Personalized entropy-informed deep learning for identifying opioid misuse

Fluctuations in pain, stress and craving are thought to contribute to opioid misuse. Developing accurate prediction models is vital for intervention and prevention efforts. In this work, we leverage physiological data and semantic analysis of electronic health records to tackle the challenge of detecting opioid misuse. Utilizing personalized hierarchical deep-learning models, we analyze trajectories of predicted pain, stress and craving states with 10,140 hours of heart-rate data collected by wearables from patients on long-term opioid therapy. From these trajectories, we extract entropy features from nonlinear dynamical analysis and develop a novel relevance-based temporal fusion model of opioid misuse risk. We incorporate clinical data into a large language model to enhance opioid misuse risk detection. We then fuse these modalities to achieve an accurate opioid misuse risk assessment with area under the precision-recall curve of 0.94 ± 0.05. This study marks a substantial advancement in personalized prediction of addictive behavior by elucidating the entropic nature of underlying affective state dynamics. This study addresses opioid misuse prediction by integrating physiological data and electronic health records. Utilizing personalized deep-learning models, it achieves a high accuracy in risk assessment through entropy feature extraction and relevance-based temporal fusion, demonstrating effective intervention potential.