Identifying and Predicting Pain Flares Among Veterans with Multilevel and Multimodal Data: A Gradient-Boosted Mixed-Effects Model Study

Background and Purpose: Chronic pain affects roughly half of U.S. military veterans and fluctuates substantially day-to-day, yet clinicians and patients lack practical tools to monitor risk for pain increases. Novel, multimodal data sources including wearable devices and daily diary surveys are promising for day-to-day risk monitoring. However, most prior research has examined whether wearable and daily diary data are associated with same-day pain, leaving open whether these data can also forecast tomorrow’s pain. Another open question is whether wearable data alone may be sufficient for pain risk monitoring, a question with practical importance given the burden of repeated self-reporting. This study tested whether multimodal data can be used to identify same-day pain increases and predict next-day pain increases in veterans, and whether wearable-only models performed comparably for pain risk monitoring.

Methods: Veterans with chronic pain (n=70) completed smartphone-based daily diary surveys and wore Fitbit or Apple Watch wearable devices for up to three months, providing approximately 4,700 person-days of data. The primary outcome was a clinically meaningful pain increase, defined as a pain rating of 2 or more points above each participant’s average pain level, in line with established minimal clinically important difference thresholds. Gradient-boosted mixed-effects (GPBoost) models, which are designed for multilevel data (e.g., data with multiple days per person), were trained on 50 multiply-imputed datasets. Hyperparameters were tuned via 10-fold cross-validation per imputation and metrics pooled using Rubin’s rules. Predictors included wearable (e.g., sleep and physical activity) and daily diary (e.g., affect, stress, social interaction) features, represented as deviations from each participant’s average. Models were trained on participants’ first 80% of days and tested on their last 20% of days. Separate models were fit for same-day identification (contemporaneous) and next-day (lag-1) prediction, and fit with both combined (wearable + daily diary) and wearable-only feature sets. SHapley Additive exPlanations (SHAP) values characterized which features contributed most to predictions.

Results: GPBoost models combining wearable and daily diary data achieved Area Under the Curve (AUC)=0.74 for same-day classification and AUC=0.75 for next-day prediction, with relatively balanced sensitivity and specificity. Wearable-only models retained the majority of combined model performance (same-day AUC=0.72; next-day AUC=0.71). SHAP analyses indicated that daily affect (stress, depressed mood, anxiety), sleep (efficiency, REM, time in bed, hours asleep), and physical activity (steps, stairs) features were each among the most informative predictors across model variants.

Conclusion and Implications: Combined models achieved adequate performance for both same-day identification and next-day prediction of pain increases. Wearable-only models retained most of the performance of combined models, suggesting that wearables alone may feasibly capture much of the predictive signal otherwise obtained through daily diaries in this sample. SHAP results are in line with biopsychosocial models of pain, pointing to a complex interplay of contributing affective, sleep-related, and physical activity factors. Findings support the feasibility of lower-burden monitoring for chronic pain in veterans and could inform the development of just-in-time interventions for pain. Replication in larger and more diverse veteran cohorts is needed to establish generalizability.