NIH-Funded AI Predicts Intimate Partner Violence Risk Years Before Disclosure

A new artificial intelligence tool funded by the National Institutes of Health can predict which patients are likely to experience intimate partner violence years before they seek help. The research, published March 13, 2026, in npj Women's Health, represents a shift from reactive screening to proactive risk identification in clinical settings.

Dr. Bharti Khurana of Mass General Brigham led the development of three machine learning models trained on electronic medical records. The team used data from 841 patients enrolled in a domestic abuse intervention center, plus 5,212 non-IPV patients for comparison. All three models exceeded 80% accuracy, with the multimodal combination model reaching 88%.

On average, the models detected IPV risk more than three years before patients sought help for abuse. That detection window matters. In clinical practice, a physician reviewing a patient chart might notice patterns—frequent emergency department visits, radiology reports documenting injuries, prescriptions for pain medication—but connecting those dots requires time and attention that busy clinics rarely have. The AI automates that pattern recognition.

The research team designed distinct approaches to handle different data types. One model processed structured patient data in tables—diagnoses, medications, social deprivation indices based on zip code. A second model analyzed unstructured data from clinical notes, radiology reports, and emergency department documentation. The third, called Holistic AI in Medicine (HAIM), fused both modalities. The fusion model proved most stable and accurate across validation groups.

According to the NIH research summary, the models identified several risk factors beyond what traditional screening captures. Mental health conditions, chest pain, and painkiller use correlated with higher IPV risk. So did high social deprivation and frequent radiology tests. Patients who regularly accessed preventive services like mammograms and cervical cancer screenings showed lower risk—possibly because they have better healthcare access and feel more comfortable seeking medical care.

Current screening tools identify only a fraction of IPV cases. They rely heavily on patient self-disclosure, which many people avoid due to safety concerns, fear, and stigma. Millions of people in the United States experience intimate partner violence annually, yet most cases go undetected in healthcare settings. The AI tool doesn't replace human judgment—it flags patterns that might otherwise remain invisible during routine visits.

Mass General Brigham's press release confirms the validation process. The team tested the models on three additional patient groups not included in training data. The combination model maintained accuracy between 82-88% across these cohorts. That consistency suggests the approach could generalize beyond the initial study population, though broader validation remains necessary.

The researchers emphasized important limitations. The model should be evaluated in more general populations before clinical deployment. It is not intended to diagnose IPV. Rather, it helps healthcare providers identify patients who may benefit from discussions about IPV and support resources. The goal is never to force disclosure, but to help clinicians communicate with patients in a supportive way and connect them with resources.

Dr. Qi Duan, director of the Division of Health Informatics Technologies at NIH's National Institute of Biomedical Imaging and Bioengineering, called the tool a potential game-changing asset to public health. The research was co-funded by NIBIB grant R01EB032384 and the NIH Office of the Director. That funding structure signals institutional commitment to translating the research into practical clinical tools.

Implementation would require embedding the AI in electronic medical record systems to provide real-time risk evaluations. Clinicians would see alerts during patient visits—perhaps a subtle flag in the sidebar of their EMR interface, or a risk score displayed alongside vital signs. The physical experience matters: a doctor scrolling through a patient's history might miss subtle patterns, but an automated alert draws attention without requiring extra clicks or navigation.

The team developed guidance at their project website to help clinicians approach conversations with patients thoughtfully. They plan to use the AI models to develop a decision-support tool embedded in EMR systems. That integration work is where most medical AI projects stall—technical accuracy means little if the tool doesn't fit into existing clinical workflows.

Khurana noted that the control group in training data may have included false negatives—patients experiencing IPV who did not report it. Future training with larger, more diverse datasets over longer time periods will improve accuracy. The models were developed and validated in patients who had sought care for or disclosed IPV, which may limit accuracy in predicting IPV among individuals less likely to seek care.

Previous research led by Khurana found that women who frequently undergo imaging studies at the emergency department and have specific types of injuries are more likely to later report IPV. This new AI research identified additional risk factors, expanding the clinical picture beyond what radiologists alone might recognize.

The technology represents a fundamental shift from reactive disclosure to proactive risk recognition within routine clinical care. By analyzing patterns already present in healthcare data, the approach supports clinicians in initiating earlier, safer, and more informed conversations with patients. That's the stated goal. Whether healthcare systems actually adopt it—and whether patients trust it—remains the real question.

Whether patients actually pay for it remains the real question. No, that's not right—patients don't pay for this. Whether healthcare systems actually adopt it, and whether patients trust AI flagging them for abuse risk, is what matters. The technology works in controlled studies. Clinical reality is messier.

Artūras Malašauskas is an AI Systems Integrator with 20+ years of production-grade web engineering experience. He has designed, shipped, and scaled enterprise Python/PHP systems for logistics, SaaS, and public-sector clients. For the past year, he has focused exclusively on AI integrations: deploying open-source LLMs, building generative media pipelines (image, audio, video), and engineering multi-agent workflows for real production environments. His standard: reproducibility, security, cost-efficient inference—no vaporware. He documents and evaluates emerging AI tooling, separating verified capabilities from marketing noise. Technical editor at: muza-ai.eu, ai-verslas.lt, ai-naujinos.lt Connect on LinkedIn