From bias to business insight: evaluating machine learning approaches for fair and explainable health insurance underwriting

Machine Learning (ML) can reveal the complex nonlinear relationships between applicant attributes, offering new opportunities to strengthen underwriting risk classification in insurance. However, concerns about algorithmic opacity and subgroup bias have hindered the use of ML in underwriting. Existing studies have mostly concentrated on claims prediction and pricing, with limited attention to classification, interpretability, and fairness. We evaluated three ensemble models, viz., Random Forest (RF), XGBoost, and LightGBM across three underwriting granularities, i.e., binary, three-class, and eight-class risk classification settings, using a benchmark dataset comprising 59,381 insurance applicants. Feature selection was conducted using Boruta to construct a parsimonious and feature space. Model performances were assessed using accuracy, Cohen’s kappa, and Matthews Correlation Coefficient (MCC), across training, validation, and test sets. Interpretability was examined using SHAP-based feature attributes and a gain-based importance measure. Fairness was evaluated using the Statistical Parity Difference (SPD) and Equal Opportunity Difference (EOD) across age and BMI subgroups. Robustness was assessed through bootstrap resampling (1000 iterations), probability threshold sensitivity (0.1–0.9), and ranking generalisation assessment (RGA), operationalised via AUROC and bootstrap stability. XGBoost achieved the strongest performance in the binary setting (test accuracy = 0.831, MCC = 0.624), the performance declined with increased class granularity, indicating intrinsic limits in fine-grained risk recovery. Body Mass Index and insurance age jointly account for over 40% of model importance. Fairness audits revealed relatively mild disparities across age groups but more pronounced differences across BMI categories, underlining the need for a bias-aware underwriting design that considers these factors. Bootstrap confidence intervals confirm stability of predictive and fairness metrics, while threshold sensitivity and RGA analyses indicate robust performance and stable ranking behaviour under sampling and decision perturbations. This study presents a reproducible, governance-oriented framework integrating predictive accuracy, explainability, fairness, and robustness, providing actionable guidance for responsible ML deployment in insurance underwriting.