Predictive Linguistic markers of suicidal ideation in autobiographical memory narratives

Abstract

Suicide claims approximately 700,000 lives annually, yet risk identification remains challenging due to reliance on self-disclosure and limitations of current assessment methods. This study examined whether linguistic analysis of autobiographical memory narratives could predict suicidal ideation. Methods In total, 190 participants (137 clinical, 53 community) provided autobiographical memories to twelve emotionally evocative Turkish cue words, analyzed with Linguistic Inquiry and Word Count (LIWC-22) software to extract 90 linguistic features. Full-sample screening identified candidate predictors, interpreted with SHAP analysis. For unbiased validation, feature selection and classification were repeated within a fully nested cross-validation across nine algorithms, with predictors selected independently within each fold and performance compared against a label-permutation null. Results Discrimination was modest but consistent across all nine algorithms, with the best models reaching an AUC of 0.70, exceeding chance (permutation p < .001). In most models, individuals classified as having suicidal ideation were 1.5 to 1.7 times more likely to actually report it than the base rate. Five linguistic markers, spanning anger-, fear-, and distress-cued memories, were selected in every fold and associated with higher risk. Four remained significant after symptom-severity adjustment, and all five increased with ideation severity. Exploratory SHAP analysis revealed emotion-specific effects, with cognitive-mechanism words in happy memories and impersonal pronouns in fear contexts associated with lower risk. Conclusions Linguistic analysis of autobiographical memory content carries a modest but genuine signal for suicidal ideation, and detecting risk from non-suicide-related narratives may help overcome disclosure barriers. Future work should prioritize longitudinal replication, external validation, and clinical translation.