Predictors of Narcotic Use Treatment Readmission in the Thai Public Health Service System, Fiscal Years 2023-2025: A Big Data Analysis for Developing Continuous Care Models in the Community

Abstract

Background: Readmission to narcotic use treatment is a system-detected service outcome reflecting both unmet needs for continuing care and the system's capacity to bring people back into treatment; it is not equivalent to relapse. Thai national evidence on predictors, model performance, and the translation of such findings into continuing care models remains limited.

Objectives: To estimate and compare 365-day readmission proportions, analyse factors associated with time to readmission, develop and temporally validate prediction models, and synthesise a conceptual proposal for community continuing care.

Methods: This retrospective cohort used national data from Thailand's narcotic treatment and rehabilitation information system. One index episode per person, closed during fiscal years 2023-2024, was followed for 365 days using data through 30 September 2025. Kaplan-Meier and Cox proportional hazards models were applied, with piecewise Cox analysis for variables violating the proportional hazards assumption. Fiscal year 2023 formed the development set and fiscal year 2024 the temporal validation set. Three classification and three time-to-event models were compared, assessing discrimination and calibration following TRIPOD+AI. Findings were then synthesised with evidence on continuity of care and community-based treatment.

Results: Among 195,452 people, 21,715 (11.11%; 95% CI 10.97-11.25) were readmitted within 365 days: 7.64% in fiscal year 2023 and 14.38% in fiscal year 2024. Median time to readmission was 173 days. Factors associated with higher hazards included primary opium use (AHR 2.025) or heroin use (1.816), discharge after programme completion with ongoing use (1.800), unassessed use status at discharge (1.685), inpatient treatment (1.580), previous treatment, and mental health treatment history, with excess hazard concentrated in the first 30 days after discharge. Random Forest and XGBoost achieved AUROC 0.664 and 0.662, with a positive predictive value of 22.63% for Random Forest, and XGBoost-AFT achieved a C-index of 0.652. All models showed limited discrimination, underestimated observed risk, and performed worst among people using opium or heroin.

Conclusions: The between-year difference establishes neither a trend nor causality, and the models are not ready for individual decision-making. The findings can inform needs assessment, post-discharge follow-up planning, and continuing care development. The synthesis produced an Integrated Risk-Stratified Community Continuing Care Model (IR-CCM 365) specifying eight service components, three risk tiers, and four follow-up windows, offered as an initial conceptual proposal requiring co-design, content validation, feasibility testing, and effectiveness evaluation before implementation.