The U.S. Food and Drug Administration is all about hard science, and correlation of variables is not evidence of causation. To make a regulatory decision, FDA has long held that line, requiring evidence of actual occurrences — e.g., data from a clinical trial showing X number of patients experienced side effect Y, precise HPLC test results and the like. Until last week, that is. In the case of emergency use authorization (EUA) for the drug Anakinra, a treatment for COVID-19, on July 31, 2024, FDA acknowledged it relied on the soft science of prediction, through artificial intelligence / machine learning (AI/ML), to make a regulatory decision: the identification of the patient population most likely to benefit from taking a drug.
In “Using Machine Learning to Identify a Suitable Patient Population for Anakinra for the Treatment of COVID-19 Under the Emergency Use Authorization,” the Agency directly acknowledges this “first.” Yes, the publication of this explanation of its use of the technology aligns with the emphasis the Agency has said it will place on algorithmic transparency; and yes, the statement was an opportunity for FDA to tout its ability to play in the world of high-tech. But the true importance of this publication was as a signal to industry that, with appropriate bridging and proper validation, FDA will consider evidence facilitated by ML to be real evidence. In its conclusion, FDA even notes, “Similar approaches could potentially be applied in other situations during drug development, for example, to help with clinical trial patient selection strategies.”
Perhaps most exciting, FDA explains that its scientists even looked at the ML data in a consideration of efficacy: “The CDER team conducted additional exploratory analyses using data from the SAVEMORE trial to evaluate whether the scoring rule could help identify patients at risk for progressing to SRF, and to evaluate the efficacy of anakinra in patients likely to have suPAR levels ≥ 6 ng/mL and worse outcomes (positive for the scoring rule) and patients who were likely to have suPAR levels < 6 ng/mL and better outcomes (negative for the scoring role [sic]).” (Emphasis supplied.)
What This Means for Industry
For many years, industry has been reluctant to submit prediction-based data to FDA, but this publication may be read as an encouragement to industry to do just that. Industry knows that machine learning has enabled use of predictive technologies that, in some instances, produce evidence that proves more reliable than data that is flawed even though generated by the much-vaunted double-blind, placebo-controlled, clinical trial.
Could it be that a new day is dawning at FDA, in CDER and CBER review departments, in the policy shops, and beyond? Even if it never plays a large role in White Oak’s safety and efficacy review halls, maybe one day the evidentiary value of data generated by predictive algorithms and employed through validated processes could substitute for other long-required testing — e.g., in-process sampling and testing under 211.110(a)!
