DSRA-Net: A Disagreement-Aware Stacked Ensemble Framework for Diabetes Prediction, Risk Stratification, Clinical Cost Optimization, and Healthcare Resource Allocation using Real-World Hospital Data

Abstract

Diabetes mellitus affects over 537 million adults globally and 77 million in India alone. Existing machine learning approaches are confined to binary classification on public benchmark datasets without extending predictions to clinical decision support. This paper introduces DSRA-Net, a five-stage clinical decision-support framework: (i) disagreement-aware stacked ensemble where Dᵢ = (pₓᵍᴺ,ᵢ − pₘₗₚ,ᵢ)² is an explicit uncertainty meta-feature; (ii) clinical cost-sensitive threshold optimization minimizing L(τ) = 5·FN + FP; (iii) four-tier Risk Priority Index stratification; (iv) constrained healthcare resource allocation; and (v) SHAP-based explainability. Validated on 100,000 real-world records from Nilvastu Pathology Lab, DSRA-Net achieves 97.19 % accuracy, Recall = 70.88 %, F1 = 81.09 %, ROC AUC = 0.97959. Seven feature-combination experiments (E1–E7) confirm DSRA-Net outperforms all individual models across all feature subsets. False negatives are reduced by 35 patients per 20,000 screened (McNemar χ² = 21.02, p < 0.0001). Clinical cost is reduced by 15.89 %. Practical deliverables include a Patient Risk Report Card and a Hospital Resource Allocation Report. Five-fold CV: F1 = 0.8109 ± 0.0004. Meta-learner: ŷᵢ = σ(−4.4726 + 12.7990·pₓᵍᴺ − 0.0969·pₘₗₚ − 0.0785·Dᵢ).