How to Choose Suicidal Ideation Screening in Veterans
How should clinicians choose between a simple PHQ-2 approach and more complex prediction models when screening veterans for suicidal ideation?
Veterans have elevated risk for suicidal ideation, but screening programs must balance missed cases against overwhelming follow-up demand. This guide applies to clinical settings deciding whether the added complexity of multipredictor models is worth using instead of a simple PHQ-2-based approach.
-
Define the screening goal before selecting a model
Decide whether the intended use is broad screening, where minimizing missed cases is the priority, or a different clinical function with a different tolerance for false positives. The article emphasizes that sensitivity, specificity, and other performance characteristics should be matched to the clinical context rather than judged by AUC alone.
-
Consider the PHQ-2 sum score as the practical reference approach
Use the PHQ-2 sum score as the starting comparator because it is simple, clinically intuitive, widely implemented, and does not directly ask about suicidality. In this study, the PHQ-2 achieved AUCs of 0.79 to 0.80, and the authors used different PHQ-2 severity cutoffs to represent a standard screening approach.
-
Interpret any advantage of complex models as modest
Recognize that more complex models improved discrimination only modestly in this veteran sample. Comprehensive 32-variable models reached AUCs of 0.85 to 0.88 versus 0.79 for the PHQ-2, while clinically available 9-predictor models reached AUCs of 0.78 to 0.83 versus 0.80 for the PHQ-2.
-
Plan around false positives before implementing multipredictor models
Expect that positive screens will often not correspond to true suicidal ideation, especially if the model is used at default decision thresholds. Positive predictive values were low at 0.23 to 0.27 in the comprehensive models and 0.16 to 0.30 in the clinically available models, so implementation is only practical if the system can follow up all flagged patients.
-
Use negative results mainly for ruling out rather than confirming risk
Treat these approaches as better for excluding likely past-year suicidal ideation than for confirming it. Negative predictive values were consistently high, ranging from 0.97 to 0.98 in the comprehensive models and 0.95 to 0.98 in the clinically available models.
-
Weigh data burden and workflow feasibility
Factor in the infrastructure required to collect, preprocess, and maintain larger predictor sets. The comprehensive approach used 32 variables and had more missing-data loss, with 2,666 participants available at wave 1 and 2,659 at wave 2, compared with 3,026 and 3,052 in the 9-predictor clinically available set.
-
Prefer simpler tools when local capacity is limited
If the practice does not have robust data systems, technical support, or clinician training for interpreting algorithmic outputs, favor the simpler PHQ-2-based strategy. The authors conclude that complex machine-learning approaches did not offer clinically meaningful improvement over the PHQ-2 for predicting suicidal ideation in veterans.
-
Embed any prediction approach within a broader prevention pathway
Do not treat a positive model output as a standalone solution. The article states that there is still no consensus on what actions should follow when someone is flagged high risk and that prediction models are only useful when integrated into comprehensive suicide prevention strategies.
Clinical Considerations
- The study did not develop or validate a clinically deployable prediction tool or specify a post-screen intervention pathway.
- Sensitivity and specificity were derived from default model decision thresholds, typically probability 0.50, which may not match the threshold a screening program would choose.
- Performance may have been inflated because the same individuals were assessed at both time points rather than in a fully prospective external validation cohort.
- The findings come from a veteran-specific sample and may not generalize to other populations.
Bottom Line
For veteran suicidal ideation screening, a PHQ-2-based approach is usually the most practical choice because more complex models added only modest discrimination while increasing false-positive burden and implementation complexity.
