Interpretable Machine Learning Reveals Key Dose Trajectory Patterns Predicting Success of Acupuncture-Assisted Methadone Tapering

Background: Long-term methadone maintenance treatment (MMT) requires gradual dose reduction to mitigate adverse effects; prior research has shown that acupuncture facilitates dose tapering and alleviates opioid cravings. Although medication dosage parameters hold significant clinical value in addiction medicine, the impact of early outpatient dose data on therapeutic effects currently remains unclear. This study aimed to construct a methadone dose reduction prediction model and to analyze the clinical value of historical dose trajectories. Method: Data from two randomized trials (N=197 patients across six Chinese MMT clinics) were analyzed, with participants grouped into the acupuncture and non-acupuncture cohorts. The primary outcome was combined outcome comprising methadone dose reduction and craving score changes. Preintervention dose trajectories were derived via cluster analysis and merged with baseline features. Five machine learning models were trained using SHAP to explain the feature contributions. Subgroup analyses linking trajectories to the effects of acupuncture were conducted. Results: Methadone dose data were clustered into three trajectories. Model training included nine features from eleven variables. The CatBoost model achieved the best performance on the test set (area under the curve=0.7531, accuracy=0.8205). The SHAP summary plot revealed that the three most influential factors in methadone dose reduction were intervention type, body mass index, and dosage trajectory. Subgroup analysis showed that trajectory class 2 exhibited significantly better efficacy than class 0 at weeks 2 and 4 of acupuncture (week 2: risk difference, 20.4%, P =0.015; week 4: risk difference, 27.5%, P =0.013). Conclusions: In this study, we established a trajectory-based prediction model for MMT dose reduction and demonstrated the clinical value of historical trajectories. The results suggest that acupuncture optimally supports patients with dynamic "rise-then-decline" trajectories, advancing personalized MMT strategies.