Abstract
Background: Posttraumatic stress disorder (PTSD) is a multifactorial condition shaped by numerous biopsychosocial risk and protective factors. Machine learning methods provide powerful tools to model these complex and interacting influences on symptom expression, although many existing approaches rely on data that are impractical for large-scale application.Objective: In this study, we combined accessible psychometric measures with heart rate variability (HRV) to model vulnerability and resilience factors associated with PTSD symptoms in trauma-exposed individuals from a nonclinical sample. Importantly, a longitudinal follow-up complemented the cross-sectional design, allowing us to examine whether early psychological and physiological markers predict subsequent PTSD symptoms.Method: Regression models were implemented using a sparse multiple kernel learning (MKL) approach, allowing an estimation of the relative contribution of predictors both at the construct level (e.g. psychometric and HRV measures) and at the individual item level. The sample comprised 176 undergraduate students who completed psychometric scales assessing tonic immobility, optimism, positive and negative affect, coping strategies, and childhood maltreatment. HRV during exposure to negative stimuli was included as an autonomic predictor. PTSD symptom severity (PCL-5) was assessed cross-sectionally (T1) and after 18 months (T2).Results: MKL successfully predicted PTSD symptom severity at both time points. Across cross-sectional and longitudinal models, negative affect was identified consistently as the most influential predictor at T1 and T2. Additional relevant predictors included tonic immobility, childhood maltreatment, and coping strategies.Conclusions: These findings emphasise the potential of combining easily measurable predictors to enhance the early detection of PTSD risk and to guide the development of more personalised prevention and intervention strategies.