April 28, 2026

AI tool predicts patients at risk of intimate partner violence

Image

The tool could help clinicians more quickly connect patients with resources and support.   

Fizkes / Shutterstock

Each year millions of people in the United States experience intimate partner violence, or IPV. IPV refers to abuse from current or former spouses and partners. It can lead to life-threatening injuries, chronic pain, and mental health conditions. Many people who experience IPV do not tell health providers because of safety concerns, fear, and stigma.

Current screening tools identify only a fraction of IPV cases and often rely on patients to self-report. Recognizing IPV cases early allows for timely intervention to prevent long-term health consequences. 

An NIH-funded research team, led by Dr. Bharti Khurana of Mass General Brigham, developed and tested an artificial intelligence (AI) tool to predict patients at risk of IPV. A description and evaluation of the tool was published March 13, 2026, in npj Women’s Health.

The researchers used a type of AI called machine learning to develop three computer models that could predict IPV. They developed the models using electronic medical records from 841 patients enrolled in a domestic abuse intervention and prevention center. Records from another 5,212 non-IPV patients of similar ages and backgrounds were included for comparison. One model used structured patient data in tables, while the second used unstructured patient data from medical notes. The third model incorporated both data types.

The researchers used 80% of the patient data to train the models and then tested their accuracy on the remaining 20%. All three models were more than 80% accurate, with the combination model being the most accurate at 88%. On average, both the table and combination models could detect IPV risk more than three years before a patient sought help for IPV. 

The team confirmed the accuracy of the models in three other patient groups. The combination model continued to perform best, with accuracy between 82-88%.

Mental health conditions, chest pain, and painkiller use were linked to higher IPV risk, according to the models. So were high social deprivation and frequent radiology tests. Patients who regularly sought preventive services like mammograms and cervical cancer screenings had a lower risk of IPV. The researchers suggested this may be because patients who receive regular screenings have better access to health care and are more comfortable seeking medical care.

The researchers noted that the model should be evaluated in more general populations before it is used in clinical settings. They also emphasized that the model is not intended to diagnose IPV. Rather, it’s designed to help health care providers identify patients who may benefit from discussions about IPV and IPV support resources.

“By analyzing patterns already present in health care data, this approach supports health care clinicians in initiating earlier, safer, and more informed conversations with patients,” Khurana says. “The goal is never to force disclosure, but to help clinicians communicate with patients in a supportive way and to connect them with resources and support.”

Related Links

• Machine learning analysis of CT scans
• AI may aid screening for opioid use disorder
• AI tool predicts response to cancer therapy
• Machine learning approach detects brain tumor boundaries
• Artificial intelligence predicts heart disease risk from CT scans
• Machine learning detects early signs of osteoarthritis
• Harmful partnerships
• Relationships and safety (Office on Women's Health)
• Artificial intelligence (AI)
• About intimate partner violence

References

Leveraging multimodal machine learning for accurate risk identification of intimate partner violence. Gu J, Carballo KV, Ma Y, Bertsimas D, Khurana B. NPJ Womens Health. 2026;4(1):15. doi: 10.1038/s44294-025-00126-3. Epub 2026 Mar 13. PMID: 41836047.

Funding

NIH’s National Institute of Biomedical Imaging and Bioengineering (NIBIB), Office of the Director (OD).