Prediction of Rare Disease Medication Transition Patterns Based on an Interpretable Machine Learning Model
Author(s)
Jingyi Li, .1, Danyang Wei, .1, Xuanqi Qiao, .1, Hongfei Gu, .2, Min Hu, PhD1.
1Department of Health Economics, School of Public Health, Fudan University, Shanghai, China, 2Hongmian Cancers and Rare Disorders Charity Foundation, GuangZhou, China.
1Department of Health Economics, School of Public Health, Fudan University, Shanghai, China, 2Hongmian Cancers and Rare Disorders Charity Foundation, GuangZhou, China.
OBJECTIVES: Despite the proven efficacy of disease-modifying therapies (DMTs) for multiple sclerosis (MS), financial toxicity remains a critical barrier in resource-constrained settings. This study aims to: (1) develop an interpretable model predicting DMT transition patterns, and (2) identify critical determinants of MS medication transitions.
METHODS: A longitudinal cohort of 455 Chinese MS patients was constructed using nationally representative survey data with complete matched observations across baseline (2022) and follow-up (2025) waves. The analysis incorporated 11 dynamic features including socioeconomic status (insurance coverage transitions, household income volatility), clinical factors, and health system determinants (medication reimbursement ratio fluctuations). Comparative model evaluation was conducted using nested cross-validation across three algorithms: light gradient boosting machine (LGBM), random forest (RF), and logistic regression (LR). The optimal model was selected for SHAP analysis.
RESULTS: A total of 455 MS patients were included. Comparatively, the LGBM model demonstrated the highest predictive performance among models with an area under the curve (AUC) of 0.84, outperforming LR (AUC: 0.67) and RF (AUC: 0.70) models. SHAP analysis indicated that reimbursement ratio increases (|SHAP| = 0.4988) drove DMTs treatment initiation among untreated patients, indicating price elasticity of demand for DMTs. Health status improvements (Δself-rated health: |SHAP| = 0.8151) predominantly explained treatment discontinuation, suggesting non-adherence post-symptomatic relief. Medical assistance coverage exhibited minimal predictive power (|SHAP| < 0.1), underscoring the necessity to strengthenrare disease security in supplementary medical insurances beyond the basic medical insurance.
CONCLUSIONS: This study demonstrated the potential of machine learning techniques in medication changes prediction for rare disease patients. Policy analysis demonstrates the superior efficacy of China's basic medical insurance over supplementary health insurance in improving access to Disease-Modifying Therapies (DMTs). Implementation priorities should emphasize: (1) institutionalizing evidence-based reimbursement ratio adjustment mechanisms to ensure rational medication utilization patterns, and (2) strengthening rare disease coverage in existing supplementary insurance programs through integrated risk-sharing mechanisms.
METHODS: A longitudinal cohort of 455 Chinese MS patients was constructed using nationally representative survey data with complete matched observations across baseline (2022) and follow-up (2025) waves. The analysis incorporated 11 dynamic features including socioeconomic status (insurance coverage transitions, household income volatility), clinical factors, and health system determinants (medication reimbursement ratio fluctuations). Comparative model evaluation was conducted using nested cross-validation across three algorithms: light gradient boosting machine (LGBM), random forest (RF), and logistic regression (LR). The optimal model was selected for SHAP analysis.
RESULTS: A total of 455 MS patients were included. Comparatively, the LGBM model demonstrated the highest predictive performance among models with an area under the curve (AUC) of 0.84, outperforming LR (AUC: 0.67) and RF (AUC: 0.70) models. SHAP analysis indicated that reimbursement ratio increases (|SHAP| = 0.4988) drove DMTs treatment initiation among untreated patients, indicating price elasticity of demand for DMTs. Health status improvements (Δself-rated health: |SHAP| = 0.8151) predominantly explained treatment discontinuation, suggesting non-adherence post-symptomatic relief. Medical assistance coverage exhibited minimal predictive power (|SHAP| < 0.1), underscoring the necessity to strengthenrare disease security in supplementary medical insurances beyond the basic medical insurance.
CONCLUSIONS: This study demonstrated the potential of machine learning techniques in medication changes prediction for rare disease patients. Policy analysis demonstrates the superior efficacy of China's basic medical insurance over supplementary health insurance in improving access to Disease-Modifying Therapies (DMTs). Implementation priorities should emphasize: (1) institutionalizing evidence-based reimbursement ratio adjustment mechanisms to ensure rational medication utilization patterns, and (2) strengthening rare disease coverage in existing supplementary insurance programs through integrated risk-sharing mechanisms.
Conference/Value in Health Info
2025-09, ISPOR Real-World Evidence Summit 2025, Tokyo, Japan
Value in Health Regional, Volume 49S (September 2025)
Code
RWD106
Topic Subcategory
Health & Insurance Records Systems
Disease
SDC: Rare & Orphan Diseases