Cost-Sensitive Learning with Model Explainability for Imbalanced Medical IoT Classification Using Homomorphic Encryption

Abstract

Class imbalance in medical Internet of Things (IoT) systems presents a critical challenge, where rare but clinically relevant conditions are heavily outnumbered by normal states, leading to poor detection of minority classes. Cost-sensitive learning (CSL) effectively addresses this by assigning differential misclassification costs, yet existing approaches require model transparency that conflicts with privacy requirements in healthcare. This paper introduces ECSLF, a novel framework combining homomorphic encryption with cost-sensitive learning and explainability to enable privacy-preserving classification of imbalanced medical IoT data. Our framework makes four primary contributions: (1) Encrypted cost-matrix operations leveraging CKKS homomorphic encryption for weighted loss computation on ciphertexts with 12× computational overhead reduction compared to general FHE; (2) Encrypted model explainability via gradient-based feature attribution working entirely in the encrypted domain with 87% fidelity to plaintext methods, enabling healthcare professionals to understand predictions without exposing patient data; (3) Federated cost-sensitive learning with institutional cost heterogeneity, allowing multi-hospital collaboration while maintaining local autonomy over disease-specific cost matrices; and (4) Formal privacy guarantees proving ε-differential privacy preservation when computing cost-weighted gradients on encrypted data. Empirical evaluation on cardiac monitoring datasets demonstrates that ECSLF achieves 89% of plaintext CSL performance while maintaining differential privacy and meeting real-time clinical latency requirements (<100ms). The framework successfully reduces false negative rates for abnormal arrhythmia detection from 40% to 11% while preserving patient privacy across multi-institutional healthcare networks.