false
OasisLMS
zh-CN,en,fr,de,ja,es
Catalog
35th AM (2025) - Poster Session
Real Time Risk Prediction of Treatment Outcome in ...
Real Time Risk Prediction of Treatment Outcome in Opioid Use Disorder Treatment
Back to course
Pdf Summary
This study presents a novel approach to real-time prediction of next-week relapse risk in patients undergoing opioid use disorder treatment, focusing on the role of behavioral complexity in predictability. Using data from 2,200 patients over 24 weeks, the authors leverage Permutation Entropy (PE) to quantify weekly behavioral complexity from categorical substance use histories (opioid, cocaine, heroin). They find that low-complexity patient histories are more forecastable, while high-complexity cases show less predictability, a nuance often obscured by standard evaluation metrics.<br /><br />By stratifying patients into complexity tiers based on PE, the model routes individuals to tailored machine learning (ML) models (LSTM-based), providing risk scores with uncertainty estimates that reflect each patient’s behavioral predictability. This complexity-calibrated evaluation (using stratified AUROC and PR AUC) offers a more honest assessment of model performance, highlighting wider error margins and poorer calibration in higher-complexity groups. Clinically, this allows practitioners to interpret relapse risk in context, prioritize high-PE patients for intensified follow-up, and tailor treatment decisions such as dosing or adherence support based on complexity-informed uncertainty.<br /><br />The study underscores that uncertainty measurements correspond with behavioral complexity, suggesting that clinical decisions informed by both risk and complexity-aware uncertainty can better address relapse risk. Limitations include relatively small data size and non-comparable dosing scales across treatment types (buprenorphine oral/depot, methadone), which the team plans to address in future work by expanding datasets (e.g., MIMIC) and refining dose normalization. Multi-week-ahead predictions are also a future target.<br /><br />In summary, this approach innovatively integrates permutation entropy to quantify behavioral complexity, improving relapse risk prediction and enabling actionable, uncertainty-aware clinical decision support in opioid use disorder treatment.
Keywords
opioid use disorder
relapse risk prediction
behavioral complexity
permutation entropy
machine learning
LSTM model
uncertainty estimation
stratified evaluation
clinical decision support
substance use history
×