EMR-Derived ML Flags for Intimate Partner Violence: Radiology Operational Considerations

A machine-learning approach using electronic medical record (EMR) data was reported to flag future intimate partner violence (IPV) risk years before some patients presented to a domestic abuse intervention and prevention center.

The account describes models built from both structured EMR fields and narrative documentation, with unstructured inputs spanning routine clinical notes as well as radiology and emergency department (ED) reports. The report describes retrospective testing in time-stamped, archived EMR records, including a mean lead time of 3.68 years before some patients sought care. Overall, the work is framed as an EMR-derived risk flag evaluated in retrospectively assembled records.

Investigators trained three related approaches—tabular models from structured data, models derived from unstructured notes, and a combined “fusion” approach—using structured and notes-based IPV risk models. The structured inputs named in the source include diagnoses, medications, radiology history, hospital visits, vital signs, and zip code–level social deprivation and household income proxies, while the unstructured component draws on clinical narratives, including radiology and ED reports. Model development was described using EMR data from 841 women who visited a domestic abuse intervention and prevention center between 2017 and 2022, excluding 2020, along with 5,212 demographically matched controls who did not carry abuse-related diagnoses. In the report, the approaches are presented as modality-specific models plus a fused model intended to integrate multiple EMR modalities.

In testing, the best-performing approach was reported to be the HAIM fusion model, with an AUC of 0.88. In a separate analysis using time-stamped, archived medical records from the 2023 self-report cohort, the fusion model reportedly detected 80.6% of cases in advance, with a mean lead time of 3.68 years before patients sought care at the intervention center. These results are presented, within this dataset, as model predictions from archived EMR records that preceded the eventual center visit by years. The performance and timing findings are described as coming from retrospective evaluation of archived EMR data.

Beyond the primary training and testing process, validation was described in two further patient groups not included in the initial model development: patients at another hospital in the same healthcare network who enrolled in a domestic abuse intervention and prevention center, and patients at the original hospital who carried IPV-related diagnoses but did not enroll in the center. The source reports cohort-specific validation metrics, with the fusion model achieving AUCs of 0.82 and 0.84 in those additional groups. Prospective, real-time deployment was not evaluated in this study, although the authors state that they plan to develop an EMR-embedded decision support tool for real-time IPV risk evaluation.

Because the notes-based and fused approaches draw on narrative documentation that includes radiology reports, the source suggests a pathway by which an EMR-derived IPV risk flag could intersect with documentation-heavy environments that span ED and imaging encounters, depending on where such a score is surfaced.

The authors also note limitations tied to how cases and controls were defined, including that development and validation drew heavily from patients who sought help for IPV or carried IPV-related diagnoses, which may limit generalizability to people less likely to disclose or seek care. They also note that IPV is underreported and that the true distribution of IPV and non-IPV patients is unknown, which could affect model evaluation. The report describes future work including evaluation in broader clinical populations, further analysis of contributing factors, and development of a real-time EMR decision support tool.

Key Takeaways:

• A recent report describes multimodal EMR inputs spanning structured fields and unstructured clinical narratives, including radiology and ED documentation.
• A fused, multimodal model is reported as the top performer, with retrospective analyses of time-stamped records suggesting that risk signals could be identified years before some patients presented to an intervention center.
• The authors cite limitations related to the development population and the underreported nature of IPV, and they describe broader evaluation and real-time clinical decision-support development as next steps.