AI System Unlocks Clinical Notes for Large-Scale Medical Research

For decades, medical research has largely relied on structured billing codes and lab values, ignoring the rich narrative contained in clinicians' notes. A team led by RespondHealth, in collaboration with Drexel University and other institutions, developed a system capable of extracting facts from these notes at scale. To ensure reliability, the researchers utilized a human-in-the-loop validation process where board-certified physicians reviewed the AI output against original records. The system achieved a 99.4 percent accuracy rate, significantly outpacing the speed of human manual review while maintaining full traceability to the source text.

The researchers applied this technology to track over 16,000 patients prescribed GLP-1 medications, such as semaglutide. The analysis revealed distinct patterns in weight loss and blood sugar improvement that were previously obscured. Notably, 70 percent of the blood sugar readings used in the study were found exclusively in written notes rather than database fields. By leveraging this narrative data, the team observed that patients with healthier baseline blood sugar lost weight more efficiently, while those with poorer control saw faster improvements in their blood sugar levels. Beyond metabolic health, the system also captured improvements in depression scores and pain intensity—metrics rarely captured by standardized billing codes. This approach offers a path to broader real-world evidence gathering, as the method is applicable to any medical condition documented in clinical prose.