In a groundbreaking initiative, a team of researchers funded by the National Institutes of Health (NIH) has developed an innovative artificial intelligence (AI) tool designed to assist clinicians in identifying patients at risk of intimate partner violence (IPV). By utilizing data collected during routine medical visits, the team trained a machine-learning model capable of accurately detecting IPV, marking a significant advancement in healthcare interventions.
Intimate partner violence encompasses abuse inflicted by current or former partners, leading to severe repercussions such as life-threatening injuries, chronic pain, and mental health issues. It impacts millions across the United States, affecting both men and women at various stages of life. Unfortunately, numerous cases go unnoticed due to patients’ reluctance to disclose their situations, driven by concerns for their safety, fear, and social stigma.
In their investigation, the research team, spearheaded by scientists from Harvard Medical School in Boston, introduced three AI models aimed at detecting IPV within healthcare contexts, evaluating their effectiveness in making predictions.
“This clinical decision support tool could significantly enhance the prediction and prevention of intimate partner violence,” stated Dr. Qi Duan, Ph.D., director of the Division of Health Informatics Technologies at NIH’s National Institute of Biomedical Imaging and Bioengineering (NIBIB). “Given how prevalent these cases are, this tool has the potential to be a transformative asset for public health.”
According to the study authors, many instances of IPV remain unrecognized, which results in missed opportunities for timely interventions. They noted that existing screening tools often capture only a fraction of cases, while clinical and imaging records hold valuable information for identifying IPV risk. Radiologists, in particular, are in a unique position to recognize indicators of IPV through the patterns of physical trauma they observe.
The researchers analyzed several years’ worth of hospital data, drawing from nearly 850 affected female patients alongside 5,200 control patients who were matched by age and demographics. Given the inconsistencies in data collection across various healthcare settings, they developed two distinct AI models: one trained on structured patient data (arranged in table format) and another trained on unstructured data, such as medical notes and radiology reports. Additionally, they created a multimodal model that integrates both structured and unstructured data.
All models demonstrated high performance levels during the study; however, the multimodal fusion model surpassed those relying solely on one type of data, achieving an 88% accuracy rate. Both the tabular and fusion models were capable of identifying IPV risk over three years prior to patients’ enrollment in hospital-based domestic abuse intervention programs. While the tabular model identified IPV risk slightly earlier, the fusion model successfully detected more cases in advance.
The multimodal model exhibited more consistent performance compared to relying on just one type of data. The researchers explained that the different data types are processed separately and merged only during the prediction phase, with the tabular framework being particularly relevant in a healthcare context where data availability varies among hospitals.
The team stressed that AI tools like their machine learning models could help healthcare providers initiate timely conversations with patients regarding IPV, guiding them toward suitable support resources. These AI tools are intended to assist rather than replace clinical judgment in making definitive diagnoses.
“For decades, our healthcare system has largely depended on patient self-disclosure to identify intimate partner violence, resulting in many cases being overlooked and unsupported,” noted Bhati Khurana, M.D., the study’s senior author and an emergency radiologist at Mass General Brigham, as well as an associate professor of radiology at Harvard Medical School. “Our work signifies a pivotal shift from reactive disclosure to proactive risk recognition within routine clinical practice. By analyzing existing patterns in healthcare data, this approach aids clinicians in initiating safer, more informed, and earlier discussions with patients.”
The researchers believe that when employed with a patient-centered approach, this tool can play a crucial role in proactive IPV intervention, offering timely and effective assistance, which can ultimately enhance long-term health outcomes for those at risk. They have established guidance on the project website to assist clinicians in thoughtfully navigating discussions with patients.
“The aim is never to compel disclosure but to facilitate supportive communication and connect patients with necessary resources,” said Khurana.
The research team intends to develop AI models that will integrate into electronic medical record systems, providing real-time evaluations of IPV risk in clinical environments.
For more about IPV: About Intimate Partner Violence | Intimate Partner Violence Prevention | CDC
For further information regarding Automated IPV Risk Support: https://bhartikhurana.bwh.harvard.edu/airs
This research received co-funding from the NIBIB grant R01EB032384 and the NIH Office of the Director.
