Recent innovations from a multidisciplinary research team supported by the National Institutes of Health (NIH) are set to revolutionize clinical approaches to intimate partner violence (IPV). The team has developed an advanced machine learning tool that markedly improves the early identification of IPV risk among patients. This pioneering artificial intelligence (AI) model utilizes routine healthcare data to proactively pinpoint individuals at risk—an advancement that holds significant promise for enhancing public health and ensuring patient safety.
Intimate partner violence, which encompasses abuse from current or former partners, is a widespread and troubling public health issue in the United States, affecting millions of people regardless of gender. The repercussions of IPV can lead to serious physical injuries, chronic pain, and severe mental health problems. Despite its prevalence, IPV often goes undetected in clinical settings, primarily due to patients’ hesitance to reveal their experiences. Factors such as fear, stigma, and worries about personal safety contribute to this silence, highlighting the urgent need for improved diagnostic tools.
To tackle this issue, the research team—led by Harvard Medical School and joined by specialists in biomedical imaging and clinical informatics—designed and rigorously validated three AI-driven models intended to detect IPV risk using various types of medical data. These models were trained on both structured data, such as demographic and clinical variables presented in table formats, and unstructured data sourced from narrative medical notes, including radiology reports. A novel fusion approach that integrates both data types during the predictive phase was also employed.
Interestingly, the multimodal model outperformed the others, achieving an impressive accuracy of 88% in predicting IPV risk from data obtained from nearly 6,000 female patient cases. This hybrid model surpassed its counterparts that relied solely on either tabular or narrative data, emphasizing the value of combining quantitative and qualitative streams from healthcare information. Additionally, the study underscores the importance of radiological imaging as an effective diagnostic tool, as radiologists are skilled at recognizing injury patterns associated with IPV.
A remarkable feature of this technology is its early identification capabilities: both the tabular data model and the fusion model could predict IPV risk more than three years before patients enrolled in hospital-based domestic violence intervention programs. While the tabular model had a slight edge in earlier detection, the integrated multimodal model excelled at identifying a greater number of at-risk individuals overall, providing a more comprehensive screening mechanism.
From a technical perspective, the framework processes structured and unstructured inputs separately before blending predictions, resulting in robust performance across diverse clinical datasets from various healthcare settings. This approach addresses challenges related to the inconsistent availability and documentation of unstructured clinical data, allowing for adaptability and scalability within the intricate landscape of medical institutions.
The integration of such AI tools into electronic health record (EHR) systems presents transformative possibilities. Real-time risk assessments can notify healthcare providers during routine visits, encouraging timely and sensitive discussions with patients who may be vulnerable yet reluctant to disclose abuse. Importantly, the tool serves as a decision support system rather than a definitive diagnostic authority, assisting clinicians in navigating sensitive conversations and connecting patients to appropriate community resources without pressuring them.
Dr. Bhati Khurana, the senior author and an emergency radiologist at Mass General Brigham as well as an associate professor at Harvard Medical School, emphasizes that this technology represents a paradigm shift. Traditional IPV screening has largely depended on patient disclosures, often proving reactive and inadequate. This AI-driven proactive method taps into existing clinical data patterns, fostering earlier interventions that can mitigate chronic harm and effectively disrupt the cycle of violence.
Besides its clinical significance, the research team has put considerable effort into ethical implementation. They have developed comprehensive guidelines for clinicians on how to engage with patients using insights derived from the tool. These guidelines underline patient-centered communication strategies that prioritize safety, confidentiality, and support, ensuring that the technology facilitates compassionate care.
Moving forward, the research aims to integrate these AI models with clinical decision support systems on a scalable level, allowing frontline healthcare workers to seamlessly incorporate machine intelligence into their daily practices. This integration not only promises to enhance IPV risk detection but also aims to involve predictive analytics in broader frameworks for assessing complex health and social challenges.
The study’s methodological rigor is evident in its extensive dataset, which encompasses 850 IPV-affected female patients matched against 5,200 controls based on clinical and demographic metrics. The application of multimodal machine learning represents a significant milestone, transforming intricate clinical data into actionable insights and showcasing AI’s potential to address deeply rooted societal concerns within healthcare.
Moreover, the technological framework considers the variability in hospital data systems, an essential factor given the differing healthcare infrastructures across regions. This flexibility ensures the tool’s accuracy and functionality, even in environments where unstructured clinical narratives may be poorly documented, thereby minimizing obstacles to broader adoption.
Ultimately, this innovation marks the beginning of a transformative chapter in public health and clinical practice, where predictive analytics enable healthcare providers to act preemptively. By advancing beyond conventional reactive methods, this AI-driven tool could significantly lessen the incidence and repercussions of IPV, enhancing the safety and well-being of millions affected by this often-hidden epidemic.
The research was co-funded by the NIH National Institute of Biomedical Imaging and Bioengineering and the NIH Office of the Director, underscoring its strategic importance in federal initiatives aimed at integrating AI into healthcare advancements. The full study, published in *npj Women’s Health*, offers deep insights and is accompanied by resources to assist clinicians in the ethical implementation of these tools.
Subject of Research: Artificial Intelligence Applications in Healthcare, Intimate Partner Violence Risk Prediction
Article Title: Leveraging Multimodal Machine Learning for Accurate Risk Identification of Intimate Partner Violence
News Publication Date: March 13, 2026
Web References: https://bhartikhurana.bwh.harvard.edu/airs/; https://www.cdc.gov/intimate-partner-violence/about/index.html
References: Gu J, Villalobos Carballo K, Ma Y, Bertsimas D, and Khurana B. Leveraging multimodal machine learning for accurate risk identification of intimate partner violence. npj Women’s Health. 2026. DOI: 10.1038/s44294-025-00126-3
Keywords: Artificial intelligence, Machine learning, Intimate partner violence, Risk assessment, Clinical decision support, Biomedical imaging, Data fusion, Predictive analytics
Tags: AI in clinical informatics, biomedical imaging in violence detection, early identification of domestic violence, healthcare data analysis for abuse, intimate partner violence detection AI, machine learning for IPV risk, mental health impact of IPV, multidisciplinary research on IPV, NIH-funded AI health projects, overcoming patient disclosure barriers, predictive models for patient safety, public health technology innovation
