Research Letter
October 9, 2023
Analysis of Devices Authorized by the FDA for Clinical Decision Support in Critical Care
Jessica T. Lee, Alexander T. Moffett, George Maliha, et al
JAMA Intern Med. Published online October 9, 2023. doi:10.1001/jamainternmed.2023.5002
The use of predictive clinical decision support (CDS) devices (ie, those that use machine learning [ML] or artificial intelligence [AI]) has the potential to improve outcomes in critical care, but a clear regulatory framework is lacking.1 Recent guidance from the US Food and Drug Administration (FDA) suggests most CDS tools for critical illness will be regulated because of the time-sensitive nature of the decisions informed by these devices. However, growing concerns about the clinical impact of predictive CDS systems raise questions about whether current device regulatory frameworks, developed before advanced statistical learning methods were widely available, are sufficient to ensure effectiveness and safety.2
On September 22, 2021, the FDA released a public database of authorizations for medical devices that use ML or AI. We sought to identify devices that offer CDS in a critical care setting and characterize the evidence cited in their authorization.
Methods
We extracted data from the AI and ML database as of December 15, 2022, and augmented those data through the OpenFDA interface.3,4 The 2 most common FDA pathways for CDS device approval are the 510(k) pathway, which requires demonstration of substantial equivalence to a previously authorized device (hereafter, a predicate) and typically does not require submission of clinical data, and the de novo pathway, which indicates that a novel device has low to moderate risk and reasonable assurance of being safe and effective. We identified devices relevant to critical care and searched the main FDA database for additional high-profile devices that were not found in the AI/ML data set.4 For each device, we recorded whether the decision summary included clinical evidence, software code, safety evaluation, and consideration of potential performance bias among historically marginalized groups. For each device, we traced back all predicates developed before CDS systems using AI/ML methods were widely used. This case series was exempt from institutional review and informed consent by the Common Rule given its use of publicly available data.
Results
Of 521 authorizations in the FDA AI/ML database, we identified 10 that might inform care for patients with critical illness (Table 1). Of these, only 3 included citations of published data, 4 mentioned a safety assessment, and none mentioned an evaluation of performance bias (Table 2). All but 1 device were authorized through 510(k) clearance, which relies on substantial equivalence to predicates, but only 3 devices included AI/ML predicates. Notably, a high-profile sepsis-focused prediction model that ostensibly would meet criteria as a CDS device was not found in the AI/ML or full FDA databases.5We found no studies examining the clinical impact on care processes or patient outcomes for these device authorizations.
Discussion
While many prediction models might offer CDS for patients with critical illness,6 our review of the database revealed that only 10 AI/ML CDS devices have received FDA authorization. The clinical evidence for these devices ranged from completely absent to peer-reviewed assessment of model performance, and most of the devices authorized through the 510(k) pathway relied on equivalence to non-AI/ML predicates. Furthermore, at least 1 high-profile and widely implemented model5 did not appear to have received FDA authorization. While this study was limited to critical care, these findings highlight the need to update regulatory requirements to align with current knowledge about using AI/ML systems across many clinical practice settings.
Although the release of the curated FDA database4 permits easier identification of FDA-authorized devices that rely on AI/ML methods, users must look elsewhere to obtain essential information about the clinical effectiveness, safety, and performance biases of a given CDS system. The criteria for establishing equivalence for 510(k) clearances could be better adapted to both current AI/ML methods and the clinical environment of high-risk decisions for patients with critical illness.
Limitations of this study include not accounting for FDA approvals made through the recently ended precertification pilot pathway or for other devices in widespread use not present in the FDA database.
