Development and Real-World Evaluation of an Interpretable Machine Learning Framework for Predicting COPD-Related Healthcare Costs.

BACKGROUND

Chronic obstructive pulmonary disease (COPD) imposes a substantial economic burden. While machine learning (ML) enables data-driven cost prediction, real-world benchmarking comparing algorithms, data-splitting strategies, and feature optimization on continuous, skewed healthcare expenditures remains limited.

OBJECTIVE

To develop an interpretable ML framework for predicting annual COPD-related healthcare expenditures, assessing the impact of algorithm selection, train-test split ratios, and feature optimization.

METHODS

We analyzed a retrospective cohort of 396 COPD patients from a Taiwanese medical center. Annual healthcare costs were log-transformed and standardized due to extreme skewness. Five ML regression algorithms were benchmarked across four train-test split ratios. Recursive feature elimination with cross-validation (RFECV) applied within training sets identified an optimal predictor subset. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Performance was evaluated using R-squared (R2) and mean absolute error (MAE).

RESULTS

The 80:20 split yielded the most stable generalization performance. Gradient Boosting achieved the highest accuracy (test R2=0.365; MAE=0.258 on log-transformed scale, ~NTD 82,400 raw scale), outperforming baseline models. RFECV identified six key predictors-Age × BMI, malignancy, exacerbation frequency, Charlson Comorbidity Index, and DOSE index-retaining >95% of full-model performance. SHAP analysis confirmed their clinical relevance.

CONCLUSION

Gradient Boosting provides robust predictive performance for COPD healthcare costs. A parsimonious six-feature set enhances interpretability without compromising accuracy. This explainable ML framework supports cost-aware healthcare management, pending future multi-center validation.