Machine learning models using wrist actigraphy and sleep log data demonstrated satisfactory performance in predicting clinician-rated PTSD diagnosis and probable PTSD (PCL-5≥31) among trauma-exposed male service members and veterans, with a combination of subjective and objective features being most impactful.
Key Findings
Results
A machine learning model predicting clinician-rated PTSD diagnosis achieved satisfactory performance using wrist actigraphy and sleep log data.
The diagnosis model achieved an AUC of 0.83 (95% CI 0.61-1.00)
High accuracy of 88% and specificity of 96% were observed
Moderate sensitivity of 63% was achieved
Model was trained using extreme gradient boosting with leave-one-subject-out cross-validation
Sample consisted of N=36 trauma-exposed male service members and veterans (mean age 41, SD 5.3 years)
Results
A machine learning model predicting probable PTSD based on a PCL-5≥31 threshold showed comparable satisfactory performance to the clinician-diagnosis model.
The PCL-5≥31 model yielded an AUC of 0.84 (95% CI 0.71-0.98)
Balanced sensitivity of 73% and specificity of 82% were achieved
Performance was comparable to the clinician-rated PTSD diagnosis model
Data were collected over 1 week of wrist actigraphy and daily sleep logs
Results
The machine learning model predicting probable PTSD based on the PCL-5≥38 threshold performed poorly.
AUC was 0.47 (95% CI 0.24-0.69), indicating performance near chance level
The relationship between model-predicted scores and the PCL-5≥38 outcome was nonsignificant (B<0.01; P=.89)
This contrasted markedly with the satisfactory performance of models using the lower PCL-5≥31 threshold
Results
A combination of subjective and objective rest-activity features was most impactful for predicting PTSD outcomes in the best-performing models.
Both subjective features (sleep logs) and objective features (actigraphy) contributed to model performance
This pattern held for both the clinician-rated diagnosis model and the PCL-5≥31 model
Candidate features were identified using univariate feature selection prior to model training
Results
The best-performing models predicted PTSD outcomes even when accounting for co-occurring non-PTSD mental health diagnoses, supporting discriminant validity.
Model-predicted scores were significantly associated with clinician-rated PTSD diagnosis (B=0.19; P=.002) after accounting for other mental health diagnoses
Model-predicted scores were significantly associated with probable PTSD based on PCL-5≥31 cutoff (B=0.24; P=.003) after accounting for other mental health diagnoses
Linear regression was used to assess discriminant validity of model-predicted scores relative to other mental health diagnoses
The PCL-5≥38 model showed no such discriminant validity (B<0.01; P=.89)
Background
The study identified discrepancies between clinician-rated PTSD diagnosis and probable PTSD derived from self-report measures as a key motivation for using objective wearable data.
Prior work has focused on predicting probable PTSD based on self-report measures rather than clinical diagnosis
Discrepancies exist between clinical diagnoses and probable PTSD derived from self-reports
The PCL-5 was used with two cutoffs (≥31 and ≥38) to represent different thresholds for probable PTSD
The study specifically targeted clinician-rated diagnosis as a primary outcome to address this gap
Methods
The study sample consisted exclusively of trauma-exposed male service members and veterans, collected over a one-week monitoring period.
N=36 participants with mean age of 41 years (SD 5.3 years)
All participants were male service members and veterans
All participants were trauma-exposed
Data collection involved wrist actigraphy and daily sleep logs over 1 week
Extreme gradient boosting models were trained using leave-one-subject-out cross-validation
What This Means
This research suggests that wearable devices that track movement and sleep, combined with daily self-reported sleep information, can help identify whether trauma-exposed military veterans have post-traumatic stress disorder (PTSD). The researchers collected one week of wrist activity monitor data and sleep diary entries from 36 male service members and veterans, then used a type of machine learning called extreme gradient boosting to build models that predicted PTSD. The best models—one predicting a formal clinical PTSD diagnosis and one predicting probable PTSD based on a widely used questionnaire score—both performed well, correctly classifying most participants and showing accuracy rates around 83-88%.
A key finding was that combining both objective data (from the wearable device) and subjective data (from self-reported sleep logs) worked better than either type of data alone. The models were also able to distinguish PTSD from other mental health conditions, suggesting the rest-activity patterns being detected are somewhat specific to PTSD rather than just reflecting general mental health problems. However, a model using a stricter PTSD questionnaire score threshold performed no better than chance, highlighting that how PTSD is defined matters greatly for what machine learning tools can detect.
This research suggests that wearable technology combined with simple daily sleep diaries could one day support PTSD screening and assessment, potentially making it easier to identify individuals who need clinical evaluation without requiring lengthy in-person assessments. However, the study had a small sample of only male veterans, so further research with larger and more diverse groups is needed before these tools could be used in real-world clinical settings.
Check Your Own Numbers
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.