» Back to Table of Contents

Innovation Update

Enhancing Suicide Risk Prediction with Theory-Guided Natural Language Processinge

Suicide continues to pose a major public health challenge, particularly among Veterans, for whom rates of suicide remain significantly higher compared to rates for the general population. Accurate risk prediction is vital for delivering timely interventions, yet current approaches often rely exclusively on structured electronic health record (EHR) data, which may not contain critical nuances available in clinical progress notes and therefore have low predictive accuracy. Our project, titled “Can suicide theory-guided natural language processing of clinical progress notes improve existing prediction models of Veteran suicide mortality?” resulted in three key findings, each highlighting the transformative potential of theory-guided natural language processing (NLP) in the domain of suicide risk assessment.

Finding 1. Suicide Theory Concepts Are Documented in Clinical Progress Notes. Suicide theory provides a framework for understanding critical factors that drive suicidal thinking and behavior. Our study adopted the 3-Step Theory of suicide (3ST), which is based on the concepts of hopelessness, pain (especially psychological pain), connectedness, and capability for suicide. Despite their absence from structured EHR data, our analysis revealed that these concepts are frequently documented in clinical progress notes.1 This finding underscores the unique richness of unstructured text as a resource for understanding patient thinking and behavior that goes beyond what is represented in structured EHR data.

Finding 2. Suicide Theory Concepts Can Be Quantified as a Measure of Intensity. Using NLP techniques, we extracted mentions of presence and absence of hopelessness, psychological pain, connectedness, and capability for suicide from progress notes, which were translated into quantifiable metrics that reflect intensity of these concepts during a given timeframe.2 These scores represent a novel approach to suicide risk assessment, offering a more nuanced and clinically rich perspective on patient risk compared to traditional structured data sources such as diagnoses, procedure codes, and healthcare utilization.

Finding 3. Suicide Theory Concept Scores Identify Unique Patients at Risk of Suicide. Interpreting the intensity scores for hopelessness, psychological pain, connectedness, and capability for suicide within the 3ST framework identifies patients with a lower disease burden, both in terms of physical and mental health, who are less likely to be identified as at risk for suicide based on structured EHR data. This suggests that the assessment of suicide risk can be improved through theory-guided NLP of clinical progress notes. This improvement might enable healthcare providers to better target interventions such as safety planning, counseling, or closer monitoring – resources that are often constrained and cannot be extended to all patients equally.

Our work highlights the broader implications of incorporating real-world unstructured clinical text into predictive analytics and is now being translated into practice through a partnership with the Program Evaluation Resource Center (PERC), an evaluation center in the VA Office of Mental Health. PERC seeks to leverage our research to better capture the complex nature of suicide risk for use in VA clinical decision support tools.

Furthermore, the use of suicide theory as a guiding framework ensures that the insights generated by NLP are grounded in an evidence-based understanding of suicidal thoughts and behavior. This approach not only enhances the interpretability of the findings but also aligns closely with clinical priorities, resulting in models that are more actionable and have face-validity for healthcare providers.

Our work exemplifies the potential of combining theory-driven insights with modern computational methods to address one of the most pressing challenges in mental healthcare. By using the detailed observations recorded in clinical notes, we are paving the way for more personalized, precise, and effective strategies to reduce suicide mortality among Veterans.

Reference

1. Meerwijk EL, et al. , “Development of a 3-Step Theory of Suicide Ontology to Facilitate 3ST Factor Extraction from Clinical Progress Notes,” Journal of Biomedical Informatics 2024;150:104582.
2. Meerwijk EL, et al., Computing 3-Step Theory of Suicide Factor Scores from Veterans Health Administration Clinical Progress Notes, (Submitted for publication).