The pursuit of personalized medicine, tailoring treatments based on a patient’s unique genetic profile, is a crucial goal in healthcare. However, a recent Yale-led study found limitations in current mathematical models used to predict treatments. Analyzing multiple schizophrenia treatment trials, researchers discovered that these algorithms predicted outcomes within specific trials but failed when applied to different ones.
Published in the Science journal on Jan. 11, the study challenges the conventional approach to algorithm development. Adam Chekroud, an adjunct assistant professor of psychiatry at Yale School of Medicine, emphasizes the need for algorithms to prove effective in at least two settings. While optimistic, he acknowledges the challenges faced by medical researchers.
Schizophrenia, affecting about 1% of the U.S. population, highlights the necessity for more personalized treatments. The hope lies in new technologies, like machine learning and artificial intelligence, to develop algorithms better predicting treatment efficacy. However, due to the high cost of clinical trials, most algorithms are tested on a single trial, limiting their effectiveness across different patient profiles and treatments.
Chekroud and his Yale colleagues aggregated data from five schizophrenia treatment trials to assess the generalizability of algorithms. Although the models worked within the trials they were developed for, they failed when applied to patients in different trials. The issue lies in the algorithms being designed for larger datasets, leading to “over-fitting” when applied to smaller clinical trial datasets.
Chekroud emphasizes the need to develop algorithms similarly to new drugs, requiring validation across multiple times or contexts. The study raises questions about the broader application of personalized medicine in cardiovascular disease and cancer. Increased data sharing among researchers and additional data banking by healthcare providers are suggested to enhance the reliability of AI-driven algorithms.