Maestría en Analítica Aplicada · 2026

The "how" over the "who": A machine learning approach to understanding clinical judgment in acute suicide risk assessment

Resumen

Background: Suicide is a pressing public health crisis in Bogotá, Colombia. Despite escalating rates, a gap exists in data-driven research using localized, real-world clinical data. This study aimed to train and interpret a machine learning model to predict acute suicide risk by leveraging official epidemiological records from Bogotá's public health surveillance system. Methods: This retrospective study analyzed 24,536 anonymized suicide risk cases (2021–2022) from the SISVECOS registry. From an initial 74 variables, 38 critical sociodemographic and clinical features were selected. Various algorithms were evaluated using 10-fold cross-validation, with the final model selected based on the macro F1-score. Model interpretability was achieved using feature importance analysis. Results: A Linear Discriminant Analysis classifier demonstrated high performance, achieving a final test set F1-score of 0.8771, an accuracy of 92.88%, and an AUC-ROC of 0.9568. The model's logic was driven by the mechanisms of the suicide mechanism; acute intoxication and high-lethality methods were the most powerful discriminators of a high-risk classification. Conversely, traditional sociodemographic factors showed minimal impact. These findings provide a clear, data-driven proxy for the Interpersonal-Psychological Theory of Suicide construct of acquired capability, the critical factor separating suicidal thought from lethal action. Conclusion: This study successfully trained a model that codifies the complex triage logic of clinicians assessing acute suicide risk. The results confirm that in high-stakes scenarios, the "how" (method lethality) is prioritized over the "who" (patient demographics). This supports a view of the suicide attempt as a discrete leap continuum from ideation to action, validating a binary approach for acute risk assessment. By using existing national health data, this work provides a blueprint for a scalable decision-support tool and insight to standardize clinical assessment and inform suicide prevention strategies in Colombia.