Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China
- 1School of Public Health, Key Laboratory of Environmental Pollution Monitoring and Disease Control, Ministry of Education, Guizhou Medical University
- 2Guiyang Public Health Treatment Center
Abstract
Background: Inpatient traditional Chinese medicine (TCM) costs are an important component of hospital expenditure in China, yet their cost structure and drivers remain poorly characterized. This study aimed to identify the drivers of total TCM cost among inpatients in the TCM department of a tertiary hospital using interpretable machine learning.
Methods: This retrospective study used 1,532 de-identified inpatient records (2023–2025) from the TCM department of a tertiary hospital in Guiyang, China. Because the model includes cost-component variables (the total cost of TCM treatment items and the total cost of Chinese herbal medicines) as predictors, it is framed explicitly as a cost-structure explanatory model rather than a prospective prediction model, and its explanatory power partly reflects these cost components. Because the outcome was right-skewed, total TCM cost was log(1 + x)-transformed. Seven algorithms (linear, ridge, lasso, elastic net, decision tree, random forest and XGBoost regression) were compared using repeated 5-fold cross-validation (5 repeats) and an independent 8:2 hold-out test set. Model performance was assessed by RMSE, MAE and R ² , and permutation importance and Shapley additive explanations (SHAP) quantified variable contributions.
Results: The cohort had a median age of 61 years and 61.4% female patients; the median total TCM cost was 3,446.72 yuan (IQR 1,616.00–5,324.05). The random forest achieved the best performance (test R ² = 0.820 on the log scale; back-transformed RMSE = 1,408.6 yuan, R ² = 0.746), outperforming XGBoost and all regularized linear models. Permutation importance and SHAP consistently ranked the total cost of TCM treatment items, length of hospital stay and the total cost of Chinese herbal medicines as the three leading predictors (approximately 81% of the total SHAP importance), whereas demographic and clinical factors contributed little.
Conclusion: The total cost of TCM treatment items is the dominant driver of total TCM cost, followed by length of hospital stay and the cost of Chinese herbal medicines. These findings support refined cost accounting and standardized management of TCM treatment items. The reported contributions reflect associations within the model rather than causal effects; given the single-center design and internal validation only, multicenter external validation and admission-stage models are required before any clinical or administrative use.
Related
- Interpretable machine learning for identifying poor physical fitness in adults across three high-altitude regions of Tibet: a cross-sectional study
- Eye-movement-based machine learning for identifying Parkinson’s disease dementia in a retrospective cohort: an interpretable model-development study
- Radiomics-based interpretable machine learning model for identifying high-risk patients with small bowel obstruction: a multicenter study.
- Discovery of Q-Markers in traditional Chinese medicine based on machine learning: Application and verification in Xiaoyao Heji.
- Interpretable machine learning for identifying mild cognitive impairment among Chinese patients with obstructive sleep apnea: a multicenter cross-sectional study
BibTeX
@article{li2026interpretable,
title = {{Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China}},
author = {Xi Li and Shouyan Chen and Mei Gong and Changyang Chen and Jia Chen and Jiangping Zhang},
journal = {Frontiers in Public Health},
year = {2026},
month = oct,
doi = {10.3389/fpubh.2026.1937712},
url = {https://www.frontiersin.org/articles/10.3389/fpubh.2026.1937712/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China
AU - Xi Li
AU - Shouyan Chen
AU - Mei Gong
AU - Changyang Chen
AU - Jia Chen
AU - Jiangping Zhang
PY - 2026
DA - 2026/10/21
JO - Frontiers in Public Health
DO - 10.3389/fpubh.2026.1937712
UR - https://www.frontiersin.org/articles/10.3389/fpubh.2026.1937712/full
AB - Background: Inpatient traditional Chinese medicine (TCM) costs are an important component of hospital expenditure in China, yet their cost structure and drivers remain poorly characterized. This study aimed to identify the drivers of total TCM cost among inpatients in the TCM department of a tertiary hospital using interpretable machine learning. Methods: This retrospective study used 1,532 de-identified inpatient records (2023–2025) from the TCM department of a tertiary hospital in Guiyang, China. Because the model includes cost-component variables (the total cost of TCM treatment items and the total cost of Chinese herbal medicines) as predictors, it is framed explicitly as a cost-structure explanatory model rather than a prospective prediction model, and its explanatory power partly reflects these cost components. Because the outcome was right-skewed, total TCM cost was log(1 + x)-transformed. Seven algorithms (linear, ridge, lasso, elastic net, decision tree, random forest and XGBoost regression) were compared using repeated 5-fold cross-validation (5 repeats) and an independent 8:2 hold-out test set. Model performance was assessed by RMSE, MAE and R ² , and permutation importance and Shapley additive explanations (SHAP) quantified variable contributions. Results: The cohort had a median age of 61 years and 61.4% female patients; the median total TCM cost was 3,446.72 yuan (IQR 1,616.00–5,324.05). The random forest achieved the best performance (test R ² = 0.820 on the log scale; back-transformed RMSE = 1,408.6 yuan, R ² = 0.746), outperforming XGBoost and all regularized linear models. Permutation importance and SHAP consistently ranked the total cost of TCM treatment items, length of hospital stay and the total cost of Chinese herbal medicines as the three leading predictors (approximately 81% of the total SHAP importance), whereas demographic and clinical factors contributed little. Conclusion: The total cost of TCM treatment items is the dominant driver of total TCM cost, followed by length of hospital stay and the cost of Chinese herbal medicines. These findings support refined cost accounting and standardized management of TCM treatment items. The reported contributions reflect associations within the model rather than causal effects; given the single-center design and internal validation only, multicenter external validation and admission-stage models are required before any clinical or administrative use.
SN - 2296-2565
ER -