<![CDATA[In modern health insurance systems, accurate medical expenditure prediction is vital for financial solvency and equitable premium distribution. However, deployment is often hampered by the trade-off between predictive accuracy and model transparency. This study presents a unified, interpretable machine learning regression framework to evaluate personalized health insurance charges using structured tabular data. We benchmarked six predictive architectures (five tree ensembles and a multi-layer perceptron control) on a real-world dataset (n=1,338). Experimental results demonstrate that Gradient Boosting achieved superior performance with R2=0.8789, RMSE = $4,335.47, MAPE = 28.49%, and an operational model footprint of only 170 KB. In contrast, standard deep learning (MLP) failed catastrophically (R2 = −0.3947) due to severe right-skewness and nonlinear tabular feature interactions. Model interpretability via SHAP values identified smoker status (48.2% Gini importance) and BMI interactions as primary cost drivers. Future research will focus on evaluating hybrid TabNet-Boosting architectures and integrating longitudinal temporal claims to capture evolving risk profiles across multinational cohorts.]]>
Copyrights © 2026
