Advancing Suicide and Self-Harm Research with AI: A Focus on Safety-Net Psychiatric Hospitals
Identifying suicides and self-harm is a critical medical and public health priority, especially among underserved populations in safety-net psychiatric hospitals. Despite the urgency, clinical data to support rapid research progress — particularly using artificial intelligence (AI) — is limited. This article explores innovative efforts to enhance our understanding of suicidal events and related factors, as documented in clinical notes within these settings. The goal is to establish a gold standard suicidality corpus, conduct comprehensive manual content analysis, and employ natural language processing algorithms for automated text classification.
Methods and Analysis
To address this challenge, a multidisciplinary panel developed an annotation guideline focused on capturing four key suicide-related factors: suicidal ideation (SI), suicide attempt (SA), suicide exposure, and nonsuicidal self-harm. Utilizing a clinically validated annotation process, an annotated corpus of 500 notes was created. Subsequent cohort analysis provided insights into demographic and suicidality distributions. A pre-trained language model facilitated automatic classification, underscoring its potential to revolutionize data analysis in psychiatric research.
Results
The annotated corpus achieved a remarkable Cohen’s kappa of 0.95, indicating high data quality and consistency in the annotation process. Upon deidentification for data sharing, analysis revealed that 79.4% of notes contained one or more suicide-related terms. Notably, suicidal ideation and suicide attempt were the most frequently co-occurring terms, at 35.6%. The cohort’s mean age was 33.4 years, with 51.7% identified as men and 75.8% as single individuals. Key stressors impacting this cohort included unemployment (24.2%), homelessness (12.0%), limited access to healthcare (5.4%), and legal challenges (5.0%).
Four primary findings emerged to enhance suicidality documentation: implicitness, conflict, ambiguity, and incomplete definitional coverage. The base model demonstrated satisfactory performance in multi-label classification with a micro-averaged F1 score of 0.70.
Diploma
The near-perfect agreement among annotators underscores the robustness of the proposed annotation process and data quality. The cohort analysis provides critical insights into the distribution and documentation of suicidality findings. Data modeling highlights the transformative potential of AI-powered methods for evaluating extensive clinical notes, offering a promising avenue for future research initiatives.
To explore this innovative approach further, visit the source Here.
“`
