Psycholinguistically Grounded Explainable Screening for Suicide Risk Assessment in Short-Form Text: A Transparent Feature-Fusion Approach

Automated identification of individuals exhibiting suicide-related language in digital text carries profound clinical and ethical weight. While transformer-based architectures yield impressive accuracy figures on curated benchmarks, their representations remain inaccessible to practitioners who must justify and assume responsibility for clinical decisions. This paper proposes a multi-class stratification architecture grounded in computational psycholinguistics, deliberately constructed around three transparent feature families: lexical salience scores derived through term-frequency weighting, hand-crafted cognitive-behavioural markers anchored in established clinical theory regarding suicidal cognition, and affective polarity measurements extracted via two complementary rule-based sentiment tools. A controlled feasibility corpus of 200 annotated samples—distributed across Low, Medium, and High risk strata—serves as the experimental substrate. Three inherently interpretable classifiers are benchmarked under stratified five-fold cross-validation with recall accorded priority weighting to reflect the asymmetric consequences of undetected high-risk instances. Linear SVM achieves the strongest aggregate performance (weighted F1 = 0.7156, recall = 0.7250, precision = 0.7216), while Logistic Regression exhibits the lowest fold-to-fold variance. Prediction transparency is demonstrated through coefficient ranking and LIME-based local feature attribution, enabling practitioners to audit the evidence underlying each classification. The framework is explicitly positioned as a pre-diagnostic screening signal requiring mandatory qualified-professional review prior to any action.