Detection of suicidal ideation with machine learning using biopsychosocial and linguistic correlates

Suicide represents a global public health crisis, causing over 700,000 deaths annually. Traditional suicide risk assessment relies heavily on direct disclosure, which is often limited by stigma and shame. Individuals may deny suicidal ideation when directly asked by healthcare providers, with this rate reaching up to two-thirds in some studies. This dissertation presents three complementary machine learning studies that predict suicidal ideation without relying on directly suicide-related predictors (previous suicide attempts, explicit suicidal thoughts). This approach aims to overcome current assessment barriers while ensuring comprehensive coverage of dimensions of the biopsychosocial model. The first study used psychiatric symptom variables, the second study employed psychosocial factors outside of psychopathology, and the third study utilized variables obtained from linguistic analysis of autobiographical memory narratives cued by emotionally evocative words as predictors. In Study 1, an artificial neural network was trained using data from 924 university students with the DSM-5 Self-Rated Level 1 Cross-Cutting Symptom Measure, and external validation was tested with data from 361 university students. Study 2 examined 190 participants (clinical and control groups) using non-psychopathological measures including childhood adversities, attachment patterns, coping strategies, emotion regulation, and personality traits, with external validation on 84 participants. Study 3 investigated 190 participants' autobiographical memory narratives using Linguistic Inquiry and Word Count (LIWC-22) software with nested cross-validation. All studies employed explainable machine learning techniques (SHAP analysis) and multiple algorithms including neural networks, random forest, and support vector machines. Study 1 achieved AUC = 0.80 (external validation: 0.79), identifying personality functioning, depressed mood, and anxiety as top predictors. Study 2 demonstrated superior performance with AUC = 0.878 and F1 = 0.819, maintaining strong external validation (AUC = 0.791, F1 = 0.549), with loneliness as the strongest predictor alongside protective effects of conscientiousness and extraversion. Study 3 achieved the highest discriminative performance (AUC = 0.908, F1 = 0.727), revealing emotion-specific linguistic patterns including protective effects of impersonal pronouns in fear contexts and cognitive mechanisms in positive emotional memories. This dissertation demonstrates that suicidal ideation can be effectively predicted across multiple levels of the biopsychosocial framework without relying on direct suicide-related disclosure. The convergence of interpersonal disconnection as a critical factor across Studies 1 and 2, combined with Study 3's identification of cognitive processing patterns, provides comprehensive insights into suicide risk mechanisms. The performance hierarchy (linguistic analysis > psychosocial factors > psychiatric symptoms) suggests that novel approaches capturing complex behavioral patterns may offer superior predictive capability. These findings support the development of interpretable screening tools using multiple data sources. These findings demonstrate the importance of developing interpretable screening tools using multiple data sources. These tools can strengthen clinical assessment by reducing self-report limitations and contribute to early detection and prevention of suicidal behavior.