HHO-optimized VAE-transformer framework for robust clinical phenotyping and EHR reconciliation in diabetes management

Effective Type 2 diabetes management requires reliable Electronic Health Record (EHR) auditing. While hybrid generative-attention models can predict medication usage to ensure documentation integrity, they often suffer from training instability under suboptimal hyperparameters. Furthermore, the efficacy of metaheuristic optimization in stabilizing these complex architectures remains systematically under-investigated. We propose a Variational Autoencoder (VAE)-Transformer framework optimized via Harris Hawks Optimization (HHO). The task is rigorously structured as a supervised mapping from non-medication clinical covariates (X) to a binary medication status ($$Y\in \{\text{0,1}\}$$). To strictly prevent data leakage, target-revealing variables (e.g., insulin administration and medication changes) were explicitly excluded from $$X$$. The VAE extracts noise-resistant latent embeddings from raw tabular logs, the Transformer captures complex clinical dependencies, and the HHO algorithm adaptively fine-tunes the continuous learning rate to maximize system stability. Evaluated on the Diabetes 130-US Hospitals dataset (> 100, 000 encounters) via strict patient-level grouped cross-validation, the model achieved 90.03% accuracy, 98.21% precision, an F1-score of 0.91, and an AUC of 0.953. Specificity exceeded 99% with only $$32$$ false positives. Ablation studies validated the synergistic contribution of all modules, while convergence analysis confirmed rapid loss stabilization within five epochs. Metaheuristic optimization of the learning rate substantially enhances the generalizability of hybrid deep learning models. The HHO-enhanced VAE-Transformer provides a highly reproducible, leakage-free framework for trustworthy AI-driven EHR reconciliation and automated clinical phenotyping. A novel VAE-Transformer framework for automated clinical phenotyping and EHR reconciliation Continuous learning rate optimized via the Harris Hawks Optimization (HHO) metaheuristic algorithm Achieved a superior accuracy of 90.03%, outperforming tabular deep learning baselines HHO-optimized architecture demonstrates state-of-the-art performance in complex clinical mapping Provides a computationally feasible for clinical integration, hyperparameter-robust tool for automated healthcare registries