Detection of Posttraumatic Stress Disorder With Rest-Activity Data: Machine Learning Approach Using Wearable and Self-Report Data

Abstract Background Growing evidence suggests that disruptions in rest-activity rhythms may serve as relevant markers of posttraumatic stress disorder (PTSD). Despite the emergence of machine learning methods applied to actigraphy and self-report data, few studies have used these approaches to identify individuals with clinically diagnosed PTSD. Prior work has focused on predicting probable PTSD based on self-report measures, yet discrepancies exist between clinical diagnoses and probable PTSD derived from self-reports. Objective This study explored whether wrist actigraphy and sleep logs could be used to accurately predict clinician-rated PTSD diagnosis and probable diagnosis of PTSD based on established self-report cutoffs (PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition [PCL-5] ≥31 and ≥38) among trauma-exposed service members and veterans. We also explored which features were most strongly predictive of each outcome and whether models were able to predict PTSD diagnosis even when accounting for other mental health disorders. Methods Wrist actigraphy data and daily sleep logs were collected over 1 week from trauma-exposed male service members and veterans (N=36; mean age 41, SD 5.3 y). Candidate features were identified using univariate feature selection. Extreme gradient boosting models were trained using leave-one-subject-out cross-validation to predict the diagnosis of PTSD and probable diagnosis of PTSD based on 2 self-report cutoffs (PCL-5≥31 and ≥38). Performance metrics were then calculated at the person level. Linear regression was used to assess the discriminant validity of model-predicted scores and each PTSD outcome specifically, relative to other mental health diagnoses. Results Machine learning models predicting PTSD diagnosis and probable PTSD based on the PCL-5≥31 threshold demonstrated satisfactory performance in this sample. The diagnosis model achieved an area under the curve (AUC) of 0.83 (95% CI 0.61‐1.00), with high accuracy (88%) and specificity (96%) and moderate sensitivity (63%). The PCL-5≥31 model yielded comparable performance (AUC=0.84, 95% CI 0.71‐0.98) with balanced sensitivity (73%) and specificity (82%). For both models, a combination of subjective and objective features was the most impactful. These models were able to predict PTSD even when accounting for non-PTSD mental health diagnoses, as model-predicted scores were significantly associated with 2 outcomes: clinician-rated PTSD (B=0.19; P =.002) and probable PTSD based on a PCL-5≥31 cutoff (B=0.24; P =.003). In contrast, the model predicting probable PTSD based on the PCL-5≥38 threshold performed poorly (AUC=0.47, 95% CI 0.24‐0.69), with a nonsignificant relationship between model-predicted scores and the outcome (B<0.01; P =.89). Conclusions Both subjective and objective rest-activity features may improve the prediction of PTSD. Further research is needed to validate these findings and explore the use of integrating wearable sensor data and subjective information to support PTSD assessment.