TOPLINE:

In emergency department (ED) patients with suicidal ideation, augmentation of an electronic health record (EHR)-based suicide risk score with psychosocial factors identified from clinical notes was linked to improved prediction of 90 ‑ day suicide attempts. Chronic stress was the strongest predictor.

METHODOLOGY:

• Researchers conducted a retrospective prognostic study using EHR data from 4661 patients discharged from the ED after presenting with suicidal ideation across two centers — Vanderbilt University Hospital (VUH; n = 3382; mean age, 26.1 years; 51.8% men; 67.8% White individuals) and Regional Health Systems (RHS; n = 1279; mean age, 34.5 years; 55.9% men; 80.6% White individuals) between 2018 and 2024.
• Researchers evaluated the Vanderbilt Suicide Attempt and Ideation Likelihood (VSAIL) score alongside six psychosocial factors — homelessness, financial insecurity, chronic stress, social isolation, loneliness, and adverse childhood experiences — extracted from clinical notes using a validated natural language processing (NLP) approach.
• The primary outcome was a suicide attempt within 90 days after the ED visit.
• Model performance was compared across three approaches — VSAIL alone, psychosocial factors alone, and a combined model — using the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), and classification metrics such as the positive predictive value (PPV), the negative predictive value, sensitivity, and specificity.

TAKEAWAY:

• Within 90 days, suicide attempts were reported in 4.7% of visits at VUH vs 2.7% at RHS.
• The addition of psychosocial factors to VSAIL significantly improved model performance compared with VSAIL alone, increasing the AUROC at VUH (from 0.645 to 0.734; P < .001) and RHS (from 0.547 to 0.680; P < .001). The median AUPRC also increased at VUH (from 0.083 to 0.122; P < .001) and RHS (from 0.029 to 0.054; P < .001).
• The combined model also improved prediction in the group with the highest risk for suicide attempt: the median PPV increased from 0.093 to 0.143 (P < .001) at VUH and from 0.042 to 0.112 (P < .001) at RHS, while maintaining specificity > 0.90 at both sites.
• Chronic stress was the strongest predictor of suicide attempt (P < .001), followed by adverse childhood experiences (P < .001) and a higher VSAIL score (P = .04).

IN PRACTICE:

"Incorporating psychosocial factors extracted from clinical notes was associated with higher predictive performance, bringing it toward published benchmarks for clinically actionable levels. Chronic stress emerged as the strongest predictor in the augmented suicide risk prediction model," the authors wrote.

"Identifying more suicide-relevant psychosocial factors will be crucial for further improving these models and achieving more accurate and clinically actionable suicide risk predictions," they added.

SOURCE:

The study was led by Hyunjoon Lee, MS, Vanderbilt University Medical Center, Nashville, Tennessee. It was published online on March 04, 2026, in JAMA Network Open.

LIMITATIONS:

The study was limited by a predominantly White population, restricting the generalizability of the findings. The NLP approach did not distinguish past from current suicidal events. Only six psychosocial factors were included; future work should incorporate additional factors such as reasons for living. When psychosocial factors were not documented, a value of zero was assigned, possibly reflecting unrecorded evidence rather than true absence and potentially biasing effect estimates. Additionally, the EHR data captured only encounters within Vanderbilt University Medical Center.

DISCLOSURES:

The study was funded by the FDA and the National Institute of Mental Health. One author reported receiving grants from the National Institutes of Health during the conduct of the study, personal fees from YuiMedi outside the submitted work, and a pending patent. Detailed disclosures are provided in the original article.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.