JMIR Publications

ABSTRACT

Background:

Posttraumatic stress disorder (PTSD) is a debilitating psychiatric condition that affects around 8% of adults in the United States across the lifespan. PTSD symptoms vary significantly across time in ways that are poorly captured by traditional assessment methods. Though research on the use of passively sensed digital biomarkers from smartphones to predict mental health symptoms shows promise, limited research has explored the prediction of PTSD symptoms.

Objective:

This study evaluated the use of individualized machine learning models trained on passively sensed digital biomarkers related to location and physical activity to idiographically predict PTSD symptoms across time.

Methods:

A sample of adults experiencing clinically significant PTSD symptoms (PTSD Checklist for DSM-5 (PCL-5) score ≥ 31; N = 76. participated in a 21-day intensive longitudinal study. During the study, participants completed four daily EMA surveys related to their PTSD symptoms, anxiety symptoms, and depression symptoms. In addition, participants provided continuous access to their passively sensed data from their smartphone’s accelerometer, pedometer, and GPS. Individualized long short-term memory (LSTM) networks were trained to predict PTSD symptoms using passively sensed features related to location and physical activity. Model evaluation, both across the sample and within-person, was performed using R2 and mean absolute error (MAE).

Results:

Individualized models captured a majority of the total variance in PTSD symptoms across time (R2 values = .54-.69). In contrast, individualized models accounted for substantially less within-person variance in PTSD symptoms over time (mean R2 values = .09-.12). While model-predicted scores were significantly related to self-reported PTSD symptoms, after accounting for co-occurring depression and anxiety symptoms (ps < .001), co-occurring mental health symptoms were stronger predictors of self-reported PTSD symptoms compared to model-predicted symptoms. Notably, those with greater symptom instability over the 21-day period had poorer model performance. There was significant within-person agreement in the ranking of impactful features that predicted different PTSD symptoms, but weak between-person agreement in the ranking of impactful features that predicted PTSD symptoms, demonstrating between-person heterogeneity in the features that were most impactful in predicting PTSD symptoms across time.

Conclusions:

Individualized machine learning models using passively sensed smartphone data can capture a majority of the total variance in PTSD symptoms across time, but within-person performance in idiographic prediction of PTSD symptoms is more variable. Integrating passively sensed digital biomarkers from wearable technology may help improve idiographic prediction of symptoms. Improving dynamic PTSD assessment and intervention requires a deeper understanding of the factors that influence intra-individual changes in PTSD symptoms across time, as well as the passively sensed digital biomarkers that can be used to predict those changes.