<?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[Swift Medical Technology - Medium]]></title>
        <description><![CDATA[Technical updates from a health tech startup changing the face of wound care - Medium]]></description>
        <link>https://medium.com/swift-medical-technology?source=rss----5677fcad069e---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Swift Medical Technology - Medium</title>
            <link>https://medium.com/swift-medical-technology?source=rss----5677fcad069e---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:31:53 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/swift-medical-technology" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Creating the future of wound prognostics using machine learning]]></title>
            <link>https://medium.com/swift-medical-technology/creating-the-future-of-wound-prognostics-using-machine-learning-c8037471999?source=rss----5677fcad069e---4</link>
            <guid isPermaLink="false">https://medium.com/p/c8037471999</guid>
            <category><![CDATA[forecasting]]></category>
            <category><![CDATA[survival-analysis]]></category>
            <category><![CDATA[wound-care]]></category>
            <category><![CDATA[pressure-ulcer]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Rishabh Gupta]]></dc:creator>
            <pubDate>Tue, 21 Sep 2021 20:39:56 GMT</pubDate>
            <atom:updated>2021-09-24T21:09:17.073Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Swift Medical assists clinicians in treating tens of thousands of acute and chronic wounds everyday. Contrary to acute wounds, chronic wounds fail to heal in a timely manner and affect millions of people across the world. These wounds result in decreased quality of life due to pain, suffering and other associated complications such as infection and amputation. Moreover these wounds are very expensive to treat and therefore end up costing health systems huge sums of money every year. This can reach up to $97 billion for some health systems according to a 2018 estimate<strong>¹</strong>. At an individual wound level, a single Stage II Pressure Injury could end up costing $40000, whereas a Stage IV Pressure Injury could cost approximately $90000<strong>²</strong>, indicating that the net per-person hospitalization cost increases with wound severity. The incidence of such wounds are expected to grow due to the aging population and rising prevalence of co-morbidities like diabetes and obesity in the general population. Therefore, early detection of a chronic wound and tracking its progression and severity can enable clinicians to apply appropriate interventions, avoid adverse complications and reduce some of the preventable expenses.</p><p>Generally, healthcare providers track wound healing progression and severity using various wound assessment tools such as Pressure Ulcer Scale for Healing (PUSH)<strong>³</strong> and Bates-Jensen Wound Assessment Tool (BWAT)<strong>⁴</strong>. However, these tools come with their set of shortcomings. As such these tools rely on very few, subjective and sometimes imprecise wound characteristics to quantify wound progression sometimes accounting for just 60% of the variance in the dataset<strong>⁵</strong>. Moreover these tools only provide a snapshot of the wound’s progression on the current day, not considering its historical performance or performance of identical wounds. To overcome these challenges we leverage our massive and growing dataset collected over time by clinicians using Swift Skin and Wound App to develop interpretable machine learning models. These are critical to the creation of ethical wound prognostics tools that can accurately track and forecast wound healing progression, leading to:</p><ul><li>Precise triaging of patients through wound severity quantification and risk stratification</li><li>Better decisions about treatments and other clinical interventions for individual wounds/patients</li><li>Increased compliance of patients to treatment regimes as they are better informed about the likely progression of the wound and</li><li>Overall improvement of wound care and outcomes</li></ul><p>In the next few sections we will provide a brief discussion on prognostics and the state-of-the-art techniques used at Swift to build prognostic models for wounds.</p><h3>What is Prognostics?</h3><p>In healthcare, prognostic techniques allow us to study the risk of an individual developing a particular health outcome over a specific period of time, given the clinical and non-clinical profile of the patient<strong>⁶</strong>. Outcomes correspond to specific events such as death or complications that arise from disease progression. Researchers generally focus on developing multivariate prognostic models that can help identify predictors of outcomes, highlight relative importance of these predictors and predict risk of outcomes accurately. One group of machine learning models that satisfy these criteria is survival analysis models.</p><p>Survival analysis (also known as time-to-event analysis) models focus on modeling expected duration of time until an event (outcome) occurs. These model the “survival time”, time between origin and final events of interest of an entity, for example time from manufacture of a component to its failure, time between start and end of a government’s term. Such models are widely used in clinical research and have found interesting applications in other areas such as customer churn, asset liquidity risk and component lead time prediction in a variety of industries. These models allow easy interpretability of predictors’ importance which is really important for providing actionable insights to clinicians. Moreover, these models adapt well to censored data, where data is unknown beyond a particular time instance resulting from patients dropping off from the trial, which is very commonly observed in healthcare research.</p><h3>Prognostics at Swift Medical</h3><p>At Swift Medical, we are most interested in modeling time until certain wound progression related outcomes, specifically wound closure or healing. While building prognostic models for modeling time-to-heal we need to keep a good balance between model accuracy and interpretability, while considering various biases and sources of noise (e.g., data censoring) that contribute to our dataset. Wound progression is characterized by a number of factors not outlined here for brevity. However, the first task of modeling wound prognosis is to identify most relevant features and their relative importance in modeling time-to-heal. Once identified, these features can then be used to train a multivariate model for predicting a risk score that quantifies probability of delayed (or sometimes not) healing, and can be used to triage patients and aid clinicians in deciding the appropriate interventions. Following risk quantification, forecasting models can be leveraged to understand the dynamics of wound progression, which can help provide actionable insights to clinicians while treating complex wounds.</p><h4>Risk Stratification Models</h4><p>In order to analyze and model survival times of wounds, defined as time until wound closure, we leverage a multitude of tools available under the wide umbrella of survival analysis. Kaplan-Meier estimator is one such model, which is used to quantify the probability of survival of a wound beyond a specific time. It is most useful when the predictor is categorical.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*egb8MXj-qA1q3w5ysFr9zA.jpeg" /><figcaption>Kaplan-Meier Plot showing the differences in wound survival probabilities over time for Pressure Ulcers and Abrasion wounds.</figcaption></figure><p>However, many of the predictors in general practice are ordinal, interval or numerical in nature. For such situations, another model known as Cox Proportional Hazards (Cox-PH) regression model is applied. Cox-PH model allows for direct interpretation of model weights and its main product is a risk index, leveraged in patient risk stratification. Generally, a stepwise forward-backward feature selection step is applied to reduce the dimensionality of the dataset before the final model is trained. Cox-PH model performance is generally quantified using a metric called Harell’s c-index or concordance index, computed as a ratio of concordant pairs and all possible evaluation pairs used in risk stratification<strong>⁷</strong>. At Swift, we regularly leverage the Cox-PH model to train our risk stratification models which aid in accurate patient triaging.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*txhmTmmxPXGOkcvo6yea-w.jpeg" /><figcaption>Feature weights computed using Cox-PH model for a wound risk stratification model</figcaption></figure><h4>Trajectory Forecasting Models</h4><p>On the other hand, for forecasting wound progression researchers have traditionally focussed on using simpler and more interpretable models like piecewise linear regression models<strong>⁸</strong>. However, leveraging more sophisticated time series forecasting tools, such as Bayesian Regression models, Hidden Markov Models (HMM), Recurrent Neural Networks (RNNs), Long short Term Memory (LSTM), can help improve the accuracy of time-to-heal prediction and wound progression trajectory forecasting. Currently, we rely on our award winning technology that adapts to patient specific data and accurately forecasts time-to-heal for pressure injuries. Our model can determine the time-to-heal for a large wound (&gt;3cm²) with a mean absolute error of approximately three weeks, leading to better wound progression tracking .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TDttNqFSkMYQ1dQpXkVsPg.gif" /><figcaption>Wound trajectory forecasting model based on mixed effects model.</figcaption></figure><h3>Future of Prognostics at Swift Medical</h3><p>Although models like Cox-PH have an advantage of being more interpretable, these are not very good at taking predictor interaction into account. Recent advances in survival analysis such as Random Survival Forests<strong>⁹</strong> and DeepSurv<strong>¹⁰</strong> have shown to provide better accuracy in clinical prognosis of certain medical conditions. Not only are these models more accurate, but these also open avenues for building personalized intervention recommendation models. Furthermore, advanced tools like causal inference and reinforcement learning can provide extra sophistication needed to build dynamic personalized wound treatment regimes. Thus, assisting clinicians by providing actionable insights and deeper understanding of the effects of various interventions available for treating wounds. We are excited to utilize these machine learning tools and our rich dataset to improve the quality of care for our patients and apply them to entirely new predictive capabilities beyond the scope of wound healing.</p><h3>References</h3><ol><li>Sen, Chandan K. “Human Wounds and Its Burden: An Updated Compendium of Estimates.” <em>Advances in wound care</em> vol. 8,2 (2019): 39–48. doi:10.1089/wound.2019.0946</li><li>Ocampo, Wrechelle et al. “Economic Evaluations of Strategies to Prevent Hospital-Acquired Pressure Injuries.” <em>Advances in skin &amp; wound care</em> vol. 30,7 (2017): 319–333. doi:10.1097/01.ASW.0000520289.89090.b0</li><li>NPIAP. “Pressure Ulcer Scale for Healing (PUSH)” <a href="https://npiap.com/page/PUSHTool">https://npiap.com/page/PUSHTool</a> (Accessed on: 7th Sept 2021)</li><li>Bates-Jensen, Barbara M. et al. “Reliability of the Bates-Jensen wound assessment tool for pressure injury assessment: The pressure ulcer detection study.” <em>Wound repair and regeneration : official publication of the Wound Healing Society [and] the European Tissue Repair Society</em> vol. 27,4 (2019): 386–395. doi:10.1111/wrr.12714</li><li>Stotts, Nancy A., et al. “An instrument to measure healing in pressure ulcers: development and validation of the pressure ulcer scale for healing (PUSH).” <em>The Journals of Gerontology Series A: Biological Sciences and Medical Sciences</em> 56.12 (2001): M795-M799.</li><li>Moons, Karel GM, et al. “Prognosis and prognostic research: what, why, and how?.” <em>BMJ </em>338 (2009).</li><li>Royston, Patrick, and Douglas G. Altman. “External validation of a Cox prognostic model: principles and methods.” <em>BMC medical research methodology</em> 13.1 (2013): 1–15.</li><li>Gardner, Sue E. et al. “A prospective study of the PUSH tool in diabetic foot ulcers.” <em>Journal of wound, ostomy, and continence nursing : official publication of The Wound, Ostomy and Continence Nurses Society</em> vol. 38,4 (2011): 385–93. doi:10.1097/WON.0b013e31821e4dbd</li><li>Ishwaran, Hemant, et al. “Random survival forests.” <em>The annals of applied statistics</em> 2.3 (2008): 841–860.</li><li>Katzman, Jared L., et al. “DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network.” <em>BMC medical research methodology</em> 18.1 (2018): 1–12.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c8037471999" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swift-medical-technology/creating-the-future-of-wound-prognostics-using-machine-learning-c8037471999">Creating the future of wound prognostics using machine learning</a> was originally published in <a href="https://medium.com/swift-medical-technology">Swift Medical Technology</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PHIPA modernization accelerated in a time of COVID]]></title>
            <link>https://medium.com/swift-medical-technology/phipa-modernization-accelerated-in-a-time-of-covid-4c69a75085aa?source=rss----5677fcad069e---4</link>
            <guid isPermaLink="false">https://medium.com/p/4c69a75085aa</guid>
            <category><![CDATA[health-technology]]></category>
            <category><![CDATA[phipa]]></category>
            <category><![CDATA[covid19]]></category>
            <category><![CDATA[data-privacy]]></category>
            <dc:creator><![CDATA[Raul Rupsingh]]></dc:creator>
            <pubDate>Mon, 25 May 2020 16:46:00 GMT</pubDate>
            <atom:updated>2020-05-25T16:46:00.906Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*HUWYGputoPf5iO7u-RBr2A.jpeg" /></figure><p>Bill 188, the <em>Ontario Economic and Fiscal Update Act</em>, updated in March of 2020 had several significant statue changes in response to the COVID-19 pandemic, and included important amendments to the Personal Health Information Protection Act (PHIPA).</p><p>These changes have implications for Electronic Service Providers (ESP), such as Swift Medical, as well as Health Information Custodians (HIC) across Ontario.</p><p>Here’s my summary of some of the significant changes:</p><ul><li>Increased penalties for data privacy and security breaches, where fines have been increased by 2x</li></ul><blockquote><em>PHIPA fines are now $200k for an individual</em> <em>and</em> <em>$1M for organizations</em></blockquote><ul><li>The Information and Privacy Commissioner (IPC) has been given <em>expanded powers for administering and enforcing PHIPA</em> compliance</li></ul><blockquote>The IPC can fine HICs that work with identified non-compliant software vendors</blockquote><ul><li>More rigorously defined requirements for electronic audit logs from HICs that can be inspected periodically by the IPC</li></ul><blockquote>Meaning that HICs will have to enforce the same standards upon their service vendors</blockquote><ul><li>A new class of entities called “Consumer Electronic Service Providers” who provide health service to individuals and not HICs</li><li>De-identification rules move beyond just IPC best practices to a minimum legal standard for entities as part of their PHIPA compliance</li><li>Re-identification guidelines for certain identifiers in particular healthcare use cases for verification and patient linking which will be an important interoperability requirement for the future of Ontario Health Teams sharing data across continuum</li></ul><p>Generally these amendments to PHIPA significantly increase HIC liability, as previously Custodians could only face monetary sanctions for committing offences that were approaching deliberate breaches of the Act. Now with expanded scope for enforcement, the IPC can fine organizations for <em>not implementing good policies and procedures</em> around the protections of Protected Health Information.</p><p>Whether you’re a software vendor or a healthcare custodian it’s important to note that PHIPA and the IPC can <em>hold you as an individual employee liable</em> if you authorize or knowingly allow an offence to continue.</p><p>Taking the information security practice bar from deliberate breach to encompass negligence on behalf of Providers or Custodians is an important step for Canadian healthcare and brings enforcement more in line with our US counterparts.</p><p>Ultimately this is a boon for individual consumers, who have additional protections enforced by the Canadian legal framework, as well as expanded rights to pursue civil actions or class action lawsuits in situations of privacy breaches.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4c69a75085aa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swift-medical-technology/phipa-modernization-accelerated-in-a-time-of-covid-4c69a75085aa">PHIPA modernization accelerated in a time of COVID</a> was originally published in <a href="https://medium.com/swift-medical-technology">Swift Medical Technology</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Design Heuristics for Decomposing Monoliths]]></title>
            <link>https://medium.com/swift-medical-technology/design-heuristics-for-decomposing-monoliths-bcbbdba60b87?source=rss----5677fcad069e---4</link>
            <guid isPermaLink="false">https://medium.com/p/bcbbdba60b87</guid>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[microservices]]></category>
            <dc:creator><![CDATA[Raul Rupsingh]]></dc:creator>
            <pubDate>Sun, 24 May 2020 16:41:01 GMT</pubDate>
            <atom:updated>2020-05-24T16:41:00.972Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/630/1*q9BfEgFjUWCx00cKbbfvcw.jpeg" /></figure><blockquote>“When you design a system…, then if the features can be broken into … loosely bound groups of relatively closely bound features, then that division is a good thing to be made a part of the design. This is just good engineering.”</blockquote><blockquote>— Tim Berners-Lee</blockquote><p>Decomposing monoliths is a trendy topic in software architecture these days.</p><p>I saw a great talk at #SACON by Vladik Khrononov that had some important caveats and design heuristics for approaching decomposition that I thought I’d share.</p><p>As we decompose software services from a monolithic big ball of mud, we tend to apply Domain Driven Development towards defining Bounded Contexts based on business domains. From there we can continue the decomposition further towards microservices, and then even further still.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*pbjAMQT55LWD56kF" /></figure><p>As you tell from this graphical depiction, there is a clear tradeoff between service size and system complexity &amp; maintenance, as with a proliferation of nano-services you gain the benefit of high modularity but with the cost of high communication &amp; coordination between services.</p><p>A service API is the interface through which you exchange information in and out of a given service. A microservice presents a micro-interface, reducing the surface area of that API which helps to reduce coupling between services, making that service easier to understand, more fault tolerant and less likely to require modification as the system changes.</p><blockquote>”Global complexity …is the complexity of the overall structure of a program or system. i.e., the degree of association or interdependence among the major pieces of a program” — Glenford J. Myers</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vAibgCQNNnifQZ11" /></figure><p>As companies migrate their systems towards microservices it’s tempting to go overboard, and in doing so create a distributed monolith which actually increases system complexity. Whereby now you have to modify many small services to get a new feature shipped instead of making changes in a single monolithic code base.</p><p>But how do you know how far you should take your decomposition?</p><p>Vladik proposed several design heuristic to help guide architects making these types of engineering tradeoffs, and here’s a summary:</p><p>1. Decompose to Bounded Contexts</p><ul><li>Don’t implement logically conflicting models in the same service, instead try to decompose to the bounded context level</li></ul><p>2. Don’t First Law of Distributed Object Design</p><ul><li>“<em>Don’t distribute your objects</em>” — Martin Fowler</li><li>Unless there’s a clear business goal try not to decompose bounded contexts further at this point</li><li>Remember that you may not actually need microservices, so don’t decompose unnecessarily</li></ul><p>3. Logically segment your domains into Core, Generic and Supporting</p><ul><li>Buy or adopt Generic subdomains as these are mundane and complex problems everybody has and are already solved by other software vendors</li><li>Core subdomains represent your business’s competitive advantage and what you hope to deepen and grow</li><li>Supporting subdomains are not complex or interesting, offer no competitive advantage but are necessary to support the Core subdomains</li></ul><p>4. Don’t Rush</p><ul><li>Extract your Core subdomains and strictly adhere to their boundaries</li><li>Decompose further only when you acquire deep domain expertise</li><li>The best designs require learning and iteration, almost no one gets this right the first time</li></ul><p>5. Supporting Subdomains</p><ul><li>Because these change less frequently it’s often safe to decompose subdomain boundaries early for supporting services</li></ul><p>6. Evaluate Consistency Requirements</p><ul><li>What are the dependencies between your services?</li><li>Does one service need to read the last write of the other?</li></ul><p>• Could use two services with synchronous communication</p><ul><li>Is eventual consistency ok?</li></ul><p>• Could use 2+ services with asynchronous communication</p><ul><li>Do you need to control concurrency?</li></ul><p>• Perhaps put these into the same service</p><ul><li>A good heuristic is that if your microservice design requires distributed atomic transactions, just <em>don’t do it</em> and combine those into a single service.</li></ul><p>7. Public / Private Events</p><ul><li>Categorize events as public or private, where public events represent your “service door” and remember that you can collapse multiple private events into public event(s) too.</li><li>Keep domain knowledge inside of services and expose only relevant state-altering events</li></ul><p>8. Make Events Explicit</p><ul><li>Eliminate ambiguity so clients don’t have to make assumptions about your events or interfaces</li></ul><p>9. Evaluate Reasons for Change</p><ul><li>If two services regularly need to change together that’s a design smell that perhaps they should be the same service</li></ul><p>10. Evaluate Service Doors</p><ul><li>Microservices should have a micro “service door” or interface</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bcbbdba60b87" width="1" height="1" alt=""><hr><p><a href="https://medium.com/swift-medical-technology/design-heuristics-for-decomposing-monoliths-bcbbdba60b87">Design Heuristics for Decomposing Monoliths</a> was originally published in <a href="https://medium.com/swift-medical-technology">Swift Medical Technology</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>