In a pioneering advancement that merges artificial intelligence with healthcare, researchers at Mass General Brigham have devised innovative machine learning tools capable of identifying individuals at risk of intimate partner violence (IPV) well before traditional clinical recognition. This breakthrough study, published in the esteemed journal npj Women’s Health, illuminates a critical pathway toward proactive intervention, potentially transforming how medical professionals approach one of the most pervasive public health crises globally.
At the heart of this research is a suite of AI models designed to analyze an extensive array of data extracted from electronic medical records (EMRs). Unlike conventional screening methods, which rely heavily on patient disclosure and overt signs of abuse, these algorithms sift through structured and unstructured medical information, detecting subtle patterns and risk indicators that often precede visible symptoms of IPV. Remarkably, the system demonstrated the ability to predict IPV cases nearly four years before victims engaged with specialized domestic violence treatment centers.
The driving force behind this innovative work is Dr. Bharti Khurana, an emergency radiologist and researcher at Mass General Brigham’s Department of Radiology and Trauma Imaging Research and Innovation Center. Dr. Khurana highlights the potential of these AI tools to aid clinicians in early identification and intervention. By flagging at-risk patients well in advance, healthcare providers could initiate crucial conversations, offer support earlier, and stem the severe mental and physical traumas often associated with IPV.
Intimate partner violence remains a widespread issue, with statistics indicating that over one-third of women and 10% of men will experience IPV during their lifetimes. Despite its high prevalence, many survivors refrain from disclosure within healthcare settings due to fear, social stigma, or complex dependencies on their abusers. Existing research underscores the importance of sensitive, trauma-informed approaches within clinical environments, where trust and privacy significantly influence the likelihood of survivors sharing their experiences.
The interdisciplinary collaboration at the core of this project involved teaming up with experts from the Massachusetts Institute of Technology, led by Dr. Dimitris Bertsimas. Together, the team curated and analyzed EMR data from 673 women who had accessed a domestic abuse intervention center alongside a control group of 4,169 demographically matched individuals who had no reported IPV history. This robust dataset served as the foundation for training and optimizing three distinct machine learning models.
The models developed take advantage of diverse data modalities: a tabular model analyzing well-structured data like diagnoses, prescription histories, and socioeconomic factors derived from geographic indicators; a notes-based model scrutinizing narrative clinical documentation including radiology reports and emergency department notes; and most notably, a fusion model known as Holistic AI in Medicine (HAIM), which integrates both structured and unstructured datasets to enhance predictive accuracy.
Performance evaluation of these algorithms revealed their remarkable potential in a clinical context. On a test cohort of 168 patients with known IPV disclosures and 1,043 control subjects, all models exhibited high accuracy in risk prediction. The fusion HAIM model outperformed other configurations, achieving an impressive 88% accuracy rate. Furthermore, longitudinal assessments using timestamped EMR data demonstrated that the fusion model could identify over 80% of cases well in advance, on average more than 3.7 years before individuals sought specialized IPV care.
Beyond initial testing, the study’s robustness was further validated across two additional patient cohorts, completely independent of the training datasets. This cross-validation effort confirmed that the AI tools maintained their high sensitivity and reliability in diverse clinical populations, underscoring their potential for broader application across healthcare systems.
Earlier investigations led by Dr. Khurana had pinpointed imaging patterns — such as repeated emergency department visits related to specific injuries — correlated with later IPV disclosure. This present work extends those findings by uncovering further health-related risk factors through machine learning. Notably, patients with chronic pain syndromes, mental health disorders, and frequent emergency visits displayed elevated IPV risk profiles, whereas engagement with preventive healthcare services like mammographies and immunizations appeared to be protective factors.
The researchers acknowledge certain limitations within their study. The primary dataset consisted of individuals who had already sought IPV-related care or disclosed abuse, which might reduce sensitivity when applied to populations less open about their experiences due to persistent barriers. Moreover, control group participants likely included some undetected IPV cases, representing false negatives that can attenuate model accuracy. Continual refinement of these AI tools with larger, more heterogeneous datasets over longer durations is planned to enhance predictive capabilities further.
This promising research signifies a paradigm shift in the intersection of technology and violence prevention. By empowering healthcare professionals with advanced AI-driven insights, the initiative aims to facilitate early intervention strategies that could fundamentally reduce IPV’s toll on individual lives and society. The integration of such systems into routine clinical workflows raises the prospect of making covert abuse visible, thereby enabling timely, life-saving support.
The funding for this project was generously provided by the National Institute of Biomedical Imaging and Bioengineering and the National Institutes of Health Office of the Director, highlighting the growing interest and investment in machine learning applications within healthcare domains. As these AI models evolve, there is keen anticipation for their adaptation and deployment within broader medical infrastructures to augment screening protocols and foster a safer, more responsive healthcare ecosystem.
Dr. Bharti Khurana encapsulates the societal relevance of the work by emphasizing that earlier recognition and intervention may help circumvent the profound psychological and physical consequences wrought by intimate partner violence. This pioneering effort stands as a testament to the transformative potential of artificial intelligence in detecting hidden health risks and supporting vulnerable populations through the power of predictive medicine.
Subject of Research: People
Article Title: Leveraging multimodal machine learning for accurate risk identification of intimate partner violence
News Publication Date: 13-Mar-2026
Web References: https://www.nature.com/articles/s44294-025-00126-3
References: Gu et al. “Leveraging multimodal machine learning for accurate risk identification of intimate partner violence” npj Women’s Health, DOI: 10.1038/s44294-025-00126-3
Keywords: Domestic violence, Machine learning, Artificial intelligence, Radiology
