Using machine learning to identify individuals at risk for intimate partner violence

Researchers at Mass General Brigham have developed a series of artificial intelligence (AI) tools that uses machine learning to identify individuals who may be at risk for intimate partner violence (IPV) using information from their electronic medical records (EMRs). In a study published in npj Women's Health, the researchers report the tools could detect IPV up to four years before the individual sought care at a domestic violence treatment center. The findings highlight its potential for proactive screening and supporting healthcare providers in initiating earlier conversations about IPV with patients.

Our research offers proof of concept that AI can support clinicians in flagging possible abuse earlier. Earlier identification of intimate partner violence and future risk may enable clinicians to intervene sooner and help prevent significant mental and physical health consequences."

Bharti Khurana, MD, MBA, principal investigator, corresponding and senior author, founding director of the Trauma Imaging Research and Innovation Center and emergency radiologist, Mass General Brigham Department of Radiology

More than one-third of women and 1 in 10 men will experience IPV in their lifetimes; yet, despite its high prevalence, people rarely disclose IPV to health providers due to fear, stigma, or financial or psychosocial dependence on the person abusing them. Prior research shows that people experiencing IPV are more likely to disclose abuse if asked privately by a trusted health provider in a trauma-informed manner.

To promote earlier identification and intervention by health providers, Khurana's research team, working with collaborators at Massachusetts Institute of Technology (MIT) led by Dimitris Bertsimas, PhD, trained three machine-learning models using 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, as well as 4,169 demographically matched controls who did not report IPV.

The three AI models tested included a tabular model using structured EMR data such as diagnoses, medications and social deprivation index based on zip code; a notes model using unstructured clinical notes, and radiology and emergency department reports; and a fusion model combining both data types called Holistic AI in Medicine (HAIM).

When tested on 168 patients who visited the IPV intervention and prevention center in the same timeframe and 1,043 controls, all three models displayed high accuracy, with the fusion model achieving the highest (88%). When tested with time-stamped, archived medical records, the fusion model could predict 80.5% of cases in advance-on average more than 3.7 years before patients sought care.

The models were then validated with data from two additional patient groups that weren't included in the training or testing data, and controls, finding similarly high accuracies.

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 for IPV: People with mental health disorders, chronic pain, and frequent emergency department visits were more likely to experience IPV, whereas patients who regularly accessed preventive services like mammograms and immunizations had a lower risk.

The authors note the AI tools were developed and validated in patients who had sought care for or disclosed IPV, which may limit accuracy in predicting IPV in individuals who are less likely to seek care or disclose IPV to providers. In addition, the control group in the training data may have included false negatives, or patients who were experiencing IPV but did not report it, which could reduce model accuracy. Future training with larger, more diverse patient datasets over longer time periods will improve its accuracy, noted Khurana.