Overview
- Mass General Brigham researchers with MIT collaborators built three EMR-based machine-learning models, including a Holistic AI in Medicine (HAIM) fusion approach that combines structured and unstructured data.
- In held-out testing, the fusion model reached 88% accuracy and, on archived time-stamped records, identified 80.5% of future cases on average more than 3.7 years before patients sought specialized care.
- Training drew on EMRs from 673 women who visited a domestic abuse intervention center and 4,169 demographically matched controls, with validation on two additional patient groups showing similarly high accuracy.
- Patterns associated with higher risk included mental health disorders, chronic pain, and frequent emergency department visits, while consistent preventive care such as mammograms and immunizations correlated with lower risk.
- Authors caution that models were built on patients who disclosed or sought care and that controls may include unreported cases, calling for larger, more diverse, prospective studies and trauma‑informed, privacy‑conscious workflows, with plans to explore EMR-embedded decision support.