of questions that were tough to classify with the existing model. For instance, we had not seen questions like:</p><blockquote>What is the probability of getting cervical cancer when you have sex at around maybe 18 to 20 years?</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*2LvwZAdmQR6Pb5M4qepCWg.png" /></figure><p>We needed our crack team of humans to go back and label a growing pile of unclassified messages to further train the system, so we built an <strong>MTurk-style app with Shiny and Flexdashboard</strong> (see <a href="https://towardsdatascience.com/building-your-own-mturk-style-app-in-r-using-shiny-flexdashboard-3beacba50e09">here</a> for a tutorial with code).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c9BR71YQR94L8-xlGUUY5g.png" /></figure><p>After classifying the messages, we had a better picture of the full distribution of topics. In addition to the questions about contraception we expected based on our marketing (mint circles), we also fielded questions about topics we did not advertise much if at all (rose circles), such as sexual health and sexually transmitted diseases. This chart is guiding how we prioritize the development of new automated conversation modules.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*bQ-i0u-EFlE_Ds-YVlVqtA.png" /></figure><p>While we got our start catering to the health needs of adolescent girls and young women, we quickly found a high demand among men and responded by expanding our service to meet their needs as well. So what do men want to know?</p><p>Compared to our female users, men are more interested in questions related to sexual health, relationships, and STIs. But overall, contraception is still the dominant theme, even among men. We chat with loads of men who want to support their partners in family planning:</p><blockquote>My girl fears all the other contraceptives…she prefers taking P2. What are the risk?</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*T2HlKlGt_KHlU1bjnny1Pw.png" /></figure><p>Listen, analyze, refine. That’s our process.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=35cc9fec12a7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nivi-inc/what-do-users-want-35cc9fec12a7">What do users want?</a> was originally published in <a href="https://medium.com/nivi-inc">Nivi</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building your own MTurk-style app in R using Shiny+Flexdashboard]]></title>
            <link>https://medium.com/nivi-inc/building-your-own-mturk-style-app-in-r-using-shiny-flexdashboard-3beacba50e09?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/3beacba50e09</guid>
            <category><![CDATA[shiny]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[mturk]]></category>
            <category><![CDATA[classification]]></category>
            <category><![CDATA[rstats]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Wed, 26 Sep 2018 00:17:15 GMT</pubDate>
            <atom:updated>2018-09-26T00:54:59.762Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PlPGjdztkdZe8EZ1RF0BdQ.png" /></figure><h3>Conversational Health</h3><p>At <a href="https://www.nivi.io/">Nivi</a>, we’re hard at work building a great chatbot for health. We started our journey with a simple interactive voice response service in Kenya that helped users identify and obtain methods of family planning that fit their goals. It worked great, but it was a pretty structured process. We ask you some questions, you press a key to respond to these automated audio prompts, and we text you a referral. Perfect if this specific referral is all you need, but follow-up conversations with users suggested that we were missing a bigger opportunity to provide some value.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dh-lVBCIVQEqr5lOprkG-A.jpeg" /></figure><p>We thought it would be interesting to open up communication and invite users to send free form text messages. We would then try to determine what they want and attempt to meet their needs. And like that, <strong>askNivi</strong> was born. Users send SMS messages to our toll-free shortcode, and we pipe these texts to an internal app that lets our customer success team read and respond. It was a success! Messages started pouring in. This presented a new challenge: <em>responding to all of these messages</em>.</p><h3>What Are You Trying to Say?</h3><p>To keep pace with the demand and offer users something truly special, we set out to turn askNivi into a platform for automated conversations about health. The first challenge for any bot maker is to figure out what the heck a user wants to do or know. We turned to the <a href="https://wit.ai/">Wit.ai</a> natural language engine for help classifying incoming requests.</p><p>To get started, we simply fed a few hundred messages to Wit’s web app and manually tagged each message with an <a href="https://medium.com/@brijrajsingh/chat-bots-designing-intents-and-entities-for-your-nlp-models-35c385b7730d">intent label</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yBalLHXOi_FTZ4-mY-NORA.png" /><figcaption>Screenshot from Wit.ai.</figcaption></figure><p>In the next iteration, we integrated Wit (<a href="https://wit.ai/docs/http/20170307">API docs</a>) into our chat app workflow. We started passing incoming requests to Wit, and Wit classified the message with one of our intent labels (or not, in the case of recall errors). Back in our chat app, a member of our customer success team would validate the suggestion and continue the conversation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lEfs8wceZ8CMhHd7A_IVnw.png" /><figcaption>askNivi chat app in action.</figcaption></figure><h3>Managing Human Intelligence Tasks</h3><p>This workflow has served us well, but we quickly developed a backlog of messages that did not fall into one of our existing intents and therefore went unclassified. It was easy enough to identify new intent labels to close the gap, but we then had a “human intelligence task” on our hands, to use the language of Amazon’s <a href="https://www.mturk.com/">Mechanical Turk</a> platform: we needed people to read several thousand bite-sized messages in our unclassified pile and match them to the newly updated list of intents.</p><p>We considered a range of options, starting with MTurk. At first it seemed like an ideal solution: (Step 1) Create a project with discrete HITs; (Step 2) Amazon serves each message to multiple raters (“Workers”) who use a simple web interface to classify the message; (Step 3) Analyze the data for agreement between raters. Technically speaking, the MTurk platform would get the job done, but with a corpus of messages written in English, Swahili, and a mashup of the two, we came to the conclusion that it would be difficult (impossible?) to filter the pool of MTurk workers to find the right folks for the task.</p><p>At the other end of the difficulty spectrum, we also thought about just creating a spreadsheet with messages and a dropdown menu of intents. But what would be the fun in that? Besides, we had a better idea that would not be very hard to implement.</p><h3>Our Desired Specifications</h3><ol><li>Create a web app that would serve raters with one message at a time and ask them to classify the intent of the message.</li><li>Stop serving specific messages once agreement is reached across 2 raters.</li><li>Don’t serve repeat messages to raters.</li></ol><h3>Getting Started with Shiny+Flexdashboard</h3><p>I’m going to show you how to use R Markdown and the <a href="https://rmarkdown.rstudio.com/flexdashboard/shiny.html">flexdashboard</a> package to create a simple shiny app to run a MTurk-style process.</p><p>I’m a diehard R user, lover of <a href="https://rmarkdown.rstudio.com/">R Markdown</a>, <a href="https://www.tidyverse.org/">tidyverse</a> convert, and general fan of <a href="https://www.rstudio.com/">RStudio</a>. I had been looking for an excuse to develop my <a href="https://www.rstudio.com/products/shiny/">Shiny</a> skills for creating interactive documents, and this seemed like perfect opportunity.</p><p>If you’re not familiar with Shiny, the developers describe it as:</p><blockquote>…an open source R package that provides an elegant and powerful web framework for building web applications using R. Shiny helps you turn your analyses into interactive web applications without requiring HTML, CSS, or JavaScript knowledge.</blockquote><p>They sure did get the marketing right for me: a data analyst who wants to build a web app, but can’t because he googles how to specify font colors each and every time he is forced to use HTML and CSS!</p><h4>Step 1: Get a free account at shinyapps.io</h4><p>Unlike static R Markdown outputs that can served by any web server, interactive documents with Shiny apps require a Shiny server. Right now there are three options: (a) host your own <a href="https://www.rstudio.com/products/shiny/shiny-server/">Shiny server</a>; (b) host your own instance of <a href="https://www.rstudio.com/products/connect/">RStudio Connect</a>, a broader platform for sharing your data science products; or (c) put your app on <a href="http://www.shinyapps.io/">shinyapps.io</a>, a hosted service from RStudio. I’ll show you how to do Option C.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sVZcx3tW5ZeaeU4Ew1YpRA.png" /><figcaption>shinyapps.io plans.</figcaption></figure><h4>Step 2: Setup your local machine</h4><p>If you don’t already have R installed, a shiny app is not the best first project, but go ahead and <a href="https://cloud.r-project.org/">download R</a> while you are here. Download the <a href="https://www.rstudio.com/products/rstudio/download/">RStudio desktop IDE</a> while you are at it.</p><p>You’ll also need to install a few packages:</p><p>install.packages(c(&quot;flexdashboard”, “shiny”, “rdrop2&quot;, “tidyverse&quot;, “shinyWidgets&quot;, “DT&quot;))</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*fP8160PHPy3zhcTqa6RTQw.png" /><figcaption>RStudio projects.</figcaption></figure><p>My preferred workflow is to create a git repo and create a new <a href="https://support.rstudio.com/hc/en-us/articles/200526207-Using-Projects">project</a> in RStudio that maps to this directory on my local machine. Projects let you forget about working directories and have some other nice features.</p><p><a href="https://github.com/ericpgreen/Nivi-classify-example">Here’s a link to the Github repo</a> that I created for this example.</p><h4>Step 3: Setup persistent remote storage (required for Option C)</h4><p>If you are not using a self-hosted option, you need to setup some form of remote data storage because, currently, shinyapps.io <a href="http://docs.rstudio.com/shinyapps.io/Storage.html#understanding-the-storage-lifecycle">does not store data generated from one Application Instance to another</a>. Dean Attali has a <a href="https://deanattali.com/blog/shiny-persistent-data-storage/">great guide</a> for solving this problem.</p><p>I decided to use the <a href="https://github.com/karthik/rdrop2">rdrop2</a> package to connect my Shiny app to my Dropbox account. To authenticate with Dropbox, run the following commands <em>once</em> in the R console:</p><ol><li>library(rdrop2)</li><li>drop_auth()</li><li>token &lt;- drop_auth()</li><li>saveRDS(token, file = “droptoken.rds”)</li></ol><p>This generates a token that you will upload to shinyapps.io with your app. Remember, there are other approaches. See Dean’s summary. (Shoutout to <a href="https://github.com/karthik/rdrop2/issues/61#issuecomment-423341288">Clayton Yochum</a> for helping me get things running.)</p><p>On Dropbox, create a folder to store the results and <em>copy</em> over the master.csv and raters.csv file from <a href="https://github.com/ericpgreen/Nivi-classify-example/tree/master/dash">my repo</a> to this folder (leave copies in the local repo). <strong>Do this before going to the next step.</strong></p><p>To run my example without any changes, create this folder in your top-level Dropbox directory and name it dash.</p><p>[<em>Fair warning: Using Dropbox as the remote storage solution does create a race condition. It’s technically possible for one rater to overwrite another rater’s submission if both raters submit at the same exact moment. In our use case this is not a big threat, however.</em>]</p><h4>Step 4: Design your app and dashboard in a .Rmd file</h4><p>I’ve done this step for you in the classify-example.Rmd file in my <a href="https://github.com/ericpgreen/Nivi-classify-example">example repo</a>. If you installed all of the required packages in Step 2, you should be able to open the document in RStudio, hit “Run Document”, and get the plain vanilla app to open.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zmo9EadSOY6MK0og4YQIWA.png" /><figcaption>Example flexdashboard with embedded shiny app.</figcaption></figure><p>The default setup in line 27 (remote &lt;- 0) is to use local storage so you can get the basic app running, but as explained previously, this will not work on shinyapps.io. You need to switch to remote storage. If you’ve completed Step 3, you can change remote from 0 to 1 in line 27, which will tell R to interact with Dropbox at several points in the file where there is a check for (remote==1).</p><p>(If you tested locally with remote &lt;- 0 before saving a copy of dash/master.csv to Dropbox/dash/master.csv, grab <a href="https://github.com/ericpgreen/Nivi-classify-example/tree/master/dash">a fresh copy from my repo</a> and save it to Dropbox so the remote example will work properly.)</p><h4>Step 5: Customize your content</h4><p>With your app up and running locally, it’s time to replace the content in the master.csv and raters.csv files on Dropbox. In master.csv, just paste or pipe in your messages, assign unique message IDs, and set all other values to NA. Updating raters.csv is straightforward.</p><p>You should also replace some content in the .Rmd file. Search for the vector of intents for inputId = “classification” and replace it with your own intents (search for two instances in the file). If you remove or edit the final two elements of the vector, “Another option not in this list” or “Cannot make sense of text”, search for this text in the file so you can update some conditional logic about when input selection boxes are shown/hidden.</p><h4>Step 6: Publish your app</h4><p>The last step is to publish to your Shiny server. If you’re using shinyapps.io, follow <a href="https://shiny.rstudio.com/articles/shinyapps.html">these instructions</a> to install.packages(&#39;rsconnect&#39;), obtain your token on the website, and configure RStudio to publish directly to your shinyapps.io account (or self-hosted server).</p><h4>Step 7: Make it better</h4><p>Here are a few ideas for iterating on this example:</p><ol><li>Integrate rater confidence. In this example, raters indicate their degree of confidence when labeling each message, but this information is not used.</li><li>Use a better remote storage option that avoids a race condition.</li><li>Create a process to evaluate more than just exact agreement between raters. Maybe you want to push raters to use very nuanced intents that are similar yet distinct. Requiring an exact match between raters could be limiting.</li><li>Replace dropdown user selection with <a href="https://github.com/MarkEdmondson1234/googleID">authentication</a>.</li></ol><h3>Learning More About R Markdown, Shiny, and Flexdashboard</h3><p>This is a fairly simple example that should enable you to play around and explore what parameters control how shiny apps look and function. To learn more, check out the <a href="https://rmarkdown.rstudio.com/">awesome resources from Rstudio</a> and get help over at <a href="https://community.rstudio.com/">Rstudio Community</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/135bebd9867961de420ab34cc9302bb7/href">https://medium.com/media/135bebd9867961de420ab34cc9302bb7/href</a></iframe><p>(This gist is named with a .R extension, but this is only for the syntax highlighting. This is really a .Rmd file.)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3beacba50e09" width="1" height="1" alt=""><hr><p><a href="https://medium.com/nivi-inc/building-your-own-mturk-style-app-in-r-using-shiny-flexdashboard-3beacba50e09">Building your own MTurk-style app in R using Shiny+Flexdashboard</a> was originally published in <a href="https://medium.com/nivi-inc">Nivi</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I’d go with points, not bars]]></title>
            <link>https://medium.com/data-science/id-go-with-points-not-bars-f48b64030564?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/f48b64030564</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[rstats]]></category>
            <category><![CDATA[data-visualization]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Wed, 12 Sep 2018 15:04:33 GMT</pubDate>
            <atom:updated>2018-09-12T17:46:43.709Z</atom:updated>
            <content:encoded><![CDATA[<p>Writing in <a href="https://www.nytimes.com/2018/09/10/upshot/its-hard-for-doctors-to-unlearn-things-thats-costly-for-all-of-us.html?smtyp=cur&amp;smid=tw-upshotnyt">The Upshot/NYT</a>, Aaron Carroll features <a href="https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2195119">this 2015 study</a> by Niven et al. published in <em>JAMA Internal Medicine</em> that shows how the medical practice of tight glycemic control increased after one study suggested it was a good idea, but did not decrease after a second, larger study reversed the guidelines. Carroll’s point is that the only thing harder than convincing physicians to start something might be convincing them to stop.</p><p>But let’s get to the real interesting stuff: the plots!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Uri2UkjQ7OG03OckAYSJoA.jpeg" /></figure><p>The original figure is on the left, and the NYT/Upshot version is on the right. In typical NYT fashion, they strip down the plot to reveal something more pleasing.</p><p>They stick with a bar chart—maybe to be as true as possible to the original—but I think points work best here. This is part personal preference, part <a href="https://www.edwardtufte.com/tufte/books_vdqi">Tufte-inspired wisdom</a> about data-ink ratios. The basic rationale is this: all of the data rest in the tips of bars, so why not just plot the tips?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/807/1*WYOkZwxuSGtCqMXnLU0Ofg.png" /></figure><p>I created this plot in R using the ggplot2() package. The JAMA authors might have shared the underlying data somewhere, but I did not find anything. So I used <a href="http://arohatgi.info/WebPlotDigitizer/app3_12/">WebPlotDigitizer</a> to extract approximate values from the published figure. Here’s the code if you want to tinker.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/22696b0272f5bd640a5d844007b17438/href">https://medium.com/media/22696b0272f5bd640a5d844007b17438/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f48b64030564" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/id-go-with-points-not-bars-f48b64030564">I’d go with points, not bars</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My wife couldn't sleep]]></title>
            <link>https://medium.com/data-science/my-wife-couldnt-sleep-550e206cd80a?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/550e206cd80a</guid>
            <category><![CDATA[clinical-research]]></category>
            <category><![CDATA[stillbirth]]></category>
            <category><![CDATA[baby]]></category>
            <category><![CDATA[pregnancy]]></category>
            <category><![CDATA[research]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Sat, 11 Aug 2018 14:53:35 GMT</pubDate>
            <atom:updated>2018-08-13T13:56:03.119Z</atom:updated>
            <content:encoded><![CDATA[<h3>My wife couldn’t sleep</h3><h4>She did not ignore the itch, and neither should you</h4><p>She was itchy and tired. She was also 29 weeks pregnant with our second child. Certainly not a unique struggle. Being uncomfortable and exhausted is pretty much the definition of the third trimester.</p><p>But Eve knew something was not right. When the itching started, sleeping gave way to nocturnal scratching—and Googling. At first what she found online was reassuring—pregnancy does have a way of making you itchy due to increased blood flow to the skin—but her itching seemed different. Much more intense and widespread. Then she read about a relatively rare condition of the liver that can cause <strong><em>severe itching</em></strong> in pregnancy. This disorder is associated with some pretty bad outcomes for babies, so Eve decided to ask her OB. The doc thought the chances were pretty small it was anything serious, but she ordered the tests just to be safe.</p><p>The liver functioning results came back first because the samples were run locally. Only a few labs in the country evaluate bile acid levels, so Eve’s samples had to be shipped out, a process that took 3 to 7 days. As a reward for our patience, we were told her numbers looked fine.¹ Nothing to worry about it seemed.</p><p>Another two weeks passed. I traveled to Kenya and Liberia for work. Eve went to a conference in DC. Her itch tagged along and shared a room. But lying awake in those late-night scratch-fests, she could not shake the feeling that something was wrong.</p><p>She repeated the tests and received her diagnosis on June 8 (Week 31): <strong><em>Intrahepatic Cholestasis of Pregnancy</em></strong>, or ICP.</p><h3>What is ICP?</h3><p>The liver (of the Greek <a href="https://en.wiktionary.org/wiki/hepato-"><em>hepato</em></a>) is a pretty neat organ. When you scarf down a burger at <em>Five Guys</em>, your liver produces bile to breakdown the fat. For reasons that are not fully understood, ICP interrupts this process and creates a buildup of bile acids in the blood (<a href="https://en.wikipedia.org/wiki/Cholestasis"><em>cholestasis</em></a>). If not managed, ICP is associated with an increased risk of stillbirth and respiratory issues at birth. Whether excess bile acid in the blood causes these problems or is just a marker of some other disease process is not clear.</p><p>The only “cure” for ICP is delivery; most women see their liver functioning return to normal when no longer pregnant. “Managing” ICP during pregnancy typically involves planning for an early delivery and use of the medication <a href="https://en.wikipedia.org/wiki/Ursodeoxycholic_acid">Ursodeoxycholic acid</a> (a.k.a., UDCA, Ursodiol, Actigall) that may ease the itching and reduce the amount of toxic bile acids in the blood. Serum bile acid levels greater than 10 <em>µ</em>mol/L mark the onset of ICP, and 40 <em>µ</em>mol/L is the threshold for “severe” ICP.</p><h3>Researching the optimal week for delivery</h3><p>When Eve’s results came back pointing to ICP, we tried to learn everything we could about the condition. I landed on the website for <a href="http://www.icpcare.org/">ICP Care</a> on June 9, which happened to be the day before <a href="http://www.icpcare.org/get-involved/icp-awareness-month/">ICP Awareness Day</a>. Here’s what grabbed my attention:</p><blockquote>Without active management the risk of stillbirth during an Intrahepatic Cholestasis of Pregnancy can be as high as 15%. However, with active management, including the medication Ursodeoxycholic Acid and early delivery (typically by 36–37 weeks gestation), the risk is thought to be about the same as an uncomplicated pregnancy, meaning less than 1%. <strong>The mechanisms that lead to stillbirth are not fully understood, and it is impossible to predict which pregnancies are at risk.</strong> While studies have found that higher bile acid levels appear to bring higher risk of certain complications such as respiratory distress and fetal distress, studies have not been able to identify the same relationship for stillbirth. Therefore, all ICP pregnancies are recommended to be delivered early.</blockquote><p>We were confident that the overall risk of stillbirth <em>should</em> be low since Eve was being “actively managed” with UDCA and had tentative plans to deliver at 37 weeks, but two questions led us to start digging into the research:</p><ol><li>Is 37 weeks the <em>optimal</em> time to deliver?</li><li>Eve’s bile acid level at diagnosis was 25 <em>µ</em>mol/L. What if it started to rise? Is there a threshold that should push us to deliver even earlier?</li></ol><h4>What does the risk of stillbirth look like by week?</h4><p>When it comes to ICP, the “optimal timing” for delivery is the point at which the risk of doing <em>something</em> is less than the risk of doing <em>nothing</em>. Wait too long to give birth, and you risk stillbirth. Act too soon, and you risk infant death or complications following a live birth.</p><p>It should give you some insight into our risk calculation that we started our research journey by looking for studies that estimated the risk of stillbirth by week. The largest study we found on stillbirth rates was a <strong><em>retrospective cohort</em></strong> of 1.6 million pregnancies of women with and without ICP from the state of California.</p><p>In a cohort study, researchers identify people who meet certain criteria and follow them over time to see who develops the outcome of interest. In this particular study, <a href="https://www.ncbi.nlm.nih.gov/pubmed/25687562"><strong>Puljic and colleagues</strong></a> selected women from a vital statistics registry based on their exposure—ICP or no ICP—and extracted linked data on maternal, fetal, and infant outcomes. This is considered a <em>retrospective</em> (or <em>nonconcurrent</em> or <em>historical</em>) cohort design because the pregnancies and outcomes had already happened by the time the authors started their investigation.</p><p>In their dataset of 1.6 million pregnancies, the researchers found 5,545 cases of ICP, an incidence of 0.35%. This is lower than what other studies have found, meaning it’s possible that some proportion of actual ICP pregnancies were not coded correctly. Or maybe the other studies overestimated the incidence of ICP and this study got it right. It’s hard to know.</p><p>The authors present some results in tabular form that I think are better represented as figures, so I attempted a few plots (code is <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>). The following graphic charts the risk of stillbirth by week for women with and without ICP. It shows that women with ICP have a higher risk of stillbirth at every week compared to women without ICP, but the weekly numbers for the ICP group are so small that the estimates are very uncertain (the yellow band is the 95% confidence interval for ICP, and it includes all of the purple control values which are very precisely estimated). The clearest finding is a substantial increase in the risk of stillbirth for women with ICP at the end of term. By definition these cases were not being fully “managed” with early delivery since they did not deliver before 38 weeks, but we don’t know from the data what percentage were taking UDCA.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eq1Ru_K6v3Vqv4qOhuhf9w.png" /><figcaption>Data for this plot come from Table 2 in Puljic et al. (2015). The risk of infant and fetal death by each additional week of expectant management in intrahepatic cholestasis of pregnancy by gestational age. <a href="https://www.ncbi.nlm.nih.gov/pubmed/25687562"><em>American Journal of Obstetrics and Gynecology</em>, <em>212</em>(5), 667-e1</a>. Plot code <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>.</figcaption></figure><h4>What about the flip side—the risks of delivering too early?</h4><p>This retrospective cohort study by <a href="https://www.ncbi.nlm.nih.gov/pubmed/25687562">Puljic and colleagues</a> also offers partial data on the risks of delivering early. Among the group of women with ICP, the authors calculated the weekly risk of infant death associated with delivery—a.k.a. doing something—and compared it to what they call the composite mortality risk of <em>expectant management</em>—a.k.a. waiting and doing nothing. The composite risk is the combined risk of stillbirth in a given week plus the risk of infant death if delivery occurs the following week. <em>In other words, they compared the risk of delivering in a given week compared to the risk of waiting one more week.</em></p><p>Here’s my attempt to plot their data. The authors also plot the point estimates in the paper but not the uncertainty, and I think we need to see both elements. The lines cross at Week 36, suggesting that this could be the optimal point at which the risks of acting are lower than the risks of waiting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P5OIH1i2qRgg7PGeyMQn5w.png" /><figcaption>Data for this plot come from Table 3in Puljic et al. (2015). The risk of infant and fetal death by each additional week of expectant management in intrahepatic cholestasis of pregnancy by gestational age. <a href="https://www.ncbi.nlm.nih.gov/pubmed/25687562"><em>American Journal of Obstetrics and Gynecology</em>, <em>212</em>(5), 667-e1</a>. Plot code <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>.</figcaption></figure><p><strong>Of course what’s missing from this risk profile is perinatal<em> morbidity</em>—bad outcomes for moms and babies that stop short of mortality.</strong> These can include conditions like cerebral palsy and respiratory distress syndrome, outcomes that are more likely with premature deliveries. It’s not possible to get a true evaluation of the risks of early delivery without considering morbidity.</p><p><a href="https://www.ncbi.nlm.nih.gov/pubmed/25371372"><strong>Lo and colleagues</strong></a> help us out in this respect with their <em>decision-analytic model</em>. In this study, the authors simulated outcomes for a hypothetical cohort of 10,000 women with ICP based on (a) maternal and neonatal life expectancies, (b) maternal and neonatal utility values for several bad outcomes found in the literature, and (c) published probabilities of neonatal and maternal morbidities, stillbirth with ICP, and neonatal death at 35–38 weeks gestation for four different delivery strategies:</p><ol><li>delivery 48 hours after administering steroids (to promote lung development), without amniocentesis</li><li>delivery 48 hours after administering steroids, with amniocentesis</li><li>amniocentesis and re-testing, no steroids</li><li>immediate delivery (no amniocentesis, no steroids)</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/941/1*BmJwWOIAK20gksRxC7Alxw.png" /><figcaption>Data for this plot come from Lo et al. (2015). Intrahepatic cholestasis of pregnancy and timing of delivery. <a href="https://www.ncbi.nlm.nih.gov/pubmed/25371372">Journal of Maternal-Fetal &amp; Neonatal Medicine, 28(18), 2254–2258</a>. Plot code <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>. This plot shows additional QALYs relative to the minimum QALY value, which is delivery at 38 weeks following amniocentesis and steroids. This minimum value is set to 0 in the figure.</figcaption></figure><p>The primary outcome was total <strong><em>quality-adjusted life years</em></strong>, or QALYs. One year of good health is equal to 1 QALY, so the more projected QALYs per person the better. The main results of the simulation are shown here. <strong>Delivering immediately at 36 weeks without amniocentesis or steroids appears to maximize QALYs relative to all other strategies.</strong></p><p>The authors conclude:</p><blockquote>When considering the tradeoffs, while there is a higher association of neonatal morbidity with delivery at 36 versus at 37 weeks gestation or later, there is a decreased risk of [stillbirth].</blockquote><p>The following plots show the trends behind the QALY calculations and this conclusion. The risk of outcomes like cerebral palsy and respiratory distress syndrome are highest before 37 weeks, but this is offset by an increased risk of stillbirth in weeks 37 and 38.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wMjTOcDnDg7hG3tDqmasSQ.png" /><figcaption>Data for this plot come from Lo et al. (2015). Intrahepatic cholestasis of pregnancy and timing of delivery. <a href="https://www.ncbi.nlm.nih.gov/pubmed/25371372">Journal of Maternal-Fetal &amp; Neonatal Medicine, 28(18), 2254–2258</a>. Plot code <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>. These plots show the patterns of risk for three specific outcomes according to the model: the risks of cerebral palsy and respiratory distress syndrome decrease as babies mature in the womb, while the risk of stillbirth among women with ICP increases every week.</figcaption></figure><p>While not a slam dunk, these two studies made us lean toward delivering at 36 weeks instead of 37 weeks as planned. We were further convinced of this when Eve’s next bile acid test came back at 85 <em>µ</em>mol/L the following week.</p><p>When we got this result, I was reading a newly published 2018 review of ICP care published by Dr. Elizabeth Livingston and her colleagues at the Division of Maternal Fetal Medicine in Duke’s OB/GYN Department. In this review, <a href="https://www.ncbi.nlm.nih.gov/pubmed/29480924"><strong>Livingston and colleagues</strong></a> write:</p><blockquote>Given the increased risk of stillbirth in the setting of ICP, induction of labor is often recommended at 37 weeks of gestation to balance the risk of iatrogenic preterm delivery against the risk of fetal mortality. <strong>As there is a markedly increased risk of stillbirth when bile acid levels reach 100 μmol/L or greater, delivery in these cases may be considered at an earlier gestational age, such as 35 to 37 weeks; however, these decisions should be individualized with careful patient counseling.</strong></blockquote><p>Eve was a patient in the Duke Health system, so we met with Dr. Livingston to review our options. Livingston felt comfortable with delivery at 36 weeks given Eve’s high result. But what if her levels crossed the 100 <em>µ</em>mol/L mark?</p><h4>What is the relationship between bile acid level and stillbirth?</h4><p>The paper by <a href="https://www.ncbi.nlm.nih.gov/pubmed/29480924"><strong>Livingston and colleagues</strong></a> reviews the latest evidence about this relationship and summarizes the main hypotheses about how bile acids might cause stillbirth. One idea is that excess bile acids constrict the blood vessels in the placenta. This might be one reason why stillbirth in ICP seems to be the result of an acute event leading to a lack of oxygen and death, not a chronic deterioration of fetal wellbeing. If so, it helps to explain why fetal monitoring—which Eve did twice weekly—does not reduce the risk of stillbirth in ICP (but is still recommended).</p><p>Right now, high bile acid levels may be the best diagnostic indicator of increased risk for stillbirth, but this is still a matter of some debate. A key piece of evidence comes from a prospective cohort study of women with severe ICP conducted in the UK and published in 2014 in the journal <a href="https://www.ncbi.nlm.nih.gov/pubmed/23857305"><em>Hepatology</em></a>. In this study, the cohort was made of up 669 women with <strong>severe ICP</strong> (defined as bile acid levels ≥ 40 <em>µ</em>mol/L at any time during pregnancy) and the outcome of interest was stillbirth.</p><p>Here are some notable findings from the severe ICP cohort:</p><ol><li>Most of these women were treated with UDCA (72%), and the average woman delivered at 37.5 weeks. Therefore, you could say that “active management” was fairly common, but certainly not universal.</li><li>10 of the 669 women experienced a stillbirth, an incidence of 1.5%. (represented by red plus signs in the figure below)</li><li>6 of the 10 stillbirths happened <strong><em>before</em></strong> 37 weeks.</li><li>Stillbirths seem to cluster in women with bile acid levels above 100 <em>µ</em>mol/L.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/618/1*OVl8PYt0GXmUTb5mB1lOlQ.png" /><figcaption>Source: Geenes, V. et al. (2014). Association of severe intrahepatic cholestasis of pregnancy with adverse pregnancy outcomes: a prospective population‐based case‐control study. <a href="https://www.ncbi.nlm.nih.gov/pubmed/23857305"><em>Hepatology</em>, <em>59</em>(4), 1482–1491</a>. Figure 4(C). The estimated probability and 95% CIs of stillbirth in relation to the maternal serum bile acid level, based on simple logistic regression.</figcaption></figure><p>The authors also looked at stillbirth outcomes in two national comparison groups of women without ICP. At 1.5%, the incidence of stillbirth among the severe ICP cohort was about 3 times as high as the incidence of stillbirth among the non-ICP comparison groups.</p><h4>Eve’s trajectory</h4><p>Ultimately, Eve’s bile acid levels would never cross the 100 <em>µ</em>mol/L mark. Her peak level was 85.6 <em>µ</em>mol/L. The final test result prior to delivery was 38.4 <em>µ</em>mol/L.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fmzyu9Yur5ESKz-dUt1mZw.png" /><figcaption>Plot code <a href="https://gist.github.com/ericpgreen/e0d34b65cba75d070ee91a92c5f4434a">here</a>.</figcaption></figure><h3>Our little itch arrives</h3><p>After a stressful few weeks, Annie Grace was born on July 9, Eve’s birthday. 36 weeks and 1 day. We felt incredibly fortunate. If you’ve been diagnosed with ICP, the available evidence suggests that chances are great that you will have a good outcome too, especially if your doctor prescribes UDCA and schedules your delivery at 36 or 37 weeks (or sooner if it makes sense given your situation).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fnh8PQPxDjlEMMKsu8mQeA.jpeg" /><figcaption>Annie Grace</figcaption></figure><p>Unfortunately, not everyone is so lucky. Early in our journey I read Donna Benavides’s <a href="https://www.whattoexpect.com/wom/pregnancy/my-life-changing-experience-with-cholestasis-of-pregnancy.aspx">story</a>. She was diagnosed with ICP at 36 weeks and scheduled to be induced at 38 weeks.</p><blockquote>I was confused and scared, this being my first pregnancy. I read more about cholestasis of pregnancy online and printed out some information to discuss with my husband because the statistics on the article I read noted a small percentage of loss of the baby due to ICP. <strong>The day after my reactive [non-stress test] — just two days from being induced — we lost our son Jorden.</strong></blockquote><p><a href="http://www.icpcare.org/about/team-volunteers/">Donna</a> would go on to found <a href="http://I also read Donna Benavides&#39;s story. She was diagnosed with ICP at 36 weeks and scheduled to be induced at 38 weeks.  I was confused and scared, this being my first pregnancy. I read more about cholestasis of pregnancy online and printed out some information to discuss with my husband because the statistics on the article I read noted a small percentage of loss of the baby due to ICP. The day after my reactive [non-stress test]- just two days from being induced - we lost our son Jorden. Donna would go on to found ICP Care and help countless families make sense of a frightening diagnosis.">ICP Care</a> and help countless families make sense of a frightening diagnosis. Now she tells other women to fight the itch.</p><p>Donna’s doctor did not know to prescribe UDCA. While ICP is better understood today, it’s still common to read stories on ICP Care’s <a href="https://www.facebook.com/groups/ICPCare/about/"><strong><em>Itchy Moms</em></strong></a> Facebook group of women who struggle to get their providers to test for ICP and take their symptoms seriously.</p><p>Be empowered to get tested. Prepare for your appointment with this <a href="http://www.icpcare.org/resources/dr-discussion-guide/">Dr. Discussion Guide</a>, and share <a href="http://www.icpcare.org/resources/educational-brochures/">these resources</a> with your provider.</p><p><strong><em>Don’t ignore the itch.</em></strong></p><p>[1] Eve’s first bile acid level was actually 11.5 <em>µ</em>mol/L, which is above the cutoff that the <a href="https://www.ncbi.nlm.nih.gov/pubmed/29480924">Wood et al. (2018)</a> recommend for diagnosing ICP. Her result was classified as “normal”, possibly because the reference range in Duke MyChart for pregnant women ages 15 to 45 was 3.7–14.5 <em>µ</em>mol/L.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=550e206cd80a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/my-wife-couldnt-sleep-550e206cd80a">My wife couldn&#39;t sleep</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Alluvial Diagrams]]></title>
            <link>https://medium.com/data-science/alluvial-diagrams-783bbbbe0195?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/783bbbbe0195</guid>
            <category><![CDATA[r]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Wed, 01 Aug 2018 15:30:10 GMT</pubDate>
            <atom:updated>2018-08-02T18:54:53.995Z</atom:updated>
            <content:encoded><![CDATA[<h4>Plotting pathways over time</h4><p>A paper I co-wrote was published last week in the <a href="https://onlinelibrary.wiley.com/doi/pdf/10.1111/jcpp.12955">Journal of Child Psychology and Psychiatry</a>. This paper reports the results of a secondary analysis of the impact of a school support intervention on depression among orphans in Kenya. In the process of preparing the data for <a href="https://ericpgreen.github.io/JCPP2018/jcpp.html">public archiving</a>, I thought it would be neat to try visualizing our longitudinal data on depression as an alluvial diagram.</p><p>The figure below is an example of an alluvial diagram of the Titanic dataset created by Michał Bojanowski, author of the <a href="https://github.com/mbojan/alluvial">alluvial package</a> for R. The diagram flows backwards from whether passengers survived. For instance, you can trace back from Survived==&quot;No&quot; to see that most of the Titanic crew did not survive.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*ov7EnESNigBdhdJ3NYBoBg.png" /><figcaption>Alluvial diagram of the Titanic dataset <a href="https://cran.r-project.org/web/packages/alluvial/vignettes/alluvial.html">created by Michal Bojanowski</a>.</figcaption></figure><p>I’m going to show you how to use this package to create a different type of alluvial diagram that depicts how individuals flow across time points.</p><h3>The Data</h3><p>My colleagues designed a cluster randomized trial in Kenya and recruited more than 800 adolescents from 26 primary schools to participate in a study of school support (i.e., free tuition, school uniforms, and nurse visits). As part of a baseline survey, kids were asked a subset of questions from a questionnaire called the <a href="http://cesd-r.com/">CESD-R</a> to assess depression severity. Half of the schools were then randomly assigned to receive the school support intervention, and the research team followed the kids over four years.</p><p>In our secondary analysis, we created an indicator of depression that was the average of five CESD-R questions that were rescaled to have a range of 0 to 3. Higher scores indicate greater severity.</p><h3>Group Means Over Time</h3><p>Here is Figure 1 from our paper showing the mean depression score by study arm and year. It’s a concise summary of the trial data, but it does not give us a sense of how individuals changed over time. All of that information is collapsed into the means.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*Y0tf1qrIzMD1WOE0QEAukw.png" /></figure><h3>Change in Depression Status from Year 1 to Year 4</h3><p>The line plot shows that, on average, the groups were similar in terms of severity until Year 4 when the control group’s score increased, creating separation from the treatment group. So let’s explore what’s happening in Year 4.</p><p>One way of looking at the data is to label participants as “possibly depressed” or not based on their yearly severity score and compare the study arms on this status in Year 4 relative to where they started in Year 1. This is what we did in Figure 3, labeling everyone as:</p><ul><li><strong>Never depressed</strong> (below the cutoff score for possible depression in Years 1 and 4)</li><li><strong>Still depressed</strong> (above the cutoff in Years 1 and 4)</li><li><strong>Became depressed</strong> (below the cutoff in Year 1, but above in Year 4)</li><li><strong>Remission</strong> (above the cutoff in Year 1, but below in Year 4)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/948/1*GLymh7P-wRzKxXFSZtFaPA.png" /></figure><h3>Individual Change</h3><p>This figure hints at the fact that participants have different pathways to their Year 4 depression scores, but the data are still summarized by group. Let’s try dropping all aggregation and showing the scores for each adolescent at each time point.</p><p>Here’s the script to get the raw data and wrangle it into the right shape for plotting:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/340bc48c14d9cc8727b99f7df06147f9/href">https://medium.com/media/340bc48c14d9cc8727b99f7df06147f9/href</a></iframe><p>And now we can plot each person’s yearly score as a line:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4776f63ae03f7955f608d3eba5f4f2a7/href">https://medium.com/media/4776f63ae03f7955f608d3eba5f4f2a7/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/1*uMHX2PKuUWLEvYxV5iLfUw.png" /><figcaption>Individual scores over time faceted by study arm. The blue line is participant #2, and the red line is participant #3. Everyone else is grey.</figcaption></figure><p>What a mess! Data science may be both art and science, but this plot is is squarely in the art camp.</p><h3>More Nuanced Aggregation</h3><p>Now I’ll use the alluvial package to lump participants together by pathway. Here’s the code:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/bad8664c1fc021dbfbb6119a45ca88cb/href">https://medium.com/media/bad8664c1fc021dbfbb6119a45ca88cb/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CXKMJmgje2W3pQGUyfA9Zg.png" /></figure><p>This plot does not facet on study arm to keep things simple, but such faceting is possible. The bigger picture to grasp is how alluvial plots can trace all possible pathways that participants take in a longitudinal dataset.</p><p>For instance, adolescents who never scored above the depression cutoff form a straight-line grey band at the bottom from that goes from “not depressed” in Year 4 all the way back to “not depressed” in Years 3, 2, and 1. There is a parallel <em>always depressed</em> group represented by the straight-line red band at the top. The waves in the middle are participants who changed status one or more times.</p><p>The next time you see reports based on pre/post data, you might wonder what is happening in between observations. If you actually collect longitudinal data, try plotting this change with an alluvial diagram.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=783bbbbe0195" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/alluvial-diagrams-783bbbbe0195">Alluvial Diagrams</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fighting glioblastoma with recombinant poliovirus]]></title>
            <link>https://medium.com/@ericpgreen/fighting-glioblastoma-with-recombinant-poliovirus-5bef1ebc7f41?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/5bef1ebc7f41</guid>
            <category><![CDATA[health]]></category>
            <category><![CDATA[brain-cancer]]></category>
            <category><![CDATA[research-methods]]></category>
            <category><![CDATA[clinical-trials]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Fri, 29 Jun 2018 17:08:05 GMT</pubDate>
            <atom:updated>2019-08-23T18:52:32.573Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DQizWXx_a5QIchYWYft6ow.png" /><figcaption>Check out this great report from <a href="https://www.cbsnews.com/news/trial-shows-poliovirus-may-be-used-to-treat-deadly-brain-cancer-glioblastoma/">CBS News</a>.</figcaption></figure><h4>A look at this Phase 1 trial design</h4><p>A study published recently in the <a href="https://www.nejm.org/doi/full/10.1056/NEJMoa1716435?query=featured_home"><em>New England Journal of Medicine</em></a> shows that patients with recurrent glioblastoma who were infused with a modified poliovirus at <a href="https://tischbraintumorcenter.duke.edu/">Duke University</a> <strong>were more likely to be alive 3 years later</strong> compared to similar patients who did not receive the treatment. The publication gained the attention of several media outlets, including the <a href="https://www.washingtonpost.com/news/to-your-health/wp/2018/06/26/polio-virus-treatment-increased-survival-in-patients-with-deadly-brain-tumors-study-shows/?utm_term=.20b822c7ec1e">Washington Post</a>, <a href="https://www.npr.org/sections/health-shots/2018/06/26/622610333/doctors-try-genetically-modified-poliovirus-as-experimental-brain-cancer-treatme">NPR News</a>, and CBS News, who has reported on the study’s progress and profiled several patients over the past few years (see <a href="https://www.cbsnews.com/news/polio-cancer-treatment-duke-university-60-minutes-scott-pelley/">here</a> and <a href="https://www.cbsnews.com/news/trial-shows-poliovirus-may-be-used-to-treat-deadly-brain-cancer-glioblastoma/">here</a>).</p><p>I’ve followed this study with great interest. For several semesters now, I’ve showcased this trial in my course on <a href="https://themethodsection.com/">Global Health Research Methods</a> during our introduction to research and clinical trials. Now that the results are published, I decided to write a teaching note that breaks down the article and explores interesting concepts that we discuss throughout the semester. I encourage you to read the manuscript and supplemental materials for yourself.</p><p>Let’s take a closer look at the design and implementation of this amazing Phase 1 trial.</p><h3>Poliovirus in the brain. “What the hell?”</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3oEi9ziigxVxTUiFJznt7A.png" /><figcaption>ClinicalTrials.gov, <a href="https://clinicaltrials.gov/ct2/show/NCT01491893">NCT01491893</a>.</figcaption></figure><p>The poliovirus Phase 1 trial is registered on ClinicalTrials.gov (<a href="https://clinicaltrials.gov/ct2/show/NCT01491893">NCT01491893</a>) as the “PVSRIPO for Recurrent Glioblastoma” study. <a href="https://en.wikipedia.org/wiki/PVSRIPO">PVSRIPO</a> is the name of the modified poliovirus created by <a href="https://medicine.duke.edu/faculty/matthias-gromeier-md">Dr. Matthias Gromeier</a> in the 1990s when he was a postdoc (shoutout to all you postdocs out there!). When Gromeier came to Duke, he was reportedly eager to explore how PVSRIPO might work against brain tumors.</p><p><strong>That’s right. His idea was to modify the virus that causes poliomyelitis—yeah, that </strong><a href="https://en.wikipedia.org/wiki/Poliomyelitis"><strong>poliomyelitis</strong></a><strong>—and inject it into the brain.</strong></p><p>This <a href="https://alumni.duke.edu/magazine/articles/does-matthias-gromeier-know-better-way-treat-cancer">wonderful profile</a> of Gromeier by Barry Yeoman captures just how nuts this idea seemed at the time:</p><blockquote>‘Poliovirus is the virus that’s most capable of causing the most damage in the brain,’ [Gromeier] says. It invades the central nervous system and can paralyze the muscles we need to walk, swallow, and even breathe. ‘It may sound counterintuitive,’ he says, ‘but I saw this as a sign that it might be a good agent to be used in the brain.’</blockquote><blockquote><strong>‘Counterintuitive’ was an understatement.</strong> To some experts, it was outrageous. “Early on, when we first heard about it, it was ‘What the hell?’ ” says Mayo Clinic’s Russell.</blockquote><p>If you want to learn more about how the modified virus and the body team up to give GBM a run for its money, check out <a href="https://www.ncbi.nlm.nih.gov/pubmed/28931654">this 2017 paper</a> by Gromeier and colleagues published in <em>Science Translational Medicine.</em></p><h3>Clinical trial phases</h3><p>The road from crazy ideas to approved medical treatments is long, and most drug candidates don’t make it. The results of the initial poliovirus trial published in the NEJM this week have been more than 20 years in the making, and the science is just taking shape. This is because new drugs must go through several <a href="https://en.wikipedia.org/wiki/Phases_of_clinical_research">phases of research</a> prior to becoming a new medical treatment. (I focus here on drugs, but the same can be same about medical devices and procedures. There are also clinical trials of behavioral interventions, but the process looks a bit different.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6hExhq-K-GMRSWvNLvnjfA.jpeg" /><figcaption>Great overview of clinical trial phases from <a href="http://www.brainsforthecure.org/having-a-recurrence/finding-treatment/clinical-trials/clinical-trial-phases/">Brains For The Cure</a>.</figcaption></figure><h4>Preclinical and Phase 0</h4><p>Human trials are often preceded by <a href="https://en.wikipedia.org/wiki/Phases_of_clinical_research#Pre-clinical_studies">preclinical research</a> in test tubes or cell cultures (in vitro) or in animals (in vivo) to determine whether the treatment is a good candidate to become an investigational treatment. A Phase 0 study might follow with a handful of human participants who get a very low, ‘subclinical’ dose of the medication or treatment, just to give investigators an early look at what happens to a drug in the body (pharmacokinetics).</p><h4>From Bench to Bedside</h4><p>Because Phase 0 studies are relatively uncommon, Phase 1 trials are typically the point at which a drug jumps from basic research to clinical (applied) research—from bench to bedside.</p><p>The main goal of a Phase 1 trial is typically to evaluate safety and determine an optimal dose to put forward for future trials. These initial trials are often small (up to a few dozen patients) and have the shortest timeline of all clinical trial phases.</p><p>Contrary to what people often think, the word ‘trial’ does not mean ‘randomized’. In a Phase 1 trial, everyone typically gets the investigational treatment, though the dose may vary. In the poliovirus trial, for instance, participants received different doses, but everyone was treated with a clinical dose of the modified virus.</p><p>At this early stage, the focus is safety, so investigators keep detailed records of side effects and adverse events. They track clinical outcomes too, but it’s not typically expected that a Phase 1 trial will demonstrate a drug’s efficacy. That usually comes in Phases 2 and 3 with larger studies and random assignment to treatment arms.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdsfPOpE-GEs%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdsfPOpE-GEs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdsfPOpE-GEs%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8f77c9420ffac51e6ce8f6b5a4e5a310/href">https://medium.com/media/8f77c9420ffac51e6ce8f6b5a4e5a310/href</a></iframe><h3>Poliovirus Phase 1 trial design</h3><p>This was a single group, open label study in which all 61 participants received the experimental treatment. ‘Open label’ just means that every patient knew that he or she was treated. No one was blind to the treatment assignment because everyone was assigned to the treatment.</p><p>Let’s step through the Methods section to learn more about the design. You’ll probably want to reference three additional documents:</p><ol><li>The study registration on ClinicalTrials.gov (<a href="https://clinicaltrials.gov/ct2/show/NCT01491893">NCT01491893</a>)</li><li>The <a href="https://www.nejm.org/doi/suppl/10.1056/NEJMoa1716435/suppl_file/nejmoa1716435_protocol.pdf">study protocol</a></li><li>The authors’ <a href="https://www.nejm.org/doi/suppl/10.1056/NEJMoa1716435/suppl_file/nejmoa1716435_appendix.pdf">Supplemental Appendix</a></li></ol><p>Studies like this are very complex, and 3000–6000 words is often not enough space to give enough detail for curious readers to evaluate your claims, let alone attempt to replicate your study. Lots of good stuff lives in supplemental files!</p><h4>Study objectives</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RSM7-62rFyMQO_nMDqmbrg.png" /><figcaption>Desjardins et al. (2018). Recurrent glioblastoma treated with recombinant poliovirus. NEJM. <a href="https://www.nejm.org/doi/full/10.1056/NEJMoa1716435?query=featured_home">DOI: 10.1056/NEJMoa1716435</a>.</figcaption></figure><p>As a Phase 1 trial, the primary objective was to evaluate safety and determine the optimal dose for a Phase 2 study. This means that the study was designed first and foremost to answer this question. Estimating survival relative to a historical control group was a secondary objective. The authors also identified two exploratory or hypothesis-generating (rather than -testing) objectives: (a) characterize changes observed in MRI scans during the course of treatment and (b) identify potential biomarkers—in this case, something measurable in tumors—that could predict which patients will have a positive response to the treatment. As we’ll see, the data from this Phase 1 study shows that some people respond big time, but this is not universal.</p><h4>Patient recruitment</h4><p>Every trial must define inclusion and exclusion criteria. What are the essential characteristics patients must have to enroll, and what conditions rule people out? In an early stage trial like this one, these criteria help to ensure that researchers enroll the appropriate patients and limit exposing more at-risk groups to potential harm (e.g., no pregnant women or breastfeeding mothers).</p><p>In later trials that aim to evaluate efficacy in routine clinical practice—<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3181997/">pragmatic trials</a>—the list of inclusion and exclusion criteria might be shorter to maximize generalizability. To take a silly example, only enrolling left-handed red-heads named Sally might limit the extent to which the results apply to the target population of all women.</p><p>In the poliovirus trial, the key inclusion criteria were:</p><ul><li>recurrence of a WHO grade IV malignant glioma with certain features</li><li>high functioning as measured by the <a href="https://en.wikipedia.org/wiki/Performance_status#Karnofsky_scoring">Karnofsky performance test</a></li></ul><p>The full set of inclusion and exclusion criteria are listed in Table S1 of the Supplemental Appendix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/1*3Zw-sUflhx3vu6KHSriV5g.png" /><figcaption>Table S1 from the <a href="https://www.nejm.org/doi/suppl/10.1056/NEJMoa1716435/suppl_file/nejmoa1716435_appendix.pdf">Supplemental Appendix</a>.</figcaption></figure><h4>Historical controls</h4><p>While the primary objective of the trial was to evaluate safety and largest practical dose, the researchers also wanted to get an initial look at “<a href="https://clinicaltrials.gov/ct2/show/NCT01491893">the efficacy of PVSRIPO administered at the optimal dose</a>” in terms of overall survival.</p><p>Since this was a single group study in which all participants were treated, the researchers constructed a “historical” control group for comparison. These were patients previously treated at Duke who, based on medical record data, would have met the criteria for enrollment had the trial been recruiting back then. In the Supplement, the authors note two reasons for taking this specific approach:</p><ol><li>They felt that a sham infusion would have been unethical (presumably because the experimental treatment was not combined with another standard treatment in this trial; therefore, a sham infusion would have meant no treatment at all).</li><li>They believed that comparing survival rates among trial participants to published results of other treatments would also be flawed because the studies had different inclusion and exclusion criteria.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/868/1*0D3cvP4cVCalUJJW_NiqAg.png" /><figcaption>Explanation from the <a href="https://www.nejm.org/doi/suppl/10.1056/NEJMoa1716435/suppl_file/nejmoa1716435_appendix.pdf">Supplemental Appendix</a>.</figcaption></figure><p>Historical cohorts are not perfect comparisons. For instance, this study compares two cohorts of patients treated over two different time periods. It’s also likely that some patients from the historical cohort would not have chosen to participate in the trial had they had the opportunity. This means it’s possible that the patients who volunteered to be part of the poliovirus trial could be different in some unobserved ways from the some segment of the historical cohort.</p><p>But nothing is perfect. I think the use of a historical cohort was a nice addition to this Phase 1 design. Table 1 in the main paper suggests that the groups were mostly similar on observed characteristics.</p><p><a href="https://www.washingtonpost.com/news/to-your-health/wp/2018/06/26/polio-virus-treatment-increased-survival-in-patients-with-deadly-brain-tumors-study-shows/?noredirect=on&amp;utm_term=.193d0114af87">As reported in the <em>Washington Post</em></a>, Deepa Subramaniam of the Georgetown Lombardi Comprehensive Cancer Center is quoted as saying that this approach is “suboptimal”, which is true in general, but it’s not clear to me that a randomized design would have been optimal for this Phase 1 study. (Protip: use the word “suboptimal” and you will probably be right more than you are wrong.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eJiD6o0IAnK_K6hpr-w2GA.jpeg" /><figcaption>Snapshot of Table 1 from Desjardins et al. (2018). Recurrent glioblastoma treated with recombinant poliovirus. NEJM. <a href="https://www.nejm.org/doi/full/10.1056/NEJMoa1716435?query=featured_home">DOI: 10.1056/NEJMoa1716435</a>.</figcaption></figure><h4>Dose escalation and dose expansion</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/172/1*GbxSABEDYDYFZf5kUh9YYA.png" /><figcaption>Dose level for the first 16 patients.</figcaption></figure><p>Patients in this trial received 1 of 7 doses of PVSRIPO. A patient’s dose was determined by his or her order of enrollment in the trial and by a set of rules. The first person received dose level 1, the second dose level 2, etc, until reaching dose level 5.</p><p>According to the Supplement: “The starting dose was “1 x 10⁸ tissue culture infectious dose (TCID50), which is <strong>1/10th of the highest non-toxic dose in non-human primates</strong>…and <strong>1/50th of the highest non-toxic dose in NHPs in the dose range-finding toxicology study</strong>.”</p><p>The plan was for 21 patients to be administered dose level 5, but only 4 patients actually received this dose. As described in the paper, the last patient to receive dose level 5 suffered a “dose-limiting toxic event”¹: a grade 4 intracranial hemorrhage following catheter removal. This was unexpected as “preclinical data suggested that dose-limiting toxicity (DLT) would not occur at any of the five DLs evaluated”. However, the team had a clear plan for making decisions about how to proceed with dosing. Since 1 of the 4 patients who received dose level 5 experienced a dose-limiting toxic event (25%), they changed plans:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*x-fReho5575rZmdDT1Qe6w.png" /></figure><p>This trial is an example of <strong>dose-escalation</strong> and <strong>dose-expansion</strong>. Iasonos O’Quigley explain these phases in <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4758199/">this article published in <em>Nature Reviews Clinical Oncology</em></a><em>:</em></p><blockquote>Phase I trial designs increasingly go beyond their former focus on safety, and aim to identify the most-promising agents by adding dose-expansion cohorts, before moving to phase II testing. <strong>Such phase I trials now frequently include a dose-escalation phase that determines the maximum tolerated dose (MTD), followed by a dose-expansion phase to determine the recommended dose.</strong></blockquote><p>The dose-expansion phase included 52 of the 61 patients in the trial:</p><ul><li>6 patients received dose level 2</li><li>24 patients received dose level -1</li><li>15 patients received dose level -2</li><li>the final 7 patients went back up to dose level -1</li></ul><p>The authors explain this deescalation pattern as follows:</p><blockquote>…imaging changes that were suggestive of localized inflammation were managed with glucocorticoids if they were associated with neurologic symptoms. Given the protracted course of such peritumoral inflammation before tumor contraction, symptomatic patients had to continue taking glucocorticoids for extended periods, which exposed them to substantial side effects. To mitigate peritumoral inflammation, the PVSRIPO dose was deescalated gradually to dose level 2 and then to dose levels −1 and −2.</blockquote><h3>The results</h3><p>With respect to the primary objective of studying safety and finding an optimal dose, the authors note in the abstract that “dose level −1 was identified as the phase 2 dose”.</p><p>But it was the secondary objective that grabbed widespread media attention this week.</p><p>The overall survival rate among the trial cohort was 21% (95% CI, 11 to 33) 3 years after infusion compared to 4% (95% CI, 1 to 9) among the historical control group. <strong>As of March 20, 2018, 3 patients had survived more than 57 months!</strong> (See the three tick marks at the far right of the PVSRIPO line indicating censored data.)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*goC5viaO-d9M-LVmU8V0YA.jpeg" /><figcaption>Overall Survival among Patients Who Received PVSRIPO and Historical Controls. Tick marks indicate censored data. PVSRIPO denotes recombinant nonpathogenic polio–rhinovirus chimera. Source: A Desjardins et al. N Engl J Med 2018. <a href="https://www.nejm.org/doi/full/10.1056/NEJMoa1716435?query=featured_home">DOI: 10.1056/NEJMoa1716435</a>.</figcaption></figure><p>You might be wondering “where are the p-values?”. First, can’t you just appreciate the estimates and confidence intervals? Geez. Second:</p><blockquote>It is too early to evaluate our statistical hypothesis of survival at 24 months, because only 20 of the 31 patients at dose level −1 were treated with PVSRIPO more than 24 months before the data-cutoff date of March 20, 2018…the analyses presented in this article are descriptive and do not include formal statistical tests</blockquote><p>Check back for what promises to be a very interesting analysis. By my reading of Table S4 in the Supplement, 9 of the 31 patients who received dose level −1 were still alive as of the data-cutoff date (29%).</p><h3>My takeaway</h3><p>This study represents a potential breakthrough in the treatment of glioblastoma, a ray of hope against a disease that has proven very difficult to treat. A 3-year survival rate of 21% is a significant accomplishment.</p><p>Of course 21% is not 100%—44 patients have died. Not everyone responds to the treatment, but those who do have the potential to watch as this treatment goes through Phase 2, 3, and hopefully Phase 4 trials one day.</p><p>But the science is just getting started. In the Phase 2 trial that is underway at Duke (<a href="https://clinicaltrials.gov/ct2/show/NCT02986178">NCT02986178</a>), 62 patients will be randomly assigned to receive PVSRIPO alone or in combination with lomustine, a chemotherapy drug. The estimated study completion date is 2023, so you’ll have a while to wait for the next installment.</p><h3>Joining the Phase 2 study</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kjaKoB6W2o0Z8nog3NfYuw.png" /></figure><p><em>“To learn more about [the Phase 2 study], you or your doctor may contact the study research staff using the contact information provided by the sponsor [</em><a href="/service/mailto:dukebrain1%40dm.duke.edu?subject=NCT02986178,%20Pro00077024,%20PVSRIPO%20With/Without%20Lomustine"><em>dukebrain1@dm.duke.edu</em></a><em>].”</em></p><h3>Notes</h3><p>[1] Defined in the paper as “as any toxic effects of grade 3 or 4 that were not reversible within 2 weeks or any death that was considered by the investigators to be related to treatment”.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5bef1ebc7f41" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Get to know our pop-up UX lab in Nairobi]]></title>
            <link>https://uxplanet.org/healthymoms-a4c573fbfb7f?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4c573fbfb7f</guid>
            <category><![CDATA[depression]]></category>
            <category><![CDATA[design-thinking]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[chatbots]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Mon, 25 Jun 2018 14:45:10 GMT</pubDate>
            <atom:updated>2018-08-14T14:21:02.108Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ptDHv9qwOr3mSf1Hwhavzw.jpeg" /><figcaption>User testing session in Nairobi. (Photo Credit: Lulla Kiwinda)</figcaption></figure><h4>Creating a mental health chatbot to treat perinatal depression in Kenya (Part 2): User-centered design and testing</h4><p><em>This short series documents our efforts to develop and test a SMS-based treatment for perinatal depression that uses X2AI’s artificial intelligence platform to deliver an adapted version of the Thinking Healthy program. Visit </em><a href="https://healthymoms.app/"><em>healthymoms.app</em></a><em> to learn more. </em>| <a href="https://medium.com/@ericpgreen/healthymoms-36975a4b8c1b">Part 1</a> | Part 2 |</p><h3>Adapting the Thinking Healthy program for SMS</h3><p>In <a href="https://medium.com/@ericpgreen/healthymoms-36975a4b8c1b">Part 1</a>, we made the case for why we chose to adapt the <a href="http://www.who.int/mental_health/maternal-child/thinking_healthy/en/"><em>Thinking Healthy</em></a> program for <a href="http://x2ai.com/">X2AI’s mental health chatbot</a> called Tess. Here we share details about our initial testing process and a bit about what we learned.</p><h4>Thinking Healthy</h4><p>The <em>Thinking Healthy</em> program consists of 15 sessions grouped into 5 modules that span late-stage pregnancy through 10 months postpartum.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5LNQAHSx8hVPl-d-Wo2AdA.png" /><figcaption>Thinking Healthy session schedule.</figcaption></figure><p>Each <em>Thinking Healthy</em> session uses cognitive behavior therapy techniques and follows a similar format:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KyYe5XOn2qDiWnhM-eVxow.png" /><figcaption>Thinking Healthy session format.</figcaption></figure><h4>Creating SMS content</h4><p>The <em>Thinking Healthy</em> manual is a resource that guides community health workers who deliver the treatment to depressed women in one-on-one sessions. As such, the manual instructs the counselor, not the woman. As you can see in this example, the instructions are high-level steps and suggestions rather than scripted prompts, e.g., “Summarize the concepts”, “Go through the Mood Chart”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DK_1fbAijJkxhZzn7o5TsA.png" /><figcaption>Example session text from the Thinking Healthy manual.</figcaption></figure><p>This means that we had our work cut out for us to turn each session into a set of 120-character SMS conversation prompts.¹ <strong>For instance, we had to turn a prompt like this…</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gWddOfo5Nxvez7p31nzyBg.png" /><figcaption>Thinking Healthy prompt from Session 1.1, Task 3, Step 1.</figcaption></figure><p><strong>…into a conversation like this…</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/1*1YgkbJPph5RboDnIJPZB4g.png" /></figure><p>In this initial phase we tested decision tree-based conversations to let us focus on content and flow. When we move to X2’s platform, we’ll be able to take advantage of their natural language processing engine to begin developing more dynamic conversations.</p><p>Even with this highly structured approach, it was amazing to watch women’s responses. We explained that Tess is not a person and avoided creating this perception in our messages, but users seemed to partially suspend belief and treat Tess like a person in small ways. If you’ve ever thanked Alexa or Siri, then you can imagine what I mean.</p><h4>Creating a companion journal for women</h4><p>We’re designing this app with the lowest common denominator in mind: basic mobile phones with only SMS.</p><p>Services like Woebot and Tess (X2’s customizable chatbot)—see <a href="https://medium.com/@ericpgreen/healthymoms-36975a4b8c1b">Part 1</a>—offer great multimedia experiences, but they require users to own smartphones and have mobile data subscriptions. While the smartphone market in Kenya is growing rapidly, <a href="https://www.businessdailyafrica.com/corporate/Over-60pc-Kenyans-smartphones-shows-study/539550-3899256-iyu4xcz/index.html">at least 2 out of 5 people</a> would be unable to access our app today if we required smartphones and mobile data. So we’re starting with SMS.²</p><p>In order to make our SMS app more engaging, we’re developing a paper journal that women can use to document their journey. We’re currently working with an artist to create great illustrations, so we’ll give a proper introduction to this work at a later date. Our aim is to release a journal with a Creative Commons license to promote future tailoring and adaptation.</p><h3>User testing in Nairobi</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dKeNxS0Q0khcPUwM962E9w.png" /></figure><p>In May and June, the Duke team made a trip to Nairobi to begin an initial round of user testing. Our objective was to test the content and SMS experience with a group of Kenyan women.</p><p><strong>For this work we partnered with a fantastic maternity hospital in Nairobi called </strong><a href="https://jacarandahealth.org/"><strong>Jacaranda Health</strong></a><strong>.</strong> Founder and Executive Director <a href="https://globalhealth.duke.edu/people/faculty/pearson-nicholas">Nick Pearson</a> is an affiliate of my home institution at Duke University, and we share the same vision for expanding access to care for perinatal depression. Jacaranda’s emphasis on innovation and research made them a natural partner.</p><h4>The setup</h4><p>We set up our user testing lab in a meeting space next to Jacaranda’s building on the outskirts of Nairobi. In setting up our space, <a href="https://medium.com/product-labs/build-your-usability-lab-on-a-budget-b9bcd09ef23e">we took inspiration from Heewon Choi</a>, product design manager for Pivotal Labs, Tokyo. Testing materials included:</p><ul><li>Two laptops (MacBook Pro)</li><li>Two iPhones</li><li>A <a href="https://www.amazon.com/Lamicall-VC-S-US-B-Cell-Phone-Stand/dp/B01HPI5AM2/ref=sr_1_1?ie=UTF8&amp;qid=1529661963&amp;sr=8-1&amp;keywords=Lamicall+VC-S-US-B+Cell+Phone+Stand%2C+S1+Dock+%3A+Cradle%2C+Holder%2C+Stand+For+Switch%2C+all+Android+Smartphone%2C+iPhone+6+6s+7+8+X+Plus+5+5s+5c+charging%2C+Accessories+Desk+-+Black">desktop stand</a> for the prototype phone</li><li>A <a href="https://www.amazon.com/UTEBIT-Rotatable-Monopod-Smartphone-Photography/dp/B06Y46PJQH/ref=sr_1_1?s=wireless&amp;ie=UTF8&amp;qid=1529662070&amp;sr=8-1&amp;keywords=UTEBIT+Copy+Stand+Macro+Rotatable+Camera+Table+Top+Monopod+Stand+with+360+Degree+Swivel+Ball+Head+Mount+for+DSLR+Smartphone+Video+Photography+Product">vertical stand</a> with a <a href="https://www.amazon.com/Square-Jellyfish-Jelly-Tripod-Mount/dp/B01M11QFSK/ref=sr_1_1?s=wireless&amp;ie=UTF8&amp;qid=1529662032&amp;sr=1-1&amp;keywords=Square+Jellyfish+Jelly+Grip+Tripod+Mount+for+Smart+Phones+-for+all+models+of+iPhone%2C+Samsung%2C+Google+etc.">tripod mount for smartphones</a> to capture overhead recordings</li><li>A digital audio recorder</li><li>An <a href="https://www.amazon.com/Elements-Portable-External-Drive-WDBU6Y0020BBK-WESN/dp/B06W55K9N6/ref=sr_1_1?s=wireless&amp;ie=UTF8&amp;qid=1529661990&amp;sr=1-1&amp;keywords=WD+2TB+Elements+Portable+External+Hard+Drive+-+USB+3.0+-+WDBU6Y0020BBK-WESN">external hard drive</a></li><li>A <a href="http://www.airsquirrels.com/reflector/">Reflector 3</a> license</li><li>Wifi</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-dVoTdofbo-aEe4i5tldCw.jpeg" /><figcaption>Our user testing lab setup. (Photo credit: Lulla Kiwinda)</figcaption></figure><p>For each session, we seated the participant and a facilitator in front of the testing phone. A vertical camera stand held the second smartphone to record overhead video of the user’s interaction with the prototype. We positioned a MacBook Pro opposite the user and facilitator to record video of the testing environment using Quicktime player. As you can see in the photo, we dimmed the laptop screen brightness so that neither the participant nor facilitator would be distracted by the video recording.</p><p>We used the Reflector 3 software to mirror the screens of the prototype phone, the overhead video, and the laptop video via Apple AirPlay to a second MacBook Pro that recorded all three screens. A digital recorder captured audio from the testing session. This short video shows the 3-screen setup:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F276149038%3Fapp_id%3D122963&amp;dntp=1&amp;url=https%3A%2F%2Fvimeo.com%2F276149038&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F708607640_1280.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=vimeo" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/fbb76fc4594b65baf9bc3398788908de/href">https://medium.com/media/fbb76fc4594b65baf9bc3398788908de/href</a></iframe><p>One member of the Duke team sat next to the user and facilitated the session, while a second team member sat in the corner of the room, observed the session from the laptop running Reflector 3, and took notes. Our superb assistant Lilian from the Jacaranda Health team helped everything to run smoothly and contributed great insights.</p><p>In total, we conducted 23 testing sessions over 3 weeks with 10 women recruited from Jacaranda’s maternity hospital. We did not seek out women with a history of depression for this initial testing. Rather, we asked the staff at Jacaranda to recommend several pregnant women and new moms who might enjoy trying out a new service and giving detailed feedback. We ended up with a great group of women who ranged in age from 27 to 35 (<em>M</em>=30.3, <em>SD</em>=2.8). Nearly all of the women attended some post-secondary education, and all could speak and read English. Two women were pregnant, and the rest had kids under two years of age. This group does not reflect the broader demographics of this age group in Kenya, but that was not our goal at this early stage.</p><h4>SMS sandbox</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RZDR9ASmMW4grT8zqPe5Ag.png" /></figure><p>While X2AI worked on setting up our integration with Tess, we created a mock AI environment using a program called <a href="https://textit.in/flow/">TextIt</a>, a mobile messaging platform built in Rwanda. TextIt makes it easy to visually build SMS workflows like the one you see here. There is no setup cost or monthly fee, but testing in TextIt costs $0.02 USD for each incoming and outgoing message. You could build a production-ready app in TextIt, but you would pay twice for every message.</p><p>This is because TextIt requires a channel for sending and receiving messages to end users. We connected our TextIt account to a Kenya-based SMS provider called <a href="https://africastalking.com/">Africa’s Talking</a>. For about $175 USD per month, we rent a dedicated 5-digit shortcode that lets us send and receive messages in Kenya.³ Every incoming and outgoing SMS costs us about $0.08 USD currently, but will drop to $0.04 at scale. Our shortcode is toll-free, so users can message us at no cost—we pick up the check.</p><h4>What we learned</h4><p>The biggest and most important takeaway from our user testing sessions is that women liked talking with our Tess mock up. Their enthusiasm for the idea really motivated us to learn about how to make it better.</p><p>To that end, we spent a few hundred hours revising the content and approach. Cut down on the number of consecutive outbound messages here, change the phrasing there, give up on that joke that fell flat.⁴</p><h4>What comes next</h4><p>We’re partnering with Dr. Edith Kwobah (psychiatrist at Moi Teaching and Referral Hospital in Eldoret), Dr. Christine Musyimi of the <a href="http://www.africamentalhealthfoundation.org/">Africa Mental Health Foundation</a> in Nairobi, and two medical doctors completing their training in psychiatry. This fantastic team will help to further revise the content and prepare for a pilot trial with 20 depressed women recruited from Jacaranda Health’s maternity hospital.</p><p>The team at X2 is finalizing our Tess SMS integration, and we’ll complete one more round of testing in the run up to the launch of the trial. Check back for more details about our trial design and analysis plan…and our new name! Tess is known as Karim in Lebanon, Sarah on Facebook, and soon to be ?? in Kenya.</p><h3>FAQs</h3><h4>Why aren’t you creating a Swahili version?</h4><p>This is usually one of the first questions we get when we talk about our work. The short answer is that we will, but you have to start somewhere. Tess is capable of delivering services in any language, but we’re starting with English because it’s easier to demonstrate proof of concept this way. This is possible in Kenya because English is an official language and is the language of instruction in school from grade (standard) 4.</p><p>With universal primary education in place today and a plan to implement universal secondary education in 2019, future cohorts of new mothers will be much more likely to speak English than their mothers were a generation ago. We’ll soon witness a convergence of developments in AI and natural language processing that supports Swahili, high prevalence of smartphones and mobile internet, and the expansion of human capital. Automated conversational agents like Tess are emerging at a perfect moment.</p><h4>Can Tess do more than deliver the Thinking Healthy curriculum?</h4><p>Yes! X2 has a library of hundreds of conversation modules that we will be able to adapt and integrate into the <em>Thinking Healthy</em> app. Many of the available conversation modules are client-initiated, meaning a user can just say “Hey Tess” and begin a conversation. So as we gain more experience in Kenya, the <em>Thinking Healthy</em> app will deliver a structured curriculum <em>and</em> support clients who want to talk in between sessions.</p><h4>Who is funding this work?</h4><p>We have some fantastic partners who make this work possible. The <a href="https://globalhealth.duke.edu/">Duke Global Health Institute</a> and its <a href="https://globalhealth.duke.edu/topics/mental-health">Global Mental Health Working Group</a> are providing pilot funding, and several Duke programs such as <a href="https://bassconnections.duke.edu/">Bass Connections</a> are supporting student involvement. In addition, X2 is providing its Tess administration panel pro bono as part of the company’s Social Impact efforts. I am such a fan of X2’s work that I’ve joined their Ethics Board as an unpaid advisor.</p><h3>Notes</h3><p>[1] SMS allows a maximum of 160 characters, but we aimed to create messages under 120 characters to accommodate translations to other languages like Swahili that tend to use more characters to say the same thing in English. Other languages are even more restrictive. In unrelated work in Nepal, we learned the hard way that the Devanagari script<strong> </strong>requires about 2 SMS characters to encode each script character. In that work we capped English messages to 80 characters.</p><p>[2] Eventually, it will make sense to move away from SMS. When we do, there will be a significant cost savings since we will not have to pay to send and receive each message. Fortunately for us, X2’s customizable chatbot Tess works over SMS in addition to Facebook Messenger.</p><p>[3] A shared shortcode is less expensive but requires every incoming message to have a keyword for proper routing. For instance, with a shared shortcode, every time a user replies to your message they would need to type something like “DGHI” and then type their reply. Every. Time. Since our shortcode is dedicated, messages sent to us don’t require a keyword. Currently we only have this shortcode setup on <a href="https://www.safaricom.co.ke/">Safaricom</a>, Kenya’s largest telecom provider.</p><p>[4] After asking the user what she liked to do to relax, Tess said she liked to just unplug. #dadjokes</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2Fb3e808%3Fas_embed%3Dtrue&amp;dntp=1&amp;display_name=Upscribe&amp;url=https%3A%2F%2Fupscri.be%2Fb3e808%2F&amp;image=http%3A%2F%2Fapi.screenshotlayer.com%2Fapi%2Fcapture%3Faccess_key%3Dfe59908dad3baab69ffab249a2224b03%26viewport%3D1024x612%26width%3D1000%26url%3Dhttps%253A%252F%252Fupscri.be%252Fb3e808%253Fscreenshot&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/96d08ab34921bdd17986cb5c0396842f/href">https://medium.com/media/96d08ab34921bdd17986cb5c0396842f/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4c573fbfb7f" width="1" height="1" alt=""><hr><p><a href="https://uxplanet.org/healthymoms-a4c573fbfb7f">Get to know our pop-up UX lab in Nairobi</a> was originally published in <a href="https://uxplanet.org/">UX Planet</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Most people who suffer from depression don't get help]]></title>
            <link>https://uxplanet.org/healthymoms-36975a4b8c1b?source=rss-c3dc9bdc0078------2</link>
            <guid isPermaLink="false">https://medium.com/p/36975a4b8c1b</guid>
            <category><![CDATA[motherhood]]></category>
            <category><![CDATA[pregnancy]]></category>
            <category><![CDATA[mental-health]]></category>
            <category><![CDATA[depression]]></category>
            <category><![CDATA[chatbots]]></category>
            <dc:creator><![CDATA[Eric Green]]></dc:creator>
            <pubDate>Mon, 25 Jun 2018 14:44:00 GMT</pubDate>
            <atom:updated>2018-08-14T14:21:57.905Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*flRRG7Y42IBxeS5EAj2MDQ.jpeg" /><figcaption>Image credit: <a href="https://www.flickr.com/photos/untickalock/">untickalock</a> on Flickr. “<a href="https://www.flickr.com/photos/untickalock/26149038/">Help is on the way</a>”.</figcaption></figure><h3>Most people who suffer from depression don’t get help</h3><h4>Creating a mental health chatbot to treat perinatal depression in Kenya (Part 1): Why chatbots?</h4><p><em>This short series documents our efforts to develop and test a SMS-based treatment for perinatal depression that uses X2AI’s artificial intelligence platform to deliver an adapted version of the Thinking Healthy program. Visit </em><a href="https://healthymoms.app/"><em>healthymoms.app</em></a><em> to learn more.</em> | Part 1 | <a href="https://medium.com/@ericpgreen/healthymoms-a4c573fbfb7f">Part 2</a> |</p><h3>Most people who suffer from depression and other common mental disorders don’t get treatment.</h3><p>Often because they can’t. In many communities, there are few (if any) providers to pay, even if you could afford the bill. More than 3 billion of us live in a country where there is only 1 psychiatrist for every 200,000 people.¹ Psychologists are in even shorter supply.</p><p>This is a big deal. Depression is the leading cause of disability worldwide, and neuropsychiatric disorders more broadly account for nearly 1 in 5 years lived with disability globally.² More than 300 million people worldwide are thought to suffer from depression or anxiety.³</p><h4>Depression affects everyone, but pregnant women and new moms are a special group.</h4><p>When a woman experiences mood and anxiety symptoms during pregnancy or the year following delivery, we call it <strong>perinatal depression</strong>. Estimates of prevalence vary by study and setting, but generally perinatal depression is thought to impact about 1 in 10 women in high-income countries and about 2 in 10 women everywhere else.⁴ Untreated depression carries with it increased risk for suicide and a host of bad outcomes for kids. And as the following chart reaffirms, most cases go untreated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/933/1*RIC0NUJCAWXZruDA_4Q-YQ.png" /><figcaption>Drop-off in care among women with perinatal depression in mostly high-income(!) settings. (Source: Adapted from Gavin et al. (2015). Is population-based identification of perinatal depression and anxiety desirable? In Milgrom &amp; Gemmill (Eds.) Identifying Perinatal Depression and Anxiety. Wiley: Sussex.)</figcaption></figure><p>So here we have a situation of immense mental health needs and, in many places, almost no way to meet these needs. What is the solution? Training more providers and lowering social and economic barriers to treatment would be a good start, but it won’t be enough.</p><h3>One promising idea is to train ordinary folks to deliver basic care.</h3><p>It’s called <a href="http://www.who.int/mental_health/mhgap/en/"><strong>task shifting</strong> or <strong>task sharing</strong></a>. For instance, psychiatrist Dixon Chibanda and his team trained grandmothers to sit on park benches in Zimbabwe—<a href="https://www.npr.org/sections/goatsandsoda/2017/01/10/508588401/the-friendship-bench-can-help-chase-the-blues-away"><em>Friendship Benches</em></a>—and listen to people suffering from anxiety and depression. The 6-month results were impressive.⁵</p><p>Or look at program called <a href="http://www.who.int/mental_health/maternal-child/thinking_healthy/en/"><strong><em>Thinking Healthy</em></strong></a> developed in Pakistan by psychiatrist Atif Rahman and colleagues to treat perinatal depression. In this program, community health workers called <a href="http://www.who.int/workforcealliance/knowledge/resources/casestudy_pakistan/en/"><em>Lady Health Workers</em></a><em>—</em>women with high school educations and basic training in prevention and health promotion—were trained to deliver a psychosocial intervention based on the principles and techniques of <a href="https://en.wikipedia.org/wiki/Cognitive_behavioral_therapy"><strong>cognitive behavior therapy</strong></a>. CBT is a common and effective approach for treating common mental disorders worldwide. Rahman reported that the <em>Thinking Healthy</em> program halved the rate of depression 6-months post-intervention in a randomized controlled trial,⁶ though a follow-up study conducted about 7 years later suggests that the effects might be short-lived.⁷</p><h3>But how do we go from studies of 100s to programs that reach 1,000,000s?</h3><p>There is no one approach that will close the yawning gap between mental health needs and resources, but I believe technology will play a role in scale up. Here’s why.</p><p>Compared to other great public health implementation challenges that society has tackled in a very off-line way, such as the <a href="https://www.amazon.com/House-Fire-Eradicate-Smallpox-California/dp/0520274474/ref=sr_1_2?ie=UTF8&amp;qid=1529604329&amp;sr=8-2&amp;keywords=house+on+fire">eradication of Small Pox</a>, <strong>mental health treatment is relatively high touch—meaning a lot of repeated contact with the same clients. </strong>For instance, the <em>Thinking Healthy</em> program has a Lady Health Worker visit a single client at her home 15 times. At scale, this would be harder in many ways than delivering one-off vaccinations (which are already hard to do!). You have to deal with scheduling, transportation, supervision, client stigma, etc. So just as technology helps all complex systems operate more efficiently and effectively, it seems reasonable to assume that technology can help scale up traditional and task-sharing models of care.</p><h3>Is there a bigger role for technology in mental health treatment? (tldr: yes!)</h3><p>Some people will tell you that we are on the cusp of some truly amazing advances in hardware, software, and data science that will change how medicine is practiced. That artificial intelligence will usher in a new era in health care.</p><h3>NowThis on Twitter</h3><p>The Boston Dynamics robot can now open doors. Great. https://t.co/xdNYM0YqXx</p><p>Other people will tell you:</p><h3>Dr Julia Shaw on Twitter</h3><p>Artificial intelligence struggles to tell difference between fried chicken and Labradoodles. I CAN&#39;T GET ENOUGH OF THESE AI PROBLEMS.</p><p>My view is that it would be foolish to claim AI will solve every problem, but dismissing the potential of AI as empty hype is misguided. AI <em>will</em> impact the way we detect and treat many illnesses.</p><p>But what should we think about mental health chatbots?</p><h3>“The chatbot will see you now.”</h3><p><a href="https://www.newyorker.com/tech/elements/the-chatbot-will-see-you-now">This headline</a> is used so often these days that you might think bots are writing the news. But if you can get beyond the headline hype, you’ll find that there is a lot of interesting AI work underway with the potential to create a completely new channel for mental health treatment.⁸</p><p>For instance there is <a href="https://woebot.io/"><strong>Woebot</strong></a>, an automated conversational agent that will talk with you via Facebook Messenger or its apps for iOS and Android. Woebot made the news recently when AI pioneer <a href="https://medium.com/@andrewng/woebot-ai-for-mental-health-d0e8632b82ba">Andrew Ng announced he joined Woebot’s Board of Directors</a>, a move that should give everyone working on AI+mental health a boost of confidence.</p><h3>Woebot on Twitter</h3><p>Hello world! I&#39;m officially available to talk if you need me. Come find me at https://t.co/QnPABvItd4 -- I can&#39;t wait to meet you!</p><p>There is also <strong>Tess</strong>, a creation of Michiel Rauws and the team at <a href="http://x2ai.com/">X2AI</a>. Tess is a customizable chatbot that’s<strong> </strong>available 24/7 to talk with you via Facebook Messenger or SMS.</p><h3>X2AI on Twitter</h3><p>AI helps Syrian refugees through trauma https://t.co/zSheyU9WXL via @Reuters</p><p>Both chatbots are designed to deliver CBT-inspired mental health support, and both are developing an evidence base. In two different studies with college students in the US, Woebot and Tess demonstrated similar effects on symptoms of depression on the order of 10–20% reductions.⁹ This is encouraging, especially when you consider the low cost to deliver such services. <strong>Even with modest impacts, the cost-effectiveness of these digital tools could be quite high relative to traditional or task-shifting models.</strong></p><p>So my prior is that chatbots make a lot of sense for mental health treatment, especially in settings where you share your neighborhood psychiatrist with 199,999 other people.</p><h3>But could chatbots really work in Kenya?</h3><p>I had been nibbling around this question for a few years before I really even knew what a chatbot was. In the process of trying to develop a <a href="http://instedd.org/news-media/press-releases/saving-lives-at-birth-grand-challenge/">screening and referral system for maternal and newborn health</a> a few years back, we observed that Kenyan women reported more depressive symptoms to an automated voice line than they reported to our human study assistants. In a not-so-shocking result, <strong>women seemed very comfortable discussing private issues in private. </strong>This led us to think about how we could use technology to improve screening and treatment for perinatal depression.</p><h4>Start with screening</h4><p>If you glance again at the chart presented earlier, you’ll see the basic fact that was smacking us in the face at the time: fewer than 50% of cases of perinatal depression are ever recognized clinically, meaning that most women with symptoms never get screened by a professional. This number comes from high-income settings, so the situation is likely a lot worse in places like Kenya.</p><p>One reason for a lack of screening is a lack of validated tools to detect depression. Most of today’s commonly used instruments were developed in high-income settings, and research shows that simply running items through Google Translate does not guarantee their validity for measuring depression. It takes work to demonstrate that a questionnaire can discriminate between cases and non-cases and thus have value as a screening instrument, so that’s where we started—with <a href="https://github.com/ericpgreen/pdeps">a validation study</a>.</p><p>But as many people will tell you, screening without access to treatment does not improve outcomes. This is why most countries do not have policies of universal screening for perinatal depression.¹⁰ There are few options for treatment. We wanted to help change this for women in Kenya and beyond.</p><h4>Next, find a program and a platform to adapt</h4><p>We looked at the early success of <em>Thinking Healthy</em> and thought that this CBT-based program could be a natural fit for Kenya. In 2015, the WHO made a general version of the curriculum and incorporated it into <a href="http://www.who.int/mental_health/maternal-child/thinking_healthy/en/">the mhGAP evidence-based toolkit</a> to promote wide-scale adaptation and use.</p><p><em>Thinking Healthy</em> also seemed like a natural fit for X2’s Tess platform. Tess can work over SMS¹¹ and is built for CBT interventions like <em>Thinking Healthy</em>. X2 was already working to expand access to treatment around the world, and the company was not afraid of hard challenges like <a href="https://www.theguardian.com/technology/2016/mar/22/karim-the-ai-delivers-psychological-support-to-syrian-refugees">adapting the platform for Syrian refugees in Lebanon</a>. This made our choice simple.</p><p>So we reached out to Michiel at X2AI…</p><p>| <a href="https://medium.com/@ericpgreen/healthymoms-a4c573fbfb7f"><strong>Part 2 : User-centered design and testing</strong></a> |</p><h3>Notes</h3><p>[1] According to the <a href="http://www.who.int/mental_health/action_plan_2013/en/"><em>WHO Mental Health Action Plan: 2013–2020</em></a>.</p><p>[2] Based on <a href="http://www.who.int/healthinfo/global_burden_disease/estimates/en/index1.html">2015 data from the World Health Organization</a>.</p><p>[3] See <a href="http://www.who.int/mental_health/management/depression/prevalence_global_health_estimates/en/">this 2017 report</a> from the WHO for more details.</p><p>[4] See this 2016 review from Gelaye et al. published in the <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5155709/"><em>The Lancet Psychiatry</em></a>.</p><p>[5] See the results of a cluster randomized trial <a href="https://jamanetwork.com/journals/jama/fullarticle/2594719">published in <em>JAMA</em> in 2016</a>.</p><p>[6] Short-term trial results published in <a href="https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(08)61400-2/abstract"><em>The Lancet</em></a> in 2008.</p><p>[7] A <a href="https://www.thelancet.com/journals/lanpsy/article/PIIS2215-0366(15)00109-1/supplemental?code=lancet-site">follow-up study conducted about 7 years later</a> suggested that the effects might be short-lived.</p><p>[8] This is not hype unless I state that chatbots will replace doctors.</p><p>[9] The <a href="https://mental.jmir.org/2017/2/e19/">Woebot trial</a> reported a decrease in PHQ-9 (<a href="https://www.phqscreeners.com/sites/g/files/g10016261/f/201412/PHQ-9_English.pdf">pdf</a>) scores from 13.67 among the control group to 11.14 among the treatment group, a reduction of 18.5%. Using the same measure of depression, the Tess trial found a decrease of 13% in symptoms of depression, along with an 18% reduction in symptoms of anxiety (<a href="http://x2ai.com/">X2AI website</a>).</p><p>[10] Universal screening for perinatal depression is now recommended by the <a href="https://www.ncbi.nlm.nih.gov/pubmed/26813211">US Preventive Services Task Force</a> (2016) and the <a href="https://www.ncbi.nlm.nih.gov/pubmed/25932866">American College of Obstetricians and Gynecologists</a> (2015).</p><p>[11] SMS is a great choice in Kenya where mobile phone ownership is sky-high, but smartphones with apps like Facebook Messenger are not yet the norm in all parts of the country.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fupscri.be%2Fb3e808%3Fas_embed%3Dtrue&amp;dntp=1&amp;display_name=Upscribe&amp;url=https%3A%2F%2Fupscri.be%2Fb3e808%2F&amp;image=http%3A%2F%2Fapi.screenshotlayer.com%2Fapi%2Fcapture%3Faccess_key%3Dfe59908dad3baab69ffab249a2224b03%26viewport%3D1024x612%26width%3D1000%26url%3Dhttps%253A%252F%252Fupscri.be%252Fb3e808%253Fscreenshot&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=upscri" width="800" height="400" frameborder="0" scrolling="no"><a href="https://medium.com/media/96d08ab34921bdd17986cb5c0396842f/href">https://medium.com/media/96d08ab34921bdd17986cb5c0396842f/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=36975a4b8c1b" width="1" height="1" alt=""><hr><p><a href="https://uxplanet.org/healthymoms-36975a4b8c1b">Most people who suffer from depression don&#39;t get help</a> was originally published in <a href="https://uxplanet.org/">UX Planet</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>