Vikram Abbaraju, Farhan Rahman, Michael J. Cho, Jeffrey C. Liu, Yi Yang, Calista E. Murray, Afra Nawar, John A. Berkebile, Joshua Chiok, Jaiyoun Choi, Rachel Bull, Lucy Shallenberger, Viola Vaccarino, Justine W. Welsh, Santosh Patel, Amit Shah, J. Douglas Bremner, Omer T. Inan
To enable timely intervention and reduce risk of returning to use in individuals with opioid use disorder (OUD), there is a need to monitor opioid withdrawal severity using physiological biomarkers. However, in the context of opioid withdrawal, prior studies have relied on data recorded in controlled lab settings over short periods of time and have only focused on a limited set of biomarkers. To address these limitations, using multimodal wearable signals, we extracted and characterized 32 cardiorespiratory features from 16 individuals with OUD experiencing varying levels of withdrawal severity over a 7-day period. Across all participants, we found statistically significant ( p < 0.05) differences in 6 features between high and low withdrawal. For machine learning-based classification of high vs. low withdrawal, we found that performance improved when leveraging respiratory-focused and sex-/BMI-stratified models (highest macro F1 of 0.812). Finally, in 14 out of 16 participants, at least 1 feature statistically significantly ( p < 0.1) tracked changes in withdrawal scores over the 7-day period. Therefore, our analyses of these features uncovered substantial phenotype sensitivity and physiological heterogeneity across the study cohort. These findings underscore the need for phenotype-aware and personalized modeling approaches for continuous, sensor-based monitoring of withdrawal severity in the future.