8,532 AI clinical trials reveal shift from diagnosis to risk prediction

A new analysis of 8,532 AI clinical trials across 32 medical specialties, conducted by Lior Rokach and published on arXiv, provides the most detailed map yet of where artificial intelligence is actually being tested in healthcare. The study, which uses an LLM-based classifier to categorize every registered trial on ClinicalTrials.gov, reveals that 80% of AI trials were registered from 2019 onward, and 30.5% use randomized controlled designs. Imaging-based AI remains the largest category with 2,475 trials (29%), but clinical text and NLP trials grew seven-fold between 2018 and 2025, indicating a rapid expansion into unstructured clinical data.

Perhaps most telling, prognostic AI (4,324 trials) has slightly surpassed diagnostic AI (3,828 trials), suggesting the field is shifting from spotting diseases to predicting their trajectory—a move toward risk stratification and early intervention. Yet the pipeline still struggles with translational maturity: 38% of trials are retrospective validation studies and 21% are silent prospective evaluations, meaning most produce algorithmic rather than clinical evidence. Only 184 trials involve Level 4 semi-autonomous or closed-loop AI, with 68% focused on glucose management. Multimodal AI accounts for 33.6% of trials, combining imaging, omics, physiological signals, and wearables. The takeaway: clinical AI has left the lab and entered real-world testing, but it has not yet reached the scale, diversity, or outcome-focused rigor needed to prove its value at the bedside.