AI Predicts Intimate Partner Violence Risk Using Medical Records | Mass General Brigham

The potential for earlier intervention in cases of intimate partner violence (IPV) has received a boost with the development of modern artificial intelligence tools. Researchers at Mass General Brigham have created a series of machine learning models capable of identifying individuals at risk of IPV by analyzing information within their electronic medical records (EMRs). This work, published in npj Women’s Health, suggests the possibility of detecting potential abuse up to four years before a person seeks help at a dedicated domestic violence treatment center.

Understanding the Scope of Intimate Partner Violence

Intimate partner violence, encompassing physical, sexual, and emotional abuse between current or former partners, remains a significant public health concern. The prevalence is substantial; more than one-third of women and one in ten men will experience IPV in their lifetimes. Despite these figures, disclosure to healthcare providers is often low, hindered by factors like fear, stigma, and financial or emotional dependence on the abuser. This underreporting underscores the need for proactive identification strategies.

The World Health Organization classifies IPV as the most widespread form of violence against women globally. The impact extends far beyond immediate physical harm, with documented links to chronic pain, sexually transmitted infections, and a range of mental health disorders including depression, post-traumatic stress disorder (PTSD), and anxiety. Research highlights the complex interplay between IPV and overall health, emphasizing the importance of early detection and intervention.

How the AI Tools Work: A Deeper Look

The Mass General Brigham team’s approach centers on leveraging the wealth of data contained within EMRs. These records, which include clinical notes and structured data, are analyzed using machine learning algorithms to identify patterns and indicators suggestive of IPV. The study utilized both tabular clinical data and unstructured clinical notes, building both single-modality and multimodal models to account for varying data availability. The multimodal model, combining different data types, demonstrated a particularly strong ability to identify at-risk individuals, achieving an area under the curve (AUC) of 0.88 in identifying patients before they sought help at an intervention center.

Importantly, the researchers validated the model’s performance not only on patients who ultimately accessed domestic violence services but likewise on individuals who did not, and on patients from a separate hospital within the same integrated network. This validation step is crucial for assessing the generalizability and reliability of the AI tools.

Beyond Prediction: Supporting Clinical Judgment

Bharti Khurana, MD, MBA, principal investigator and founding director of the Trauma Imaging Research and Innovation Center at Mass General Brigham, emphasized that the AI tools are designed to support clinicians, not replace them. “Our research offers proof of concept that AI can support clinicians in flagging possible abuse earlier,” she stated in a press release. The goal is to enable earlier conversations about IPV between healthcare providers and patients, potentially leading to more timely interventions and improved outcomes.

The Challenge of Underreporting and the Role of Proactive Screening

The difficulty in obtaining self-reported information about IPV is a major obstacle to effective prevention and intervention. Patients may be reluctant to disclose abuse due to a complex mix of factors, including fear of retaliation, shame, and concerns about the impact on their families. This is where proactive screening, facilitated by tools like these AI models, could play a vital role. By identifying individuals at risk based on patterns in their medical records, healthcare providers can initiate sensitive and supportive conversations, creating a safer space for disclosure.

Evidence and Limitations: Interpreting the Findings

Even as the results are promising, it’s important to acknowledge the limitations inherent in this type of research. The study focused on female patients seeking help at a domestic abuse intervention and prevention center, which may not fully represent the broader population experiencing IPV. The reliance on EMR data raises questions about potential biases in data collection and documentation. For example, certain demographic groups may be less likely to access healthcare or may receive different levels of care, which could affect the accuracy of the AI models.

The AUC of 0.88, while strong, does not guarantee perfect accuracy. It indicates the model’s ability to distinguish between individuals at risk and those who are not, but false positives and false negatives are still possible. Further research is needed to refine the models and assess their performance across diverse populations and healthcare settings.

What Comes Next: Refining and Implementing the Tools

The development of these AI tools represents a significant step forward, but several key steps remain before widespread implementation. Researchers are continuing to refine the models, exploring ways to improve their accuracy and reduce bias. Future studies will focus on evaluating the tools in real-world clinical settings, assessing their impact on patient care and outcomes.

A crucial aspect of this process will be addressing ethical considerations related to data privacy and patient confidentiality. Robust safeguards must be in place to ensure that patient information is protected and used responsibly. Collaboration between researchers, clinicians, and policymakers will be essential to navigate these challenges and ensure that these tools are deployed in a way that benefits patients and promotes public health. Ongoing monitoring and evaluation will be critical to assess the long-term effectiveness and impact of these AI-driven interventions.