Researchers Identify Hidden Self-Harm Histories Using Machine Learning

Why It Matters

Under‑recorded self‑harm histories skew resource allocation and research, limiting effective suicide‑prevention strategies. Better detection can guide more accurate planning and targeted clinical interventions.

Key Takeaways

• Diagnosis codes capture only 25% of documented self‑harm cases
• Machine‑learning model estimates 7.9% self‑harm prevalence in VHA patients
• Problem lists miss self‑harm history in over 77% of coded cases
• PULSNAR method handles unlabeled data without assuming random coding
• Improved visibility can refine mental‑health resource planning and research

Pulse Analysis

Self‑harm is a leading predictor of future suicide, yet health systems often miss it because traditional data pulls rely on diagnosis codes and problem lists. In the Veterans Health Administration, those structured fields captured just 1.85% of self‑harm events, while chart reviews revealed a prevalence near 8%. This discrepancy means that policymakers and researchers may dramatically underestimate the demand for mental‑health services, leading to gaps in funding, staffing, and preventive programs.

The UNM team addressed the visibility problem with PULSNAR, a positive‑and‑unlabeled learning algorithm designed for messy real‑world records. By training on patients with known self‑harm codes and then estimating the likelihood of hidden cases among uncoded patients, the model sidesteps the assumption that missing codes equal absence of condition. This approach uncovers patterns in structured data—such as injury codes, poisoning events, or psychiatric comorbidities—that signal undocumented self‑harm, offering a scalable way to flag records for deeper review.

For health systems, especially large integrated networks like the VHA, the method promises more accurate epidemiologic estimates and better allocation of suicide‑prevention resources. While still a research tool, the technology could eventually complement existing reporting mechanisms, inform risk‑adjusted benchmarking, and support targeted outreach to high‑risk veterans. Extending the framework to other under‑coded conditions—opioid use disorder, PTSD, depression—could further tighten the feedback loop between data analytics and clinical care, ultimately improving outcomes across the mental‑health continuum.