Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China

An interpretable random forest model identified the total cost of traditional Chinese medicine treatment items as the strongest contributor to inpatient traditional Chinese medicine expenditure in a retrospective study of 1,532 hospitalization records. Length of hospital stay ranked second and the cost of Chinese herbal medicines third, together accounting for approximately 81% of the model’s total SHAP importance. The findings describe the cost structure observed at one tertiary hospital in Guiyang, China, rather than causal effects or the performance of a model for predicting bills before admission.

The study used de-identified records from the traditional Chinese medicine department of a tertiary hospital, covering admissions from 2023 through 2025. Of 1,583 records initially identified, 1,532 remained after exclusions for missing key variables, implausible or extreme costs, and out-of-pocket proportions outside the logical range of 0 to 1. The median age of patients was 61 years, with an interquartile range of 47 to 71 years, and 61.4% were women. Median total traditional Chinese medicine cost was 3,446.72 yuan per hospitalization, with an interquartile range of 1,616.00 to 5,324.05 yuan.

Researchers compared seven regression algorithms: linear, ridge, lasso and elastic net models, a decision tree, random forest and XGBoost. The outcome was transformed as log(1 + cost) because medical expenditure was strongly right-skewed, and the two skewed cost-component predictors were transformed in the same way. Hyperparameters were tuned through repeated five-fold cross-validation with five repeats on the training data, while a separate hold-out set containing 20% of records provided internal test validation. Performance was assessed with root mean square error, mean absolute error and R-squared on the log scale, with selected errors also back-transformed into yuan. Analyses were performed in R version 4.4.1.

The optimized random forest performed best in the internal test set, reaching an R-squared of 0.820, a root mean square error of 0.666 and a mean absolute error of 0.353 on the log-transformed scale. Its back-transformed test-set root mean square error was 1,408.6 yuan. XGBoost was a close second, with an R-squared of 0.818 and a back-transformed root mean square error of 1,512.6 yuan. The regularized linear models produced R-squared values of about 0.78 to 0.79, while the single decision tree ranked last, with an R-squared of 0.732, a log-scale root mean square error of 0.813 and a mean absolute error of 0.441. The random forest therefore explained about 82% of the variance in log-transformed total costs, although this was an internal estimate rather than externally validated performance.

Cross-validation results were broadly consistent with the hold-out results. The random forest’s mean cross-validated root mean square error was 0.631, compared with 0.666 in the test set, while its training R-squared of 0.967 indicated some degree of overfitting. The study authors described the training-to-test gap as modest. Predicted and observed values clustered around the one-to-one line, with residuals distributed across the observed cost range and minor deviations at the high-cost end. These results concern internal validation at one center: the study did not test the model in independent hospitals or in a temporally separate cohort.

Two complementary methods were used to interpret the winning model. Permutation importance measured how much model accuracy deteriorated when each variable was shuffled, while SHAP, or Shapley additive explanations, assigned each feature a contribution for each prediction and averaged those contributions across the dataset. Both methods placed the total cost of traditional Chinese medicine treatment items first, length of stay second and Chinese herbal medicine costs third. Their combined SHAP importance was approximately 81%. The mean absolute SHAP values were 0.7508 for treatment-item costs, 0.1836 for length of stay and 0.0997 for herbal medicine costs, compared with 0.0868 for age and still smaller values for the remaining variables.

The leading treatment-item variable represented charges for services such as acupuncture, tuina, moxibustion and acupoint application. Because these charges are themselves components of total traditional Chinese medicine expenditure, their prominence partly reflects arithmetic structure rather than an independently discovered risk factor. The same caution applies to the herbal medicine cost variable. The model therefore functions as a cost-structure explanatory tool: it shows where spending variation is concentrated, but it cannot establish that reducing a particular service or shortening a stay would cause an equivalent reduction in cost.

Length of hospital stay nevertheless had substantial predictive importance in both interpretation analyses. It had a Spearman correlation of 0.566 with total traditional Chinese medicine cost and ranked second in the permutation and SHAP rankings. Longer admissions may coincide with accumulated nursing, bed, medication, retreatment and daily service charges, but the study stressed that stay length can also reflect disease severity, treatment complexity, delayed recovery or discharge-planning constraints. Payment method and the proportion of out-of-pocket expenses made smaller contributions, while detailed reimbursement ratios, benefit packages, income and other socioeconomic measures were unavailable. The reported relationships are therefore predictive associations, not causal effects.

Descriptive comparisons found differences in median costs by sex, marital status, ethnicity, occupation, payment method and critical illness status. For example, median total cost was 3,193.08 yuan for men and 3,604.80 yuan for women. Patients classified as critically ill had a median cost of 3,830.01 yuan, compared with 3,413.57 yuan among other patients. These group differences were exploratory and cannot establish independent contributions because the models lacked several major clinical variables, including diagnosis, disease severity, comorbidities and traditional Chinese medicine syndrome differentiation. Unmeasured socioeconomic, geographic, cultural and healthcare-access factors may also have influenced the observed patterns.

For hospital managers, the immediate implication is that detailed accounting of individual traditional Chinese medicine procedures may be more informative than focusing only on medicines or broad patient characteristics. The authors also argue that standardized management of treatment items could help identify unwarranted variation in service use, and that payment systems shaped by diagnosis-related group and diagnosis-intervention packet reforms may benefit from recognizing the distinctive cost profile of traditional Chinese medicine departments. These are operational possibilities rather than evidence that a specific intervention has already reduced expenditure or improved care.

The authors identified several limits on the findings: the retrospective data came from one center, the models were internally validated only, cost components were included as predictors, major clinical information was omitted, and machine-learning importance does not demonstrate causation. They also noted that repeated admissions were not explicitly excluded, which could affect sample independence despite the methods describing each admission record as an independent unit of analysis. A genuinely prospective administrative model would need to use information available at or shortly after admission, such as age, sex, payment method, admission diagnosis, critical illness status, comorbidities and planned treatment category, rather than charges generated during the hospitalization. Before clinical or administrative use, the researchers call for multicenter external validation, including a temporally distinct validation set where feasible.

Subject of Research: Interpretable machine-learning analysis of cost drivers in inpatient traditional Chinese medicine services

Article Title: Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China

Article References: Li, X., Chen, S., Gong, M., Chen, C., Chen, J., & Zhang, J. (2026). Interpretable machine learning for identifying cost drivers of inpatient traditional Chinese medicine services in a tertiary hospital in China. Frontiers in Public Health, 14, Article 1937712. https://doi.org/10.3389/fpubh.2026.1937712

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1937712

Keywords: traditional Chinese medicine, health economics, random forest, machine learning, hospital costs, SHAP analysis, retrospective study, cost accounting, medical expenditure, Guizhou China, Interpretable, machine