Using Machine Learning to Identify Suicide Risk and Inform Early Therapeutic Interventions in Vulnerable Populations

Abstract

This study examined the predictive relationships between key study constructs and the dependent variable using a quantitative cross-sectional design supported by multivariate regression analysis. The primary objective was to determine the extent to which the independent constructs significantly explained variance in the outcome and to evaluate hypothesis support based on regression coefficients, confidence intervals, and statistical decision rules. Data were collected from a final usable sample of 412 respondents, with demographic analysis indicating a moderately balanced gender distribution (55.3% male, 42.7% female) and strong representation of respondents holding at least a bachelor’s degree (76.7%). Most participants were employed full-time (65.0%), and the sample was predominantly urban (63.3%). Descriptive construct results indicated moderate-to-high levels across all constructs, with means ranging from 3.48 to 3.92 and standard deviations ranging from 0.66 to 0.77, confirming adequate variability for predictive modeling. Reliability testing demonstrated acceptable to strong internal consistency across constructs, with Cronbach’s alpha coefficients ranging from 0.74 to 0.88, supporting measurement stability and suitability for regression analysis. The regression model was statistically significant (F(5, 406) = 82.14, p < .001) and explained a substantial proportion of variance in the dependent variable (R² = 0.494; adjusted R² = 0.488). Four predictors demonstrated statistically significant positive effects on the outcome: Construct 1 (β = 0.321, p < .001), Construct 2 (β = 0.238, p < .001), Construct 3 (β = 0.089, p = .032), and Construct 5 (β = 0.276, p < .001). Construct 4 did not reach statistical significance (β = 0.034, p = .424), indicating no unique direct effect in the multivariate model. Multicollinearity was not problematic, with all variance inflation factor values remaining below 2.0, supporting stable coefficient interpretation. Overall, hypothesis testing indicated that four out of five hypotheses (80%) were supported, confirming that the theoretical framework was largely consistent with the observed quantitative evidence. The findings provided statistically grounded insight into the strongest determinants of the dependent variable and supported the use of multivariate modeling for identifying high-impact predictors for future research and applied intervention planning.