The Challenge of Regulating Clinical Decision Support Software after 21st Century Cures
15 Pages Posted: 21 Mar 2018
Date Written: March 17, 2018
Clinical decision support (CDS) software broadly refers to software that assists healthcare providers in combining patient-specific data with general sources of medical knowledge to make better diagnostic and treatment decisions in the clinical setting. The 21st-Century Cures Act strips FDA of jurisdiction to regulate some (not all) CDS software. To qualify for this exclusion from FDA regulation, 21 U.S.C. § 360j(o)(1)(E)(iii) requires that the software must be intended to enable “the health care professional to independently review the basis” for its recommendations so that it is “not the intent that such health care professional rely primarily” on the software’s recommendations when making diagnostic and treatment decisions about individual patients. This article explores whether this is a workable standard as applied to advanced CDS software that uses machine learning to glean insights from real-world clinical experience and then applies these insights to improve the quality of patient care. We conclude that the standard Congress set out in 21st-Century Cures is potentially workable, but only if FDA takes additional steps to clarify the standards of transparency that CDS software must meet before it can escape FDA regulation. Transparency in this context includes algorithmic transparency, physician access to the underlying data that the software relies on in rendering decisions, and business transparency as reflected in the terms of contracts between CDS software vendors and users.
FDA’s recent draft guidance on CDS software leaves a crucial question unresolved. This question cuts to the very heart of what is wrong with the US healthcare system and how to fix it: Is the problem simply that doctors are not heeding existing medical evidence — for example, by ignoring warnings in FDA-approved drug labeling or failing to keep up with findings in the peer-reviewed medical literature? Or is the problem that the existing evidence base is itself inadequate and flawed — for example, because FDA-approved labeling relies on contrived clinical trials that fail to reflect real patients, or because peer-reviewed literature is skewed by publication biases that favor studies in which the treatment worked, or because clinical practice guidelines can be captured by commercial interests? FDA’s draft guidance on CDS software — perhaps as an unintended consequence — would expedite market entry for simple CDS software that promotes physician conformity with the existing medical evidence base, while imposing higher regulatory hurdles that delay the clinical translation of machine-learning software that may be our best hope to overcome flaws in current medical evidence? Is this the right path forward?
Keywords: Clinical Decision Support Software, Software as a Medical Device, FDA, Artificial Intelligence, Machine Learning
Suggested Citation: Suggested Citation