MULTIMODAL MACHINE LEARNING CLASSIFICATION OF SUICIDAL PHENOTYPES IN BIPOLAR DISORDERS

Background: Suicidal behavior is a leading cause of mortality in bipolar disorders (BD), yet risk stratification remains challenging. Given the multifactorial nature of suicide, this study aimed to characterize behavioral and neurocognitive correlates of suicidal phenotypes in BD using machine learning (ML).
Methods: 150 patients with BD and 120 healthy controls (HC) were included and classified using the Columbia-Suicide Severity Rating Scale into four groups: HC, BD without suicidal ideation/attempts (BD), with suicidal ideation (BD-SI), and with suicide attempts (BD-SA). A supervised multi-class ML framework was implemented using demographic, clinical, behavioral (mood, anxiety, impulsivity, hopelessness, affective temperaments), and neurocognitive variables (memory, executive functions [EF], inhibition, fluency, processing speed). Feature selection was applied for dimensionality reduction, and a Random Forest (RF) classifier was trained and evaluated using stratified k-fold cross-validation and external testing. Model explainability was performed using SHapley Additive exPlanations (SHAP) to quantify class-specific feature contributions.
Results: Feature selection identified 25 informative variables and RF classifier achieved an accuracy of 0.83 (weighted-F1=0.83). SHAP revealed class-specific feature patterns underlying model predictions of suicidal phenotypes (Figure1). HC showed low impulsivity, hopelessness, and preserved EF; BD showed low dysthymic/cyclothymic temperaments, low impulsivity, preserved EF and memory; BD-SI showed dysthymic/cyclothymic temperaments, moderate impulsivity, increased hopelessness, and EF and memory deficits; BD-SA showed marked impulsivity, impaired inhibitory control, dysthymic/cyclothymic temperaments, and reduced verbal fluency.
Conclusions: These findings support a neurocognitive-behavioral model of suicidal vulnerability in BD, whereby impaired top-down control facilitates the transition from ideation to suicidal attempts, with implications for risk stratification and prevention.