Medical records already contain the clues to detect domestic violence - AI can read them

More than one-third of women and one in ten men will experience intimate partner violence in their lifetimes. Most will never tell their doctor. The reasons are familiar and devastating: fear of retaliation, financial dependence on the abuser, shame, and the well-founded concern that disclosure might make things worse rather than better.

But the medical record keeps its own account. Patterns emerge over time in diagnoses, emergency department visits, injury types, medication histories, and the language clinicians use in their notes. A team at Mass General Brigham, working with collaborators at MIT, asked whether machine learning could detect these patterns early enough to change outcomes.

Training on real patient histories

The researchers trained three machine learning models on EMR data from 673 women who visited a domestic abuse intervention and prevention center at a U.S. academic health center between 2017 and 2022, alongside 4,169 demographically matched controls. The models were then tested on a separate cohort of 168 patients and 1,043 controls, and further validated on two additional patient groups not used in training or testing.

The three models took different approaches to the same data. A tabular model used only structured fields - diagnosis codes, medications, social deprivation index. A notes model worked exclusively with unstructured text from clinical notes, radiology reports, and emergency department documentation. A fusion model called HAIM combined both.

The fusion model achieved 88% accuracy on the test set and predicted 80.5% of cases in advance, on average more than 3.7 years before patients sought care at the domestic violence center. The tabular model flagged risk slightly earlier in some cases, while the fusion model caught more cases overall.

The signals hiding in plain sight

The study identified clinical risk factors that, in retrospect, make intuitive sense. People with mental health disorders, chronic pain conditions, and frequent emergency department visits were more likely to be experiencing IPV. Patients who regularly accessed preventive care - mammograms, immunizations, routine screenings - had lower risk, suggesting that consistent engagement with the healthcare system may itself be a marker of relative safety or stability.

Bharti Khurana, an emergency radiologist at Mass General Brigham and the study's senior author, has spent years studying the radiological signatures of IPV. Her previous work showed that women who undergo frequent imaging studies in the emergency department with certain injury patterns are at elevated risk. The new AI models incorporated these radiological signals alongside dozens of other data points, creating a more comprehensive picture than any single indicator could provide.

From reactive to proactive

Khurana described the AI tool as representing a fundamental shift in how healthcare systems approach IPV - from waiting for patients to disclose abuse to proactively recognizing risk patterns already present in existing data. The tool is not intended to make diagnoses or substitute for clinical judgment. It is designed to prompt clinicians to have conversations they might not otherwise initiate.

The research team developed guidance materials to help clinicians approach these conversations thoughtfully and safely. The emphasis is on creating conditions where patients feel supported enough to engage, not on pressuring disclosure.

Where the model falls short

The study has important limitations. The training data came from patients who had already sought care at a domestic violence center, which creates a selection bias. The AI may perform differently for people experiencing violence who have not yet engaged with support services. The control group could include undisclosed IPV cases, which would dilute the model's apparent accuracy.

All training and validation data came from a single academic health center. Performance at community hospitals, rural clinics, or healthcare systems with different documentation practices is unknown. The study focused on female patients, and applicability to male patients or nonbinary individuals has not been evaluated.

The fusion model requires access to both structured data and unstructured clinical notes, which may not be uniformly available or consistently documented across healthcare settings. The tabular model, which uses only structured data, may be more practical for systems with less robust documentation practices - though at the cost of detecting somewhat fewer cases.

The next step is embedding the tool directly into electronic medical record systems for real-time clinical use. Khurana's team plans to expand training on larger, more diverse datasets and assess how the tool performs across different healthcare environments and patient populations.