NIH-funded AI flags intimate partner violence risk years before patients seek help

NIH-funded AI flags intimate partner violence risk years earlier

Friday, March 13, 2026

Many patients living with intimate partner violence (IPV) don't disclose it in clinical settings. That silence leaves injuries untreated and risk unaddressed. A new NIH-funded AI decision-support tool changes the timing by predicting IPV risk using data already inside the EHR.

Built by a team led from Harvard Medical School and Mass General Brigham, the models use structured records and unstructured notes to surface risk well before patients seek formal help. The goal is simple: enable earlier, safer conversations and connect people to support.

What the researchers built

• Three models: a tabular model (structured EHR data), a text model (unstructured notes and radiology reports), and a multimodal fusion model (late-stage merge of both).
• Training data: several years of hospital records from ~850 affected female patients and ~5,200 matched controls.
• Clinical use: decision support only. These models are not diagnostic and should trigger supportive, patient-centered workflows.

Performance at a glance

• The fusion model was most accurate, at 88%.
• The tabular and fusion models identified IPV risk on average more than three years before patients enrolled at hospital-based domestic abuse intervention centers.
• The tabular model surfaced risk a bit earlier; the fusion model flagged more cases ahead of time and showed more stable performance across settings.

Why it matters: current screening misses a large share of cases. Radiology and clinical histories often contain patterns-injury frequencies, visit sequences, and note language-that point to risk even when disclosure hasn't happened.

How the fusion approach works

• Separate pipelines: structured features (e.g., ICD/CPT codes, meds, encounters) and unstructured notes (including radiology reports) are modeled independently.
• Late fusion: predictions are merged at the end, improving portability across hospitals with different documentation habits.
• Radiology advantage: recurring trauma patterns can carry strong signal that complements tabular histories.

Clinical intent and safeguards

These tools are built to support clinicians-not to declare a diagnosis. Use outputs to open trauma-informed, supportive conversations and offer resources without pressuring disclosure. The team provides guidance to help clinicians approach these talks thoughtfully.

As one NIH leader noted, given how widespread IPV is, scalable, data-driven support can meaningfully improve prevention and care. The senior study author emphasized shifting from reactive disclosure to proactive risk recognition within routine care.

Implementation notes for healthcare, IT, and development teams

• Integration pattern: consider HL7 FHIR and CDS Hooks for EHR-embedded risk prompts. Scope which FHIR resources to pull (e.g., Encounter, Condition, Observation, Procedure, MedicationStatement/Request, ImagingStudy, DocumentReference).
• Data pipeline: enable near-real-time ingestion of structured fields and clinical notes. Track provenance. Log feature versions and text-processing changes for audit.
• Model ops: version models, preserve training data schemas, and document thresholds per site. Monitor drift with rolling AUC/PPV/alert rate.
• Alert UX: minimize alert fatigue with tiered risk levels, batched notifications, and clear next-step actions (e.g., private room assessment, social work referral).
• Workflow design: map a closed-loop pathway-screen privately, assess safety, offer resources, warm handoffs, and document follow-up.
• Equity checks: measure performance across demographics and visit types. Set governance for periodic bias reviews and threshold calibration.
• Privacy and safety: apply least-privilege access, detailed audit logs, and careful note-handling since abusers may access portals in some cases. Align with HIPAA and local policy.
• Staff training: prepare clinicians on trauma-informed communication and "no-pressure" conversations. Train health information staff on notes quality and coding consistency.
• Evaluation: track clinical KPIs-time to support referral, patient-accepted services, follow-up completion, repeat-injury rates, and false-positive impact on workflows.

Practical steps to get started

• Run a retrospective validation on your EHR data to estimate site-specific performance and set thresholds.
• Pilot in a limited setting (e.g., ED or radiology-linked workflows) before broader rollout.
• Stand up a multidisciplinary oversight group: clinical leads, social work, radiology, IT, data science, compliance, and patient safety.
• Publish internal guidelines on when, how, and by whom patients are approached-and how safety checks are performed.

Who benefits in your org

• Clinicians: earlier signal to open supportive dialogue and coordinate resources.
• Radiology: structured pathways when injury patterns suggest elevated risk.
• Care management/social work: timely referrals and follow-up tracking.
• IT/Dev: a clear FHIR/CDS Hooks use case with measurable impact and defined governance.

Key limitations to consider

• Study cohort focused on female patients; performance in broader populations needs validation.
• Documentation styles vary across sites; expect threshold tuning and periodic recalibration.
• False positives/negatives carry clinical and operational costs-monitor and adjust.

Resources

• About Intimate Partner Violence | CDC
• Automated IPV Risk Support (AIRS) - clinician guidance
• AI for Healthcare
• AI Learning Path for Medical Records Clerks

Reference

Gu J, Villalobos Carballo K, Ma Y, Bertsimas D, and Khurana B. Leveraging multimodal machine learning for accurate risk identification of intimate partner violence. Nature Portfolio Journal: Women's Health. 2026. DOI: 10.1038/s44294-025-00126-3.

Bottom line: Put this model where care happens, keep the focus on patient safety and consent, and measure what matters. Earlier risk recognition gives your teams time to act-and that time can change outcomes.