Researchers Develop AI Tool to Predict Patients at Risk of Intimate Partner Violence

A groundbreaking artificial intelligence (AI) tool, developed by a team of researchers funded by the National Institutes of Health (NIH), promises to revolutionize the early detection and intervention of intimate partner violence (IPV). This automated clinical decision support system is designed to predict a patient’s risk of IPV years before they might otherwise seek help, facilitating timely interventions and potentially saving lives. Using routinely collected medical data, the team trained a sophisticated machine-learning model, demonstrating high accuracy in identifying IPV risk among patients in a comprehensive study. This innovation represents a significant shift from reactive identification to proactive risk recognition within routine clinical care, offering healthcare providers an unprecedented opportunity to address a pervasive public health crisis more effectively.

The Unseen Epidemic: Understanding Intimate Partner Violence

Intimate partner violence is a severe global public health problem that affects millions of individuals across all demographics, irrespective of gender, age, socioeconomic status, or cultural background. It encompasses physical violence, sexual violence, stalking, and psychological aggression by a current or former intimate partner. The Centers for Disease Control and Prevention (CDC) reports that approximately 1 in 4 women and 1 in 10 men in the United States have experienced sexual violence, physical violence, and/or stalking by an intimate partner during their lifetime. These figures underscore the widespread nature of IPV, which often remains hidden due to fear, stigma, safety concerns, and a lack of awareness regarding available support.

The consequences of IPV are profound and far-reaching, extending beyond immediate physical injuries to long-term chronic health issues. Victims frequently suffer from potentially life-threatening injuries, chronic pain, gastrointestinal disorders, neurological symptoms, and a host of mental health disorders including depression, anxiety, post-traumatic stress disorder (PTSD), and substance abuse. The economic burden of IPV is also staggering, encompassing direct healthcare costs, lost productivity, and criminal justice expenses, estimated to be billions of dollars annually in the U.S. alone. Despite its prevalence and devastating impact, many cases of IPV go undetected within healthcare settings, leading to missed opportunities for timely intervention and support. Current screening tools, often relying on direct patient disclosure, capture only a fraction of cases, as patients may be hesitant or unable to reveal their abusive situations for various complex reasons. This critical gap in identification has long been a major challenge for healthcare providers striving to offer comprehensive care.

Pioneering a Proactive Approach: The AI Innovation

Recognizing the limitations of existing methods, the research team, primarily led by experts from Harvard Medical School in Boston, embarked on developing an AI-powered solution. Their objective was to leverage the vast amounts of data routinely collected during medical visits to identify subtle patterns indicative of IPV risk, thereby shifting the paradigm from reactive disclosure to proactive recognition. The study introduced three distinct AI models designed for IPV detection in healthcare settings, rigorously comparing their performance in predicting the risk.

A crucial insight driving this research was the observation that clinical and imaging records often contain valuable, albeit indirect, information for detecting IPV risk. Notably, radiologists possess a unique advantage in recognizing the signs of IPV, as they frequently encounter and document specific patterns of physical trauma, such as multiple fractures at different stages of healing, specific types of soft tissue injuries, or injuries inconsistent with the reported mechanism. These patterns, when viewed collectively across a patient’s medical history, can serve as powerful indicators that might otherwise be overlooked in isolated clinical encounters.

To train their models, the researchers utilized several years of de-identified hospital data. This extensive dataset included records from nearly 850 affected female patients and 5,200 unaffected control patients, carefully matched for age and demographics. Understanding that the collection of relevant clinical data can vary significantly across different healthcare settings, the team ingeniously designed two distinct initial AI models. One model was trained exclusively on structured patient data, typically presented in table form within electronic health records (EHRs), such as demographic information, diagnostic codes, medication lists, and lab results. The second model was trained on unstructured patient data, primarily derived from free-text medical notes, including detailed physician observations, nurse’s notes, and, crucially, radiology reports. Furthermore, to harness the full power of both data types, they developed a sophisticated multimodal fusion model, which intelligently combined insights from both structured and unstructured data streams.

The Technology Behind the Prediction

The results of the study, published in the Nature Portfolio Journal: Women’s Health, demonstrated the remarkable efficacy of all developed models. However, the multimodal fusion model emerged as the clear frontrunner, consistently outperforming the models that relied solely on either structured or unstructured data. This advanced model achieved an impressive accuracy rate of 88% in predicting IPV risk. What makes this achievement particularly impactful is the timing of detection: both the tabular model and the fusion model were capable of detecting IPV risk, on average, more than three years before patients would typically enroll at hospital-based domestic abuse intervention centers. While the tabular model showed a slightly earlier recognition of IPV risk in some instances, the fusion model demonstrated a superior ability to detect a greater number of IPV cases in advance, indicating its broader utility and higher sensitivity.

The enhanced performance of the fusion model underscores the synergistic benefits of combining diverse data modalities. The scientists explained that by processing different data types separately and merging them only at the prediction stage, the model could achieve more stable and robust performance than relying on any single modality. This design also accounts for the practical realities of healthcare data, where the availability and standardization of unstructured data can vary significantly between hospitals. The tabular framework, in particular, offers a degree of resilience against such variations, making the tool potentially adaptable across a wider range of clinical environments.

It is crucial to emphasize that these AI tools are intended as clinical decision support systems, not as definitive diagnostic instruments. Their purpose is to assist healthcare providers by flagging patients at elevated risk, thereby prompting timely and sensitive conversations about IPV and facilitating connections to appropriate support resources. Dr. Bhati Khurana, M.D., senior author of the study, an emergency radiologist at Mass General Brigham and associate professor of radiology at Harvard Medical School, articulated this shift: “For decades, our healthcare system has depended largely on patient self-disclosure to identify intimate partner violence, leaving many cases unrecognized and unsupported. Our work represents a fundamental shift from reactive disclosure to proactive risk recognition within routine clinical care. By analyzing patterns already present in healthcare data, this approach supports healthcare clinicians in initiating earlier, safer, and more informed conversations with patients.”

Expert Perspectives and Ethical Considerations

The development of this AI tool has been met with significant enthusiasm from within the medical community and funding institutions. Dr. Qi Duan, Ph.D., director of the Division of Health Informatics Technologies at NIH’s National Institute of Biomedical Imaging and Bioengineering (NIBIB), highlighted the transformative potential: “This clinical decision support tool could make a significant impact on prediction and prevention of intimate partner violence. Given the prevalence of cases, the tool could be a game-changing asset to public health.” This sentiment reflects the NIH’s commitment to funding innovative research that addresses critical public health challenges through technological advancements.

The ethical implementation of such a sensitive AI tool is paramount. The researchers have been meticulous in developing guidance, available on their project website, to help clinicians thoughtfully approach conversations with patients identified as at-risk. Dr. Khurana underscored this patient-centered philosophy: “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.” This approach prioritizes patient autonomy, safety, and well-being, ensuring that the technology serves as an enabler for compassionate care rather than a mechanism for coercive intervention. Discussions around data privacy, potential biases in AI algorithms (e.g., if the training data disproportionately represents certain demographics or types of violence), and the need for continuous monitoring and refinement of the models are integral to the responsible deployment of this technology. The design ensures that the AI acts as a prompt for human interaction, maintaining the essential "human-in-the-loop" oversight that is critical in high-stakes healthcare scenarios.

Implications for Healthcare and Public Health

The implications of this AI tool for healthcare and public health are vast and potentially transformative. By enabling the identification of IPV risk years in advance, it offers an unprecedented window for intervention that could significantly improve long-term health outcomes for at-risk patients. Early intervention can lead to a reduction in the incidence and severity of physical injuries, mitigate the development or worsening of chronic pain and mental health disorders, and ultimately help individuals break free from cycles of violence. This proactive stance aligns with broader public health goals of prevention and early detection for optimal health.

From a healthcare system perspective, the tool holds the promise of increased efficiency and reduced costs. By intervening earlier, healthcare providers may avert the need for more intensive and costly treatments associated with advanced stages of IPV-related trauma and chronic conditions. The research team’s future plans involve integrating these AI models into electronic medical record (EMR) systems to provide real-time IPV risk evaluations in clinical settings. Such integration would seamlessly embed this powerful decision support directly into the clinical workflow, making it a routine part of patient care rather than an additional, separate screening process. This would require robust training programs for healthcare professionals on how to utilize the tool effectively, how to conduct sensitive conversations with patients, and how to navigate the ethical considerations inherent in this technology.

The success of this research could also influence policy changes, potentially leading to updated guidelines for IPV screening in various medical specialties. It could encourage greater collaboration between healthcare providers, social workers, and community-based domestic violence support organizations, creating a more cohesive and effective network of support for victims. The tool’s ability to detect risk across diverse data modalities also suggests its potential applicability in various healthcare settings, from emergency departments to primary care clinics, further broadening its impact.

A Call for Comprehensive Support

While the AI tool represents a monumental leap forward in IPV detection, it is crucial to remember that technology alone cannot solve the complex issue of intimate partner violence. It is a powerful assistant that must be coupled with comprehensive support systems. The project website provides valuable resources, and the CDC offers extensive information on intimate partner violence prevention. These resources are vital for clinicians to connect patients with necessary aid, including counseling, legal assistance, safe housing, and community support groups. The ultimate goal is to empower individuals to seek help and provide them with the resources they need to achieve safety and well-being. This innovative technology serves as a catalyst, urging the healthcare system and broader society to adopt a more compassionate, informed, and proactive approach to addressing IPV, ensuring that discovery truly turns into health.

This research was co-funded by NIBIB grant R01EB032384 and the NIH Office of the Director, underscoring the collaborative effort and strategic investment in advancing biomedical technologies for public health benefit. The National Institute of Biomedical Imaging and Bioengineering (NIBIB) continues its mission to improve health by leading the development and accelerating the application of biomedical technologies, integrating physical and engineering sciences with life sciences to advance both basic research and medical care. The National Institutes of Health (NIH), as the nation’s medical research agency, plays a pivotal role in conducting and supporting basic, clinical, and translational medical research, investigating the causes, treatments, and cures for a vast array of diseases.

Reference

Gu J, Villalobos Carballo K, Ma Y, Bertsimas D, and Khurana B. Leveraging multimodal machine learning for accurate risk identification of intimate partner violence. Nature Portfolio Journal: Women’s Health. 2026. DOI: 10.1038/s44294-025-00126-3