Researchers are utilizing machine learning to detect latent signs of abuse in medical records, potentially enabling earlier clinical intervention for vulnerable patients.

A research team at MIT and Mass General Brigham has developed a predictive model capable of identifying patients at risk of intimate partner violence (IPV) years before they formally disclose abuse to healthcare providers. The system, known as Automated IPV Risk Support (AIRS), leverages existing electronic health records to provide clinicians with early warning indicators that might otherwise remain obscured by standard medical documentation.

The AIRS system processes two distinct data streams to generate its risk assessments. It analyzes structured data such as diagnostic codes, medication history, and emergency department visit frequency alongside unstructured clinical notes written by physicians and social workers.

To interpret the free-text clinical notes, the researchers utilized Clinical-Longformer, a specialized language model designed to convert medical narratives into numerical data. This dual-stream approach ensures the model remains functional even when specific types of patient data are incomplete or missing across different clinical institutions.

The model architecture employs a framework called Holistic AI in Medicine (HAIM) to fuse these data streams at the prediction stage. By keeping the streams independent, the system maintains reliability even when institutional record-keeping practices vary significantly between facilities.

Testing across multiple cohorts yielded an area under the receiver operating characteristic curve (AUC) of 0.88, demonstrating high efficacy in distinguishing between IPV cases and non-cases. The study revealed that the model successfully flagged over 80% of IPV instances with an average lead time of nearly four years prior to patient self-reporting.

These results suggest that longitudinal medical records contain latent signals of domestic abuse that current screening protocols often fail to capture. The integration of socioeconomic indicators, such as zip-code-level social deprivation scores, further enhances the model’s ability to identify patients who may face higher systemic risks.

Dr. Bharti Khurana, founding director of the Trauma Imaging Research and Innovation Center at Harvard Medical School, emphasizes that the tool functions as a silent clinical support mechanism rather than a diagnostic instrument. The system is designed to surface risk scores to clinicians, providing a foundation for trauma-informed conversations rather than triggering automatic interventions.

The deployment of AIRS is currently in pilot testing at Mass General Brigham, where it is paired with structured guidance to assist medical staff in navigating sensitive patient interactions. This proactive approach aims to reduce the significant time delay between the onset of abuse and the point at which a patient feels secure enough to disclose the situation to a provider.

Despite the technical success of the model, experts caution that algorithmic tools in this domain face significant ethical and practical hurdles. Alexia Maddox, a senior lecturer at La Trobe University, points out that the current clinical definition of IPV may be too narrow to account for the complexities of coercive control. Historical precedents, such as the VioGén system in Spain, demonstrate that over-reliance on algorithmic outputs without independent human judgment can lead to dangerous outcomes for victims.

The shift toward predictive analytics in healthcare settings necessitates a rigorous evaluation of how these flags influence clinical behavior. If clinicians view an AI-generated score as a definitive diagnosis, the risk of misinterpretation or inappropriate intervention increases, potentially undermining the safety of the very individuals the technology intends to protect.

Future development will likely focus on refining the model to incorporate broader definitions of abuse, including non-physical markers of control. Stakeholders must now determine how to balance the potential for life-saving early intervention with the need for stringent data privacy and the prevention of algorithmic bias in high-stakes clinical environments. The success of these systems depends not just on the accuracy of the underlying code, but on the institutional capacity to handle the sensitive information that these models uncover.

At Global Business & Economics Journal, we contribute to the  future of business and technology with news,  innovative research, expert insights, and thought-provoking analysis.