AI predicts intimate partner violence years before disclosure, raising safety concerns

A new artificial intelligence model can detect intimate partner violence (IPV) years before victims ever disclose their abuse, raising both hopes and serious safety concerns. Developed by researchers at Harvard Medical School, MIT, and Brigham and Women’s Hospital, the system called AIRS (Automated IPV Risk Support) analyzes medical records to flag at-risk patients with 80% accuracy, sometimes identifying victims up to five years before they self-report. While the tool promises to close a critical gap in healthcare screening, experts warn that its deployment must be governed by strict ethical standards to avoid unintended harm.

More than one in three women in the United States will experience intimate partner violence in their lifetime, according to the CDC. Many arrive at hospitals with injuries, chronic pain, anxiety, or depression, yet they often wait years to disclose the root cause. Fear of the abuser, financial dependence, immigration status, and stigma prevent many from speaking up. Current screening tools rely on self-reporting, which the CDC estimates captures only a fraction of affected patients. “IPV often remains invisible within healthcare systems despite repeated patient interactions over many years,” said Dr. Bharti Khurana, founding director of the Trauma Imaging Research and Innovation Center at Harvard Medical School and an emergency radiologist at Brigham and Women’s Hospital.

The AIRS model addresses this invisibility by combining two types of data already present in electronic health records. The first is structured data,diagnoses, medications, radiology test locations, emergency visit frequency, vital signs, and a zip-code-level social deprivation score. The second is unstructured clinical notes, which include free-text records from radiologists, social workers, and emergency physicians. These notes are processed using Clinical-Longformer, a clinical language model trained on medical text. Each data stream feeds into its own classifier, and the two are combined using the HAIM framework (Holistic AI in Medicine), which fuses outputs at the prediction stage while keeping the data streams independent. This design accounts for variations in record-keeping across institutions.

On the primary test cohort, the fusion model achieved an AUC of 0.88, where 1.0 is perfect and 0.5 is no better than chance. Across all validation cohorts, including patients from a second hospital and those who never sought help from domestic violence programs, the model held an AUC above 0.8. It flagged 80.6% of IPV cases before the patient self-reported, with an average lead time of 3.68 years. Some patients were identified through records older than five years before they disclosed.

This is not the first attempt to use AI for IPV detection. Spain launched VioGén in 2007, an algorithmic risk assessment tool used by police and judges. However, at least 247 women have been killed by partners after being assessed by VioGén, with 55 scored as negligible or low risk. In the UK, the DASH questionnaire-based tool has been criticized for failing to identify the most vulnerable victims, with machine learning models significantly outperforming it. Other efforts, including NLP algorithms screening ER visits and deep learning applied to police narratives in Australia, have remained largely in pilot stages. Self-reporting remains the standard for legal and medical intervention.

AIRS is designed as a silent clinical support tool, surfacing a risk score only to the clinician, not the patient. “AIRS is not designed to diagnose IPV,” Khurana clarified. “A positive flag should never be interpreted as confirmation of abuse.” The model is currently in active pilot testing at Mass General Brigham. Alongside it, the team developed “Caring Conversations” guidance and “Empower Guides” to help clinicians respond in a trauma-informed way.

Despite these safeguards, ethical concerns remain. Alexia Maddox, senior lecturer at La Trobe University and co-chair of an IEEE Industry Connections Activity on AI for addressing family violence, noted that the study defines IPV narrowly as violence by current or former partners. However, the field has moved toward recognizing coercive control,emotional, financial, and technology-facilitated abuse,as a critical risk factor. “Coercive control, financial abuse, and technology-facilitated abuse leave almost no trace in a radiology report or an emergency department note,” Maddox said. “These prevalent forms of IPV as currently understood are largely invisible to this model.”

There is also the issue of consent. The model was built on six years of medical records, but when deployed, it will generate risk scores for patients who may not know they are being assessed. For women on temporary visas, a database linking them to IPV risk scores could carry serious implications in the current political climate. Maddox invoked VioGén as a cautionary tale: “The history of transformative technology suggests that the people closest to a technical breakthrough are rarely best positioned to see its consequences in the world. That is precisely why governance standards need to exist before deployment, not emerge from the wreckage after.”

Maddox was careful to distinguish these concerns from criticism of the research itself. “The limitations I’ve described are not really the fault of the authors,” she said. “They have done what they set out to do, and done it well.” The more pressing question, she noted, is about the gap between what a study demonstrates and what real-world applications it justifies.

Khurana’s team is already expanding the model’s scope. AIRS was trained on data from female patients, but the team is working to include queer, trans, and male survivors. “IPV affects individuals across all gender identities,” Khurana said. With NIH funding, the team is also studying long-term health risks associated with IPV, including gastrointestinal disorders, neurological conditions, and substance use disorders.

Meanwhile, the IEEE initiative Maddox co-chairs is focused on governance standards for tools like AIRS. “AI systems built to address family violence need to be sensitive to coercive control in its full range of expressions, across cultural contexts, with community knowledge built in from the start,” Maddox said. “They should be governed by frameworks that center the safety, autonomy, and dignity of the women it is designed to serve.” She admitted that the model meeting that bar does not fully exist yet, which is why building these standards will be essential to future research.