AI helped research staff review cancer treatment side effects faster

Checking medical notes for a treatment side effect involves more than finding a word. Reviewers may need to establish whether a problem is current or past, how serious it is, and whether the note links it to treatment. Those details are often buried in text. Public manuscript

Researchers tested an AI system that suggested labels and highlighted passages for people to check. In a randomized crossover experiment, 17 clinical research staff each reviewed 20 notes, using AI assistance for one set and manual review for the other. The experiment used 20 unique notes overall. Final paper

Median review time fell from 428 seconds without AI to 242 seconds with it, a 43.5% reduction. The timing analysis included 316 individual reviews after 24 reviews with missing video recordings were excluded. Those are repeated reviews by staff, not 316 different patients or notes. Timing analysis and Methods

The accuracy findings were mixed, so faster review should not be read as uniformly more accurate review. Accuracy results

The researchers separately tested the system on 884 newly arriving notes over three months. All phases took place within the same healthcare system. The work has not established that it transfers to other institutions or improves patient outcomes. Prospective evaluation

The March preprint also identifies a risk with more consistent review: people may follow the same AI suggestions, including mistakes. The study could not separate useful standardization from overreliance on the model. Preprint limitations

The peer-reviewed paper appeared on October 29. A March preprint already described the work. The evidence supports a quicker review task with people checking the source notes. Whether that leads to better care remains untested. October 29 paper · March preprint