1 - (SYM 38) Using Actigraphy and Machine Learning to Predict PTSD Diagnosis in Veterans

Background: Passive sensing through wearable technologies can provide a non-intrusive method of objectively measuring behavior related to mental health and well-being in real-world contexts. Passively sensed data can provide insight into symptoms of posttraumatic stress disorder (PTSD), such as sleep disturbances and psychomotor impairment. Less work has examined whether passively sensed actigraphy data can be used to predict PTSD diagnosis in veterans. The following preliminary research aims to understand how wrist actigraphy data and machine learning can be used to predict PTSD diagnosis in veterans.

Methods: Male veterans (N = 20) with prior exposure to trauma during their military service were recruited to participate in an observational study (mAge = 40.39, SD = 5.04; 61% White). Participants completed a baseline diagnostic assessment (Structured Clinical Interview for DSM Disorders [SCID-5]) and wore an actigraphy watch for a week. Four participants (22%) met diagnosis for current PTSD. Activity, light, and sleep data were used to compute mean, standard deviation, median, mode, maximum, minimum, skew, kurtosis, root mean square of successive differences (RMSSD), and entropy features. Extreme Gradient Boosting (XGBoost) modeling was used to predict PTSD diagnosis based on features from 2,912 hour-long intervals. Information gain (IG) values were used to identify the most important predictors. Data collection is ongoing, and final results are expected by Fall 2024.

Results: Model performance using preliminary data was fair, with an Accuracy rate of 74.6% and a Kappa value of .27 (Sensitivity = .81, Specificity = .47). Results indicated that the most important features were mean activity (IG = .09), standard deviation of activity (IG =.08), and RMSSD of activity (IG =.06). Model performance in predicting clinical PTSD diagnosis was slightly better than model performance in predicting probable PTSD diagnosis based on Posttraumatic Checklist (PCL-5) scores (PCL-5 >= 33; Accuracy = .63; Kappa = .25, Sensitivity = .63, Specificity = .64).

Discussion: Model performance using preliminary data indicated fair performance in predicting clinical PTSD diagnosis in a veteran sample. Activity-related features were the most impactful in predicting PTSD diagnosis, and performance in predicting clinical PTSD diagnosis was slightly better than performance in predicting probable PTSD diagnosis based on PCL-5 scores. The use of passive sensing and machine learning to assess PTSD in veterans will be further discussed.