<?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 Angel on Medium]]></title>
        <description><![CDATA[Stories by Angel on Medium]]></description>
        <link>https://medium.com/@vossough?source=rss-3f8d70cd668b------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*_oJCMOt7N31SOlDJVtdsXQ.jpeg</url>
            <title>Stories by Angel on Medium</title>
            <link>https://medium.com/@vossough?source=rss-3f8d70cd668b------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 13:25:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@vossough/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[COVID-19 Prediction Modeling: Analyzing Risk Factors and ICU & Ventilator Exhaustion by State]]></title>
            <link>https://medium.com/@vossough/covid-19-prediction-modeling-analyzing-risk-factors-and-icu-ventilator-exhaustion-by-state-20c9c37d22ee?source=rss-3f8d70cd668b------2</link>
            <guid isPermaLink="false">https://medium.com/p/20c9c37d22ee</guid>
            <dc:creator><![CDATA[Angel]]></dc:creator>
            <pubDate>Wed, 22 Apr 2020 18:38:06 GMT</pubDate>
            <atom:updated>2020-04-22T18:38:06.959Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*H1RsPnp9ZT8_jOlLkG2bSQ.jpeg" /></figure><p><strong>Authors:</strong></p><p>Angel Vossough, Prakhar Maini, Arsalan Vossough</p><p><em>UC Berkeley, Master in Data Science program, W200 project, Dr. Gerald Benoit</em></p><p><strong><em>Context</em></strong><em>: The CoronaVirus pandemic is a fast moving and highly dynamic situation, with volatile data. Given the scope and timeframe of this project, we’ve chosen to act with data available as of </em><strong><em>April 15, 2020</em></strong><em>. In that respect, our predictions and analysis reflect trajectories that could/would have happened without any changes in government policy &amp; enforcement or societal behavior.</em></p><h3>Abstract</h3><p>This analysis uses a combination of available data sources to project ventilator exhaustion risk by state. We further analyzed the impact of risk-factors like obesity rates, smoking rate, and population density. With time we expect that the underlying risk factors of the general population will start playing a bigger role. In this regard, we expect that there may be <em>another wave of severe-critical cases that hit the most susceptible, and create the greatest demand for critical ICU and ventilator resources</em>.</p><p>We were able to predict the short term infection rates and death rates to a with relatively high accuracy using a simple exponential smoothing model. This indicates to us although it is very hard to conclusively forecast the overall impact of this problem, <em>we can make short term judgements (1–2 weeks) with high degree of precision to aid resource allocation decisions</em>.</p><p>From the <em>infection modeling</em> perspective, we are in a <em>linear zone in most of the states in the US</em>. Deaths are still in the exponential zone at this point. Optimistically, you could conclude that the current wave of infections seems to be coming under control with measures like shelter-in-place. More pessimistically, you could conclude that our testing capacity (or usage) is linear, which is artificially capping and substantially understating the actual number of COVID-19 positive cases.</p><p>Our modeling suggests that the <em>midwestern and southern states — </em>Georgia, North and Sound Dakota, Florida and Louisiana — <em>are at highest risk of overwhelming their medical system resources in the next 30 to 45 days</em>.</p><p>We collected data on high-risk factors and concluded that the primary risk factor for COVID-19 infection (confirmed cases) was Density; the primary risk factors for death from COVID-19 are Age and Obesity.</p><h3>Introduction</h3><p>Due at least in-part to lack of systemic federal response and varying containment policies and implementation by-state, the United States now leads the world in confirmed COVID-19 cases, surpassing China and Italy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*18oBVz5-p6NWJKSzgZv6BA.png" /></figure><p>In our project, we intend to:</p><ol><li>Analyze the current status of COVID-19 spread in the US and look at the impact, in terms of confirmed cases and deaths per state</li><li>Build a projection model to forecast state-by-state cumulative confirmed cases and deaths</li><li>Analyze demographics by geography to understand the most vulnerable areas, in terms of population risk/susceptibility</li><li>Finally, we want to understand the distribution of medical resources (ICU beds and ventilators) across US states. Our goal is to come up with <strong>a risk factor indicating the risk of the medical system becoming overwhelmed in each state</strong>.</li></ol><p>We are using the latest data to forecast cumulative confirmed cases and deaths that can give us some confidence in our forecasts. We’ve done extensive literature surveys to understand key parameters such as Intensive Care Unit (“ICU”) admittance rates and ventilator availability. Overall, with proper due diligence, we want to create a defensible model for tracking medical risk across states, and in the best-case scenario,<em> it could be used by the federal government to prioritize cross-state resource allocation efforts</em>.</p><h3>Flattening the curve</h3><p>Models have estimated that up to 40–70% of the US population will be infected over the next 12–18 months unless decisive actions are taken to slow the spread of the pandemic. These collective actions — social distancing, contact tracing, “shelter-in-place”, closing non-essential businesses, travel restrictions, and mask-wearing — have been proven in other countries to help reduce the rate of growth. Without them, the great risk is that the healthcare system will be overwhelmed with a surge in demand much greater than current capacity.</p><p>The current crisis in New York City and New Orleans demonstrate cases where we don’t have sufficient hospital beds, ICU beds, and ventilators when COVID-19 cases ramp quickly. The implication is that the COVID-19 death rate can increase drastically since patients who could have been saved are lost due to insufficient resources, and triage will be needed where doctors will be choosing who to save with available resources, and who will not get the resources (and therefore, who dies).</p><p><em>Figure 1: FLATTENING THE CURVE</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/620/0*q5naZ4ZO50iVPpDa" /></figure><h3>How the US compares to other countries</h3><p>Per the chart below (Figure 2), the United States has the highest exponential growth of COVID-19 (more than 640,000 confirmed cases and over 30,000 deaths) compared to any other country in the world, with a doubling time of ~4 days. In comparison, Singapore and Japan have been able to slow the spread of the virus and flatten their curves. Our own model (Figure 3) suggests the US will exceed 1 million confirmed cases by April 27th, 2020 (with likely many more cases unconfirmed, due to insufficient availability of testing).</p><p><em>Figure 2: Worldwide spread status</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qtx_Nopj_7UpRemB" /></figure><p><em>Figure 3: US COVID-19 Spread — Holt-Winters Projection**</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/0*SnE1AooBMRHweaoi" /></figure><p><em>** Historicals as of April 15, 2020</em></p><h3>Research questions</h3><ol><li>What is the current situation (as of April 15th) for the number of tests, cases, hospitalization, and deaths, by state?</li><li>How many ICU beds are there, by state?</li><li>What is the estimated number of ventilators, by state?</li><li>When will COVID-19 driven demand reach critical mass that drives ICU bed resource exhaustion, by state? and ventilator resource exhaustion, by state?</li></ol><h3>DataSets</h3><p>We are looking at a comprehensive review of the landscape in this project. Hence several datasets are critical for the synthesis of information:</p><h3>COVID-19 time-series data</h3><p>This is the first crucial piece of information. There are many well maintained data sources available that track the spread of the virus globally and in the US. However, in choosing our source, we wanted to align our dataset closely with the official US numbers for both confirmed cases and deaths. For this purpose, we source our data primarily from the <a href="https://github.com/CSSEGISandData/COVID-19/tree/master/csse_covid_19_data/csse_covid_19_daily_reports">CSSE</a> dataset that is maintained by John Hopkins University and the Centers for Disease Control (“<a href="https://www.cdc.gov/">CDC</a>”). In addition we used the <a href="https://www.worldometers.info/coronavirus/country/us/">worldmeters.info</a> website to cross-reference the data points and ensure data sanctity. The data set available on CSSE’s website contains global information, so we created our own <a href="https://drive.google.com/a/berkeley.edu/file/d/12xFqehv7UZSvhjfIvEKLg_7Xzp0GjIaZ/view?usp=sharing">dataset</a> where we extracted and created a subset with state-by-state time-series data for confirmed cases and deaths.</p><h3>US metadata</h3><p>As we want to delve deep into the vulnerability of the US, we collected basic demographic metadata for the US at the state level. We used <a href="https://www.census.gov/data/tables/time-series/demo/popest/2010s-state-total.html">census</a> data and <a href="https://worldpopulationreview.com/states/">external data sources</a> to understand the distribution of population in general and of elderly population across the US by-state.</p><h3>Medical statistics</h3><p>Finally, we need to combine the COVID-19 specific data and US demographic data with hospitalization rates, ICU beds, and ventilator numbers by-state. For capturing the ICU bed distribution across the US we used <a href="https://khn.org/news/as-coronavirus-spreads-widely-millions-of-older-americans-live-in-counties-with-no-icu-beds/">a dataset from Kaiser Health</a>. To understand the hospitalization rates across states, we used a dataset from <a href="https://docs.google.com/spreadsheets/u/2/d/e/2PACX-1vRwAqp96T9sYYq2-i7Tj0pvTf6XVHjDSMIKBdZHXiCGGdNC0ypEU9NbngS8mxea55JuCFuua1MUeOj5/pubhtml#">CovidTracking</a>.</p><h3>Literature survey</h3><p>We did an extensive literature survey to understand the distribution of known risk-factors of COVID-19 (such as obesity, lung disease, diabetes, hypertension, etc.) so that these numbers can inform the risk_factor distribution across the US. We wanted to understand following four parameters at state level:</p><ol><li>Average rate of ventilator availability</li><li>Average hospitalization rate</li><li>Average ICU transfer rate (after hospitalization)</li><li>Average ventilator transfer rate (after ICU transfer)</li></ol><p>Literature survey helped us in creating a well maintained and calibrated dataset. We created the following <strong>prepared datasets</strong> which are being utilized in our analysis:</p><ul><li><a href="https://drive.google.com/a/berkeley.edu/file/d/12xFqehv7UZSvhjfIvEKLg_7Xzp0GjIaZ/view?usp=sharing"><em>COVID-19 projection data</em></a></li><li><a href="https://docs.google.com/spreadsheets/d/10FQcScbf3e8P5pxyQUYp5pYGan0fbB8ubi-ImYy8YZg/edit?usp=sharing"><em>Metadata</em></a></li><li><a href="https://docs.google.com/spreadsheets/d/1HaPyaoYsHTjQh8K_2ojP71Jw3rz05_ZCwoatcJROs20/edit?usp=sharing"><em>Risk-factors</em></a></li><li><a href="https://drive.google.com/a/berkeley.edu/file/d/1MDXcnFAhjxygJwFzNWa-5YEfw4mPwBzR/view?usp=sharing"><em>Risk_factor_Metadata</em></a></li></ul><p>As most of the data that we are working with is self assembled, we provide a full data dictionary at the below location to provide the full understanding.</p><ul><li><a href="https://docs.google.com/spreadsheets/d/1cldwVVfOpc2561ttdZIwdCB-ZHolgLkIHDEQDOHomis/edit?usp=sharing"><em>Data Dictionary</em></a></li></ul><h3>US COVID-19 Spread — Overview</h3><h3>Growth and spread over time</h3><p>For ease of visualization, we divided the states into roughly equal quantiles based on the severity of infection rates as of <strong>April 15, 2020</strong>.</p><p><em>Figure 4: Patient confirmations across US</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/824/0*wiEhiaqx1jISzyJQ" /></figure><p>Similarly, the patient deaths show a similar pattern.</p><p><em>Figure 5: Patient deaths across US</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/0*fln2XMhv79tmh435" /></figure><ul><li>The<strong> top two quartiles</strong> show maximum variation across all other quartiles for both confirmed patients and patient deaths. This is expected due to the presence of outliers in the top quartile.</li><li>The meteoric rise in the overall cases and deaths in the US is driven by the top quartile states. This indicates that the infection is still somewhat <strong><em>localized</em></strong> and there is still potential for massive rise in disease spread.</li></ul><h3>Geographical concentration of infection</h3><p>New York has become the global hotspot for the pandemic now, hence those numbers skew the geographical distribution.</p><p><em>Figure 6: Distribution of COVID-19 (</em><strong><em>confirmed</em></strong><em>) cases (with and without NY) Apr 15, 2020</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/0*xOgRBki0FRV0Vgs8" /></figure><p><em>Figure 7: Distribution of COVID-19 (</em><strong><em>Deaths</em></strong><em>) cases (with and without NY) Apr 15, 2020</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/985/0*8k7coW80JrBL22pC" /></figure><p>From the above historical trends and the status of the infection across US, we can say the following definitively:</p><ul><li>East and west coast states are more heavily hit by the infection than the remaining (at least by April 15, 2020). This might be attributable to the heavy concentration of population, economic activity of these regions, and higher prevalence of international travel.</li><li>The pandemic has transitioned into the midwest and southern US states now. The distribution of deaths paint this picture more clearly.</li></ul><h3>US Demographic &amp; Healthcare facilities — Overview</h3><h3>Demographic distribution</h3><p>The United States is the third-most populous country in the world, with an estimated population of 330 million as of January 28, 2020. The American population almost quadrupled during the 20th century at a growth rate of about 1.3% a year.</p><p>The median age of the total population is 38.2 years; the male median age is 36.9 years; the female median age is 39.5 years. The population is distributed by age as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/876/1*rfNv5wsq7U2b__En7lo3mA.png" /></figure><p><strong>Birth rate</strong>: 12.4 births/1,000 population (2018 est.) Country comparison to the world: 157th</p><p><strong>Death rate</strong>: 8.2 deaths/1,000 population (2018 est.) Country comparison to the world: 86th</p><p><strong>Life expectancy</strong>:</p><ul><li>Male: 77.8 years</li><li>Female: 82.3 years</li></ul><p>in the United States:</p><ol><li>California (Population: 39.7M)</li><li>Texas (Population: 29.1M)</li><li>Florida (Population: 21.6M)</li><li>New York (Population: 19.5M)</li><li>Pennsylvania (Population: 12.8M)</li><li>Illinois (Population: 12.7M)</li><li>Ohio (Population: 11.7M)</li><li>Georgia (Population: 10.6M)</li><li>North Carolina (Population: 10.5M)</li><li>Michigan (Population: 10.0M)</li></ol><p><em>Figure 8: population distribution by states</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*hT3jvWxtZt5A1Syb" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*CqMhI17ls3hOj-wE" /></figure><p><strong>Healthcare facilities distribution</strong></p><p><strong>The ICU Beds: </strong>There are 6,210 total hospitals in the US. California, Texas and Florida have the highest number of hospitals and ICU beds. About 68% of total ICU beds are equipped with mechanical ventilators. However, more than half the counties in America have no intensive care beds, posing a particular danger for more than 7 million people residing there who are age 60 and above ― older patients who face the highest risk of serious illness or death from diseases.</p><p><em>Figure 9: Number of ICU beds per capita.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*MZ7aWdRGBmHY1XKB" /></figure><h3>Ventilator exhaustion risk assessment</h3><h3>Risk factors</h3><p>Multiple risk factors impact the susceptibility of an individual needing special care (e.g., ICU bed, ventilators), including obesity rate, average age, population density, hypertension, and smoking rate in a given state. We collected data for the top 18 most affected states in the US so far and mapped the actual deaths in the states with these factors.</p><ul><li>The most correlated factor turns out to be average population density. This makes sense as COVID-19 is a highly infectious disease with community spread.</li><li>Some of the other risk factors (e.g., obesity, smoking) seem to be less strongly correlated with death rate. This is counterintuitive; a potential explanation is that COVID-19 prevalence is still sufficiently low in the population that individual hospitalizations to date don’t yet strongly correlate with geographic population metadata. With greater prevalence we’d expect this to change over time.</li></ul><p><em>Figure 10: Correlation heatmap of death rates and risk factors across 18 largest US states (excluding New York)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/622/0*lj5U6DASz7yQ2tAE" /></figure><p><em>Figure 11: Pair plots to understand relationships across death rates and risk factors (excluding New York)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/0*xdSTr5e4HwjoPAbL" /></figure><p>New York is a big outlier in terms of number of infections and deaths; adding it produces skewed data representation. Hence for the above analysis, we have excluded NY from consideration set to understand the underlying nature of the data. If we see the top row of this figure, we see that <strong>Deaths have a strong positive correlation with population density</strong> and negative correlation with <strong>Number of ICU beds and Ventilators (per capita)</strong>. Negative correlation exists between deaths and factors like smoking and obesity (which is counterintuitive, and potentially reflects early stage of COVID-19 prevalence, as discussed above).</p><h3>Risk assessment</h3><p>We have set up the following model for analysis of ventilator exhaustion risk:</p><ol><li>We estimate the demand side using our projection model for confirmed COVID-19 cases and other key assumptions</li><li>We estimate the supply of ventilators using our literature survey on ventilator distribution and their average availability</li><li>The risk factor is defined as the <strong>surplus ratio</strong> (i.e., ( supply — demand)/supply)</li></ol><p><em>Figure 12: Schematic model of ventilator risk calculation</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LVlYIyUBA48WtEVLV9CEhw.png" /></figure><p>There are four key risk factors that we are using in our model to compute the risk. The population-level risk factor analysis in the first half of this section informed us on how to pick reasonable numbers for these meta factors.</p><ul><li><strong>Ventilator availability ratio</strong>: To make a simple model, we have considered that only<strong> 70%</strong> of the ventilators available in a state are free to be utilized by the COVID-19 patients. In our literature survey, we came to understand that US hospitals put a lot of focus on efficiency in recent years which resulted in a shrinkage of spare capacity.</li><li><strong>Average hospitalization rate (per state)</strong>: The <a href="https://docs.google.com/spreadsheets/u/2/d/e/2PACX-1vRwAqp96T9sYYq2-i7Tj0pvTf6XVHjDSMIKBdZHXiCGGdNC0ypEU9NbngS8mxea55JuCFuua1MUeOj5/pubhtml#">covidtracking</a> website tracks hospitalizations across the US at the state level. We worked with the assumption that the current cumulative hospitalization to total confirmed cases ratio is a good proxy of the final rate. Hence we used this calculated number directly in our calculations.</li><li><strong>Average ICU transfer rate (per state)</strong>: The ICU transfer rate will depend a lot on the parameters we studied in the earlier section. As per the medical experts, the demographic factors such as age, smoking habits, obesity and pre-existing medical conditions make an individual more likely to be hit harder by the disease. In our model, we used a <strong>20%</strong> transfer rate to ICU per state as the baseline value. Based on the analysis of risk-factors done earlier, we increased this rate to 25% for Florida, Georgia, Massachusetts, Michigan, and Louisiana. These states have a high risk factor concentration and, in the data so-far, they show higher rates of ICU admissions.</li><li><strong>Average shift to ventilator rate (per state)</strong>: A fraction of people who are referred to ICU actually end up on ventilators, and the survival rate for patients that are ventilated is currently 20% according to medical experts. All the factors mentioned above that affect ICU admission rate also affect ventilator usage rate. Based on available data and our literature survey, we have used a base rate of <strong>40%</strong> for the percentage of ICU patients that require ventilators. For states with higher risk factors — New York, New Jersey, Florida, Georgia, Massachusetts, Michigan, and Louisiana — we have increased by a further 5–10%.</li></ul><p>When we look at the overall distribution of Number of Deaths vs. ICU per capita capacity in each state, we see that no particular trend is available. This suggests that at this point in time, the ICU capacity is not the bottleneck and other factors are the prime drivers of loss of life (e.g. other risk factors, extent of community spread, and social distancing guidelines / adherence).</p><p><em>Figure 13: plot to understand the relationships between number of deaths and number of ICU beds.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*UCNCb_N-gfgAqXWH" /></figure><p><em>Figure 14: Ventilator exhaustion risk — progression (April 4th — May 29 2020)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/773/0*5Mkak3_Hp1a1tHEB" /></figure><p>From our simulation, we observe that many US states especially on the east coast and central US are most susceptible to their medical infrastructure becoming overwhelmed. We can see that states around New York and New Jersey that have high population density are also running high risk of their medical infrastructure becoming overwhelmed, especially running out ventilator systems.</p><p>This analysis can become a key consideration for the federal government to more effectively allocate resources (e.g., ventilators from the federal reserves). As this analysis is forward-looking and can be updated based on the available data, it can be useful for supply chain management decisions.</p><h3>Projection model details</h3><h3>Initial exploration and Data Preparation</h3><p>The data for confirmed cases was available at the state level for the US since January 22, 2020. But the data was highly sparse. We used the CDC data[4] and other sources[5] to collect the COVID-19 <em>confirmed cases</em>, <em>deaths </em>and <em>recovered </em>numbers at daily granularity. The number of recovered patient time-series stopped getting reported after March 23, 2020. Hence we had to exclude that from our analysis. We also used external data sources to triangulate our daily time-series and get the number of hospitalizations or average hospitalization rates per state. These numbers were used later to estimate the<em> ventilator exhaustion risk</em>.</p><p><em>Figure 15: Obesity rate in United States</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*C2gGEKDmBdIG12LE" /></figure><p>At the next step, we plotted the cumulative number of infections and deaths per state alongside the daily change to understand their historical distribution. By analyzing historical data, we realized that for non-coastal states, the numbers started very gradually but escalated very quickly. This is a testament to the highly infectious nature of the disease.</p><p>“<em>At the start of March, with large outbreaks already reported on both coasts, officials in Louisiana had not yet identified a single case of the coronavirus. But in the weeks since, the state has been pummeled. At least 23,000 Louisianans have been infected and at least 1,200 have died.</em>”</p><p><a href="https://www.nytimes.com/interactive/2020/us/coronavirus-us-cases.html">https://www.nytimes.com/interactive/2020/us/coronavirus-us-cases.html</a></p><p>At the next step, we tried to understand if we can find natural categories (i.e. groupings) of states. After analysis, we finalized the groups of size 10 to represent the infection and death rate as for most groups the observations had similar order of magnitude.</p><h3>Key assumptions</h3><p>The standard approach to model disease spread in epidemiology is the <strong>SIR model</strong>. <strong>S</strong> stands for susceptible population, <strong>I</strong> stands for infected population and <strong>R</strong> stands for removed (i.e. cured or diseased) population. We can denote each of these aspects as an ordinary differential equation with reasonable assumptions and simulate the infection spread in a closed system.</p><p>As we are working on a tight schedule, we took a different route towards modeling this problem. We split the overall infections into different states and produced one model per state. The key assumptions embedded in our modeling are as following:</p><ol><li>The infection rates are independent across states</li><li>History contains all the information about the efforts taken by individual states and local communities to curb the infection spread</li><li>Most recent data contain better information about the future than older data</li></ol><h3>Modeling</h3><p>Keeping in mind the assumptions noted above, we experimented with Exponential Smoothing (i.e. Holt-Winters projections) and ARIMA models. Given that the time series per state is severely non-stationary and limited in time, we decided on using <strong>Exponential Smoothing models</strong>. The disease spread in most states showed an exponential pattern; a few that are early stage still demonstrate a linear pattern. Hence we decided to create a set of models per state and choose the best one based on Validation performance (RMSE). We chose a validation window of 5 days as the time-series data available is limited for most of the states. Additionally, we tried both multiplicative and additive trends for Holt-Winters projection. We chose the best one based on the performance on validation data, along with <em>double exponential smoothing</em>.</p><p>Holt-Winters projections provide a reasonably good forecast for shorter time windows but the quality diminishes over longer time horizons. Keeping this in mind, we are projecting for next two weeks (i.e. 14 days only). Our model is designed to be used as a planning tool based on the current data. As new data is available, we can update and get more accurate projections.</p><h3>Insights &amp; Conclusion</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/574/0*w3MQVdrE-TV_l8Kt" /></figure><p><em>Figure 16: Ventilator exhaustion risk — progression (April 4th — May 29 2020) | Correlation with traditional risk factor distribution</em></p><ol><li>In order to get a medium to long term perspective on the evolution of ventilator exhaustion risk, we tried to understand how the correlation between our calculated risk figure changes with the traditional risk-factors (such as obesity rates, smoking rate, and population density). We noted that at present, the COVID-19 situation in the US is heavily influenced by environmental and external factors such as population density. However, with time we expect that the underlying risk factors of the general population will start playing a bigger role. In this regard, we expect that may be <em>another wave of severe-critical cases that hit the most susceptible, and create the greatest demand for critical ICU and ventilator resources, is yet to come</em>.</li><li>We were able to predict the short term infection rates and death rates to a with relatively high accuracy using a simple exponential smoothing model. This indicates to us although it is very hard to conclusively forecast the overall impact of this problem, <em>we can make short term judgements (1–2 weeks) with high degree of precision</em>. This gives us hope in this time of despair.</li><li>From the <em>infection modeling</em> perspective, we have entered the <em>linear zone in most of the states in the US</em> [figure 3]. Deaths are still in the exponential zone at this point. Optimistically, you could conclude that the current wave of infections seems to be coming under control with measures like shelter-in-place. More pessimistically, you could conclude that our testing capacity (or usage) is linear, which is artificially capping and substantially understating the actual number of COVID-19 positive cases.</li><li>Our modeling suggests that the <em>midwestern and southern states — </em>Georgia, North and Sound Dakota, Florida and Louisiana — <em>are at highest risk of overwhelming their medical system resources in the next 30 to 45 day </em>[figure 14]. The west coast states seem relatively lower risk, in comparison.</li><li>We collected data on high-risk factors and concluded that the primary risk factor for COVID-19 infection (confirmed cases) was Density; the primary risk factors for death from COVID-19 are Age, Obesity, and Smoking. We didn’t include diabetes separately here since studies have shown it is highly correlated to obesity (85% of diabetics are obese and 30% of overweight people are diabetic). Other risk factors for death, like heart and lung disease, could be relevant, but there was very limited data for those conditions per state.</li></ol><h3>References</h3><ol><li><a href="https://en.wikipedia.org/wiki/Demographics_of_the_United_States#Structure">https://en.wikipedia.org/wiki/Demographics_of_the_United_States#Structure</a></li><li><a href="https://www.aha.org/statistics/2020-01-07-archived-fast-facts-us-hospitals-2019">https://www.aha.org/statistics/2020-01-07-archived-fast-facts-us-hospitals-2019</a></li><li><a href="https://www.cdc.gov/coronavirus/2019-ncov/need-extra-precautions/groups-at-higher-risk.html">https://www.cdc.gov/coronavirus/2019-ncov/need-extra-precautions/groups-at-higher-risk.html</a></li><li><a href="https://www.elegantjbi.com/blog/what-is-the-holt-winters-forecasting-algorithm-and-how-can-it-be-used-for-enterprise-analysis.htm">https://www.elegantjbi.com/blog/what-is-the-holt-winters-forecasting-algorithm-and-how-can-it-be-used-for-enterprise-analysis.htm</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20c9c37d22ee" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Inaccurate Data Visualization and COVID-19]]></title>
            <link>https://medium.com/@vossough/inaccurate-data-visualization-and-covid-19-1e2347286594?source=rss-3f8d70cd668b------2</link>
            <guid isPermaLink="false">https://medium.com/p/1e2347286594</guid>
            <dc:creator><![CDATA[Angel]]></dc:creator>
            <pubDate>Mon, 09 Mar 2020 06:26:58 GMT</pubDate>
            <atom:updated>2020-03-09T06:26:58.685Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>“Creating data-driven visualizations of</strong><a href="https://twitter.com/hashtag/COVID19?src=hashtag_click"><strong> #COVID19</strong></a><strong> is hard, but I worry that a lot of people are being misled because designers aren’t being transparent about the tradeoffs of their visual representations.”</strong></p><p><strong><em>_Evan Peck </em></strong><em>(Assistant Professor of Computer Science, Bucknell University)</em></p><p><strong>Problem 1: What do you do when the numbers don’t reflect reality?</strong></p><p>The problem is with visualizing data that is inaccurate &amp; incomplete because it’s <strong>based only on confirmed cases from PCR tests </strong>when there is very limited access to tests, and test accuracy is poor (73% sensitivity which means a 27% “false negative” rate). <strong>Many viewers will assume the visualization represents the total number of people with the virus</strong>.</p><p>“As you can see in nearly all visualizations, they only rely on confirmed positive tests of<a href="https://twitter.com/hashtag/COVID19?src=hashtag_click"> #COVID19</a> (top: CDC, bottom: NYTimes). As many experts have pointed out, these numbers dramatically underestimate reality. I suspect that many interpret these maps as “number of people with the virus”. That’s bad!!”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lUEZmhuuTAO0VH6s" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oZEIWW407YCGCvls" /></figure><p>“The number of *actual*<a href="https://twitter.com/hashtag/COVID19?src=hashtag_click"> #COVID19</a> cases contains tremendous uncertainty (estimates in around Seattle alone are 10x larger than the positive tests), but we aren’t visually representing ANY uncertainty because we know the number of *positive tests*.</p><p>There isn’t an easy solution here, but I was surprised that none of the maps (at least) contextualized these values with the number of<a href="https://twitter.com/hashtag/COVID19?src=hashtag_click"> #COVID19</a> tests we’re running in general, or provide caveats on how the numbers are understated/incomplete.”</p><p><strong>Problem 2: How do you help people understand the impact of time in exponential situations?</strong></p><p>There is no time in these visualizations. The problem is that you are presenting static data in a highly dynamic and exponential situation. Based on CDC visualization one could conclude really there is no problem anywhere except in China, Italy and Iran.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*C070z7BXE58Uvzph" /></figure><p>So you would probably think <strong>thank god I’m not in China, Italy or Iran,</strong> right?</p><p>Not exactly, because <strong>we are BAD at understanding exponential growth</strong> and how dangerous the current situation is. Unfortunately 10–12 weeks ago, China numbers were exactly like the United States. I am not usually a big fan of logarithmic scale since it can be confusing but in this exponential situation, it may be more useful to present data on logarithmic scale since the difference between 100 and 1000 is not actually much.</p><p>“Unfortunately, when we rely purely on raw numbers, we’re accurately portraying the values, but inaccurately portraying the situation.”</p><p>_Angel Vossough</p><p>Research Design and Data Analysis, UC Berkeley</p><p><strong>Reference:</strong></p><p><a href="https://twitter.com/EvanMPeck/status/1235576895628021760">https://twitter.com/EvanMPeck/status/1235576895628021760</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e2347286594" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>