Privacy-Preserving,Multi-Objective Whale-Bat Optimized XGBoost for Early Diagnosis of Insulin Resistance in Non-Diabetic Women: Development and Validation Using NHANES 2007–2020

Abstract

Insulin resistance (IR), a state of reduced tissue responsiveness to insulin, is a key precursor to type 2 diabetes mellitus and cardiovascular disease, yet scalable, noninvasive screening methods remain underdeveloped. Existing machine learning (ML) models for IR prediction primarily aim to maximize predictive accuracy, often neglecting critical deployment factors such as inference latency and model stability, and risking patient privacy by exposing individual biomarker contributions. We developed and validated a fully reproducible, privacy-preserving diagnostic framework for early IR detection that simultaneously optimizes predictive performance, inference speed, and model stability, while providing system-level interpretability without revealing individual biomarker details. Using publicly available data from the National Health and Nutrition Examination Survey (NHANES) 2007–2018 (n = 2,000) for model development and NHANES 2019–2020 (n = 2,144) for external validation, we built a pipeline centered on an extreme gradient boosting (XGBoost) model optimized by a novel multi-objective, bi-level Whale-Bat (WOA-BA) algorithm. This optimizer maximized area under the receiver operating characteristic curve (AUC-ROC), minimized inference latency, and reduced cross-validation variance. Compared to a baseline XGBoost model with default hyperparameters, our final model improved internal AUC-ROC by 4.8%, reduced inference latency by 33%, and halved AUC variance. It also outperforms previous ML models reported in the literature (AUC-ROC ≈ 0.82–0.88) by achieving superior accuracy and generalizability, with an internal AUC-ROC of 0.88 ± 0.02 and external AUC-ROC of 0.87. Feature-group SHapley Additive exPlanations (SHAP) aggregated predictors into Anthropometric/Demographic (42%), Lipid (35%), and Hepatic (23%) domains, providing actionable insights without exposing individual biomarkers. The low resource footprint enables privacy-preserving, real-time deployment on mobile and embedded devices. Additionally, comparative experiments demonstrated that our Whale-Bat (WOA-BA) optimizer outperformed NSGA-II and MOEA/D in simultaneously optimizing AUC, latency, and stability. These contributions establish a robust and reproducible approach to early IR screening using only publicly available data.