AI Reads Clinical Notes to Extract Uncoded Medical Data

A collaborative research initiative led by RespondHealth, in partnership with Drexel University, Stanford University, the University of Miami, the University of Pennsylvania, and the Icahn School of Medicine at Mount Sinai, has demonstrated a novel artificial intelligence system capable of accurately reading and extracting data from unstructured clinical notes at scale. Published in Nature Medicine, the study addresses a longstanding limitation in medical research: the vast majority of patient information resides in narrative physician notes rather than structured electronic health record fields, leaving it largely inaccessible for large-scale analysis. The AI platform converts prose-based documentation into computable datasets while maintaining strict traceability, linking every extracted data point to its original source sentence. To validate performance, board-certified physicians conducted line-by-line reviews of the system output against source charts across multiple electronic health record platforms. The AI achieved a 99.4 percent accuracy rate on combined precision and recall metrics, surpassing the 94.7 percent inter-rater agreement typically observed among human clinicians. Additionally, the system processes individual charts in seconds, whereas manual review required approximately seventy hours for just one hundred twenty cases. Researchers applied the technology to a real-world observational cohort of more than sixteen thousand adults initiating GLP-1 receptor agonist therapy for diabetes or weight management. The analysis revealed distinct clinical patterns. Patients entering treatment with normal baseline blood sugar achieved faster and greater weight loss, shedding approximately 7.7 percent of their body weight within twelve months, compared to 2.7 percent among those with poorly controlled diabetes. Conversely, individuals with the worst initial glycemic control demonstrated the most rapid improvements in blood sugar levels. Demographic factors also influenced outcomes, with women and patients aged twenty to thirty-nine experiencing significantly higher weight reduction than their male or older counterparts. Beyond standard metrics, the system uncovered clinically relevant data previously obscured by billing code limitations. Depressive symptom scores showed the greatest improvement among patients with baseline moderate to severe depression, while reported pain intensity and waist circumference also declined meaningfully. Crucially, seventy percent of the glycemic readings and one-third of the medication adherence records used in the study existed exclusively in narrative text. Relying solely on structured fields would have excluded thousands of patients from analysis. Researchers emphasize that the findings are observational and intended to highlight real-world medication behavior rather than establish causal relationships. The primary advancement lies in the system transparency, which allows clinicians to verify extracted facts directly against source documentation. By transforming unstructured clinical narratives into auditable, computable evidence, the technology offers a scalable pathway for generating robust real-world data across virtually all medical specialties, fundamentally expanding the scope of post-market drug monitoring and population health research.