Identifying predictors of interindividual differences in cortisol responder status to psychosocial stress using a machine learning approach.

Cortisol responses to acute psychosocial stress vary substantially across individuals, yet the underlying mechanisms remain incompletely understood. Although numerous person- and situation-related factors have been linked to cortisol reactivity, findings are inconsistent, and these factors are rarely examined jointly. To address this gap, we applied a machine learning approach-specifically, random forest classification-to identify key predictors of cortisol responder status while flexibly modeling nonlinear and interdependent relationships. The sample comprised 229 medical students who completed a stressful breaking bad news simulation. The random forest model classified individuals as cortisol responders or nonresponders (derived from baseline-to-peak increases) based on biological, psychological, and behavioral predictors, achieving an F1 score of .59 and an accuracy of 61.2%, both of which significantly exceeded performance expected under random label permutation. Sex emerged as the strongest predictor, with males showing a higher probability of mounting a cortisol response. While other predictors did not reach significance in the outcome permutation test, their observed relationships may still provide meaningful insights. Task duration was nonlinearly associated with cortisol response probability, with larger increases observed at longer task durations. Low levels of anxiety, more positive valence, and prior practical and theoretical experience in breaking bad news predicted higher response probabilities. Additionally, a neutral stress mindset and a moderate body mass index were linked to higher response probabilities, following an inverted U-shaped relationship. Notably, these relationship patterns were broadly consistent across sexes, although subtle differences in a small number of predictors warrant further investigation. Overall, these data-driven patterns highlight the complexity of cortisol responder status and underscore the value of multivariate approaches for studying these processes, while informing future research about important covariates.