Every time a case of Intimate Partner Violence (IPV) makes headlines in India, the public conversation turns, understandably, to safety, justice and punishment. Those conversations matter. But there is another question that gets far less attention: could we have recognised this risk earlier, quietly and safely, during medical care, before the violence reached a crisis point or a crime was reported?
Many women experiencing abuse seek medical care long before they seek help from the police or support services. They may arrive with an injury attributed to a fall, return repeatedly with pain or anxiety, or undergo imaging for injuries that appear unrelated when viewed one visit at a time. The warning signs may already exist across years of medical records, but remain scattered among different visits, clinicians and hospitals. Could technology help bring those fragments together and give a health-care clinician an opportunity to ask the right question sooner?
This pattern is not unique to India. IPV defined as physical violence, sexual violence, stalking, or psychological aggression by a current or former partner, is a global public health blind spot. In the US, where I practise and where our research was conducted, millions of people experience intimate partner violence every year, and more than one in three American women will experience it at some point in their lives, according to the Centers for Disease Control and Prevention. The numbers differ across countries, but the underlying pattern does not: IPV leads to life-threatening injuries, chronic pain and serious mental health conditions, yet most people who experience it never tell a health provider. Fear, safety concerns and stigma keep them silent, in Boston as much as in Bengaluru. The screening tools hospitals currently use rely almost entirely on patients disclosing abuse themselves, and so they catch only a fraction of real cases, whether in the US or in India. Studies suggest that many patients do not disclose IPV because they may not recognise their experiences as abuse or know that health care professionals can connect them with confidential support and resources. By the time IPV is formally identified, irreversible harm has often already occurred.
This is the problem my research team at Harvard and MIT set out to address. With funding from the US National Institutes of Health, we built and tested an Artificial Intelligence (AI) tool designed not to diagnose abuse, but to flag which patients may be at risk of it, years before they might ever seek help themselves. Our findings were published in the March issue of the npj Women's Health.
{{/usCountry}}This is the problem my research team at Harvard and MIT set out to address. With funding from the US National Institutes of Health, we built and tested an Artificial Intelligence (AI) tool designed not to diagnose abuse, but to flag which patients may be at risk of it, years before they might ever seek help themselves. Our findings were published in the March issue of the npj Women's Health.
{{/usCountry}}We used machine learning to build three models, trained on the electronic medical records of 841 patients enrolled in a hospital-based domestic abuse intervention and prevention centre reporting physical, sexual or psychological harm from an intimate partner, compared against more than 5,200 similar patients with no such history. One model analysed structured data, such as the tables and codes in a medical record. A second analysed unstructured data, including the free-text notes a doctor writes. The third combined both types of information.
All three models performed well in distinguishing patients with documented IPV from matched patients without documented IPV. Performance was measured using the area under the curve, or AUC, where 0.5 represents chance and 1.0 represents perfect discrimination. The structured-data model achieved an AUC of 0.85, the clinical-notes model 0.87, and the combined model performed best at 0.88. What struck us most, however, was not only the model’s performance but its timing. The structured-data and combined models could identify elevated IPV risk more than three years, on average, before a patient sought help.
Certain patterns recurred in what the AI picked up on: A history of mental health conditions, unexplained chest pain, regular painkiller use, high social deprivation, and frequent radiology imaging, often for injuries that, individually, seem minor. Patients who kept up with routine preventive care, such as mammograms and cervical screenings, showed lower IPV risk, likely because they already had better access to, and comfort with, the health care system.
It is important to be precise about what this tool is not. It cannot and should not diagnose abuse, because no algorithm can replace a conversation grounded in trust. What it can do is help a clinician notice a pattern across years of scattered visits that no one reviewing a single file in a rushed appointment could piece together. It is a prompt for a more informed, sensitive and caring conversation, not a verdict.
That distinction matters enormously in the Indian context, where the underreporting of IPV is compounded by financial dependency, family pressure and a justice system many women hesitate to approach. A tool that helps a clinician open the door to a private and supportive conversation could give a woman earlier access to information, counselling, legal assistance and other resources. The purpose is not to make decisions for her, but to give her greater knowledge, support and agency to decide what is safest for herself. Health care systems must create private, safe and supportive spaces where patients feel able to disclose IPV without fear of judgement, pressure or further harm.
We are now testing the model in broader and more diverse populations within the US and hope to evaluate it internationally in the future. The underlying idea, however, may apply across health care systems. Signs of elevated risk may already be present in a patient’s medical history, including repeated visits, recurrent injuries, persistent pain and patterns of health care use that are difficult to recognise when each encounter is viewed separately. Technology can help clinicians see these patterns earlier, but it cannot end IPV on its own. Its value lies in creating an earlier opportunity for a private conversation, access to resources and support, while preserving each woman’s autonomy to decide what is safest for her.
(The views expressed are personal)
This article is authored by Dr Bharti Khurana, associate professor of Radiology, Harvard Medical School and emergency radiologist, Brigham and Women’s Hospital, Boston.