<?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 Keith Brock on Medium]]></title>
        <description><![CDATA[Stories by Keith Brock on Medium]]></description>
        <link>https://medium.com/@biobrock?source=rss-387a294c0cf9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*siyoYpI7GuFg-95wZ31y2w.jpeg</url>
            <title>Stories by Keith Brock on Medium</title>
            <link>https://medium.com/@biobrock?source=rss-387a294c0cf9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 02:38:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@biobrock/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[Interdependence of Diagnostic Hardware on Artificial Intelligence Algorithms]]></title>
            <link>https://medium.com/@biobrock/interdependence-of-diagnostic-hardware-on-artificial-intelligence-algorithms-fb8c4721e0a1?source=rss-387a294c0cf9------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb8c4721e0a1</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[clinical-diagnostic]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[medical-devices]]></category>
            <category><![CDATA[clinical-research]]></category>
            <dc:creator><![CDATA[Keith Brock]]></dc:creator>
            <pubDate>Fri, 05 May 2023 23:21:51 GMT</pubDate>
            <atom:updated>2023-05-11T15:07:51.858Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hZmPNVbzl7_y0KtT" /><figcaption>Photo by <a href="https://unsplash.com/@nci?utm_source=medium&amp;utm_medium=referral">National Cancer Institute</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In the field of medical diagnostics, there is a growing trend of artificial intelligence start-ups who ambitiously claim to be hardware agnostic. While their software may in theory be broadly applicable across multiple hardware platforms, this competitive advantage may be difficult to leverage in practice. Developers struggle to follow through with this claim, as the diagnostic hardware, and the clinical data captured thereon, play a critical role in the algorithm training, clinical validation, and regulatory approval of diagnostic systems. While the interdependence of the platform and diagnostic algorithms are essential to ensure safety, efficacy, and reliability, this has created an innovation bottleneck and introduces business risks difficult to mitigate.</p><p><strong>Hardware-signature</strong></p><p>The diagnostic hardware is integral and inextricably tied to the clinical data it captures. Different types of platforms, such as imaging devices, blood analyzers, and electrocardiograms (ECGs), have their own unique strengths and weaknesses relative to their competitors, leaving their signature on the diagnostic data. This hardware-signature forces developers to narrow the training of machine learning algorithms to a single device or technology at a time to ensure no cross-platform bias is inadvertently introduced into the training set.</p><p><strong>Clinical Data Burden</strong></p><p>Furthermore, the quality and quantity of clinical data captured are critical for the development of high-performing diagnostic algorithms. This is especially acute in the development of diagnostic AI algorithms. The amount of data captured, and its diversity, can affect the performance of AI algorithms. The larger and more diverse the dataset, the more accurate the algorithm is likely to be. As a guideline, it is advisable to have an order of magnitude more data points (rows) as you have trainable parameters (columns). When assessing the quality of a data set, developers are generally concerned with reliability (noise, label errors, omitted values, duplicate values, etc.) or feature representation (presence of features, diversity of data, outliers, etc.). Depending on the features of the AI-driven decision support algorithm, rows can number in the tens of thousands, hundreds of thousands, or even millions, representing a significant investment in time and resources for researchers and the developers.</p><p><strong>System-based Regulatory Approval</strong></p><p>The current regulatory approval process of AI diagnostic systems is another critical aspect that closely links diagnostic hardware to an AI decision support tool. Regulatory agencies evaluate the safety and efficacy of medical devices, including AI-based software as a medical device (SaMD), as complete diagnostic systems requiring the combination of the hardware and software in the same application and approval. Every change to the hardware or software will require subsequent premarket review. In a move that shows some awareness of the innovation-inhibiting position they currently occupy; the FDA does provide guidance on a “predetermined change control plan” as part of the initial premarket submission which would outline the manufactures anticipated modifications and associated operating procedures to be used to implement those changes in a controlled manner to offset any risk to patients. However, the regulatory approval process requires consideration of the system, including the hardware on which the algorithm will be in use.</p><p><strong>Platform Selection Criteria</strong></p><p>Clearly, choice of diagnostic hardware on which the AI algorithm will be built is foundational, be it a budding AI start-up or multinational, vertically integrated diagnostic organization manufacturing both hardware and software. While the hardware-signature and the burden of clinical data collection may someday be addressed through AI innovation, it is unlikely that the system-level regulatory approval paradigm will change soon. This leaves inherent risk for software developers when selecting the hardware on which they will stake their development and company objectives.</p><p>Where this hardware is supplied by a third party with different company goals and development objectives, this risk is compounded. When selecting a hardware for development, the following business KPIs and platform criteria may be considered.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/1*1cb1OtvAXMLrRL-Hw-GI7Q.png" /><figcaption>Table 1. Business KPIs and Associated Platform Criteria</figcaption></figure><p>Given the relatively long development timelines associated with clinical data collection and machine learning training, a platform that met business needs at the start of a venture may not once minimum viable clinical performance are established, and regulatory approval is obtained. The prospect of migrating to another platform and “starting over” will feel demoralizing for a development team and perplexing to investors.</p><p>It is time to evaluate target markets, assess customer segmentation, refine the repeatable sales process, and reset investor expectations.</p><p><strong>Built to Migrate</strong></p><p>Platform migration is inevitable and is a litmus test on which all platform-agnostic AI companies will eventually pass or collapse. Building processes with platform migration in mind will be an accelerator when that day comes and will demonstrate to investors that your company is built to survive despite any third-party dependencies it may carry into development or commercialization.</p><p>Look to build on universal codebases to enable easier port from one platform to another. There will be changes and challenges to be addressed, but building on universal codebase can provide a strong advantage. As with any development activity, the migration to a new platform will require a well-considered design exit strategy. All old test cases will need to be modified, and likely new test cases will need to be added. If post commercialization, additional considerations will need to be made about how to migrate the customer-base. If upgrade will be optional for customers, expect and plan for a significant increase in the time to release maintenance or features as design exit processes will need to scale with the number of supported platforms.</p><p>Also, through careful distributed computing and delineation of what is or is not included as part of your SaMD<a href="#_msocom_6">[KB6]</a> , large parts of your software architecture could be salvaged or untouched by the platform migration or regulatory review.</p><p><strong>Pivot and Learn</strong></p><p>There is no need to throw the AI out with the bath water. Often there is an appropriate market, perhaps with a smaller total addressable market (TAM), that will allow commercialization of the approved AI algorithm on the sub-optimal platform. This will allow the organization to demonstrate product market fit, establish pricing and revenue models, and develop and refine a repeatable sales process. These critical commercialization activities are largely independent of the hardware and are required elements of a successful launch. Finally, executives within the organization need to establish the inevitability of platform migration with developers, investors, and boards of directors (BoD) early in the life of the organization. Investors and BoD may need to be educated on the challenges and necessity of platform migration.</p><p>In conclusion, the interdependence of diagnostic hardware and AI algorithms is foundational to the development and regulatory approval of diagnostic systems. The hardware-signature present in clinical data can create bias, limiting the machine learning algorithm’s applicability to other platforms. The quality and quantity of clinical data captured by the hardware are critical for the development of high-performing diagnostic algorithms, and the regulatory approval process considers the hardware and software as complete diagnostic systems. To mitigate business risks and innovation bottlenecks, developers must consider cross-functional platform selection criteria and build processes with platform migration in mind. In some cases, building on universal codebases and using distributed computing can facilitate platform migration, enabling the algorithm to be salvaged or untouched by the regulatory review. Companies must be prepared to pivot and learn, and may find appropriate markets to commercialize the AI algorithm on sub-optimal platforms, and demonstrate its product value while migrating to the next generation hardware.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb8c4721e0a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>