NIH-funded AI flags intimate partner violence risk years earlier
Researchers funded by the National Institutes of Health developed an AI-driven clinical decision support tool that predicts intimate partner violence (IPV) risk using routine medical data. In testing, a multimodal model combining structured EHR data and unstructured clinical notes reached 88% accuracy and signaled risk on average more than three years before patients typically enroll in hospital-based intervention programs.
For IT and engineering teams, this points to a practical, high-impact use case: deployable ML that augments clinical judgment, improves early outreach, and operates within existing EMR workflows. The models assist conversations and referrals; they are not intended to make definitive diagnoses.
What the study did
- Data: Several years of hospital records from nearly 850 affected female patients and 5,200 matched controls.
- Modalities: Structured/tabular data (e.g., codes, encounters, labs) and unstructured text (e.g., clinician notes, radiology reports).
- Models: One trained on tabular data, one on text, and a late-fusion model that merges both at prediction time.
- Highlights: The fusion model achieved the strongest and most stable performance (88% accuracy). Tabular slightly led on earliest detection, while fusion identified more cases in advance.
Radiology documentation contributed valuable signal, as repeated injury patterns can surface in imaging narratives. The team emphasized that AI should support, not replace, clinician judgment-especially for sensitive conversations around safety and resources.
Why this matters for engineering
Most IPV goes unreported due to safety concerns, fear, and stigma. Models that learn from routine EHR traces can help surface risk quietly and consistently, creating a window for earlier, patient-centered interventions.
Late fusion also makes the system more portable. Because hospitals vary in the depth of their notes and availability of specific fields, decoupling modality-specific pipelines and merging only at prediction time improves resilience across sites.
Practical implementation guide
- Data sources
- Structured: demographics, encounters, diagnosis/procedure codes, triage data, vitals, labs, meds, prior utilization.
- Unstructured: clinician notes, ED summaries, discharge notes, radiology reports.
- Pipelines: FHIR APIs for pull, HL7 v2 event streams for near real-time ingestion.
- Feature and text pipelines
- Tabular: standardize code systems, temporal features (event frequency, recency), simple aggregates, and trend flags.
- Text: de-identify for model training when feasible; tokenize and embed with domain-tuned NLP; handle section headers and negations in notes.
- Model architecture
- Two independent encoders (tabular and text), calibrated separately.
- Late fusion at prediction time (e.g., weighted average or learned combiner), plus post-hoc calibration.
- Provide a continuous risk score, confidence interval, and top contributing factors for clinician review.
- Workflow integration
- Non-interruptive alerts in the EMR with tiered thresholds (informational vs. actionable).
- Human-in-the-loop review by social work, nursing, or trained clinicians before outreach.
- Context-aware suppression (e.g., patient currently accompanied by a partner; avoid on-screen disclosures).
- Governance and safety
- Bias and fairness checks across age, race/ethnicity, language, and payer type.
- Clear documentation: data provenance, model limitations, failure modes, and guidance for sensitive conversations.
- Drift monitoring, periodic recalibration, and versioned rollouts with A/B evaluation on precision/recall and downstream outcomes.
- Privacy and security
- Minimum necessary PHI, role-based access, audit logs, encryption in transit/at rest, and on-prem or VPC isolation.
- IRB oversight where required; align with HIPAA, institutional policies, and patient safety protocols.
Key results to anchor your roadmap
- Fusion model accuracy: 88% in the study cohort.
- Early signal: tabular and fusion models flagged risk on average 3+ years before intervention enrollment.
- Stability: fusion outperformed single-modality models and generalized better across variable documentation practices.
Clinical context and guardrails
"This clinical decision support tool could make a significant impact on prediction and prevention of intimate partner violence," said Dr. Qi Duan, director at NIBIB. According to the team, AI support can help clinicians initiate earlier, safer, and more informed conversations with patients about IPV and connect them to resources.
"The goal is never to force disclosure, but to help clinicians communicate with patients in a supportive way and to connect them with resources and support," said Bhati Khurana, M.D. Use the model as a prompt for thoughtful, patient-centered care-not as a label.
What's next
The team plans to embed the decision-support tool inside EMR systems for real-time IPV risk evaluation. If you're evaluating similar deployments, start with a shadow-mode pilot, verify calibration, and partner closely with clinical leadership and social work to define safe escalation paths.
Resources
- CDC: Intimate Partner Violence
- Automated IPV Risk Support (AIRS) project site
- AI for Healthcare
- AI for Medical Records Clerks
Study details
IPV affects millions of people across genders in the United States and often goes undetected. This study compared tabular, text, and late-fusion ML models trained on multi-year hospital data to predict IPV risk earlier and more accurately than routine screening alone.
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.
Your membership also unlocks:
