Abstract
Healthcare insurance fraud imposes substantial financial burdens on health systems worldwide, with annual losses
estimated between 3–10% of total expenditure. Existing detection approaches frequently lack interpretability, struggle
with severe class imbalance, and do not leverage explainability methods at the ensemble architecture level. This study
proposes an explainable stacking ensemble framework for fraud detection in severely imbalanced healthcare claims
data, with the dual aim of achieving strong predictive performance and enabling transparent, systematic ensemble
optimisation through dual-level explainability. A publicly available dataset of 16,000 reimbursement records with 5%
fraud prevalence was used. Preprocessing included median imputation, outlier retention, and StandardScaler
normalisation. A two-stage feature selection pipeline: Pearson correlation filtering followed by Recursive Feature
Elimination with Cross-Validation (RFECV), reduced 82 variables to 20 predictive features. Twelve heterogeneous
base classifiers were screened and integrated via logistic regression as a meta-learner. SHapley Additive exPlanations
(SHAP) were applied at two levels: (i) feature attribution for claim-level and global interpretability, and (ii) model-
level contribution analysis to quantify base learner influence within the ensemble. Class imbalance was treated as an
experimental variable, with Synthetic Minority Over-sampling Technique (SMOTE), Adaptive Synthetic Sampling
(ADASYN), and original imbalanced data compared systematically. The full ensemble achieved a Receiver Operating
Characteristic–Area Under the Curve (ROC-AUC) of 0.9041, Precision-Recall AUC (PR-AUC) of 0.6206, and
Matthews Correlation Coefficient (MCC) of 0.5593 under five-fold stratified cross-validation. Training on the original
imbalanced data consistently outperformed resampled configurations in AUC-PR and MCC. SHAP diagnostics
identified redundant learners, permitting reduction to a seven-model configuration (Reduced-4) that preserved ROC-
AUC, PR-AUC, and MCC while reducing computational complexity by 42%. The reduced ensemble attained 81% precision with a false positive rate below 1% at the default 0.5 threshold. Financial expenditure variables, particularly
pooled payments and inpatient drug costs, emerged as dominant predictors. These results demonstrate that the primary
benefit of stacking lies in performance stabilisation and interpretability-guided model reduction rather than isolated
metric inflation. By embedding explainability at both feature and model levels, the proposed framework advances transparent and computationally efficient ensemble design for imbalanced healthcare analytics. External validation on
independently adjudicated datasets remains necessary prior to operational deployment.