AI-Based Preprocessing and R-Squared Weighted Federated Aggregation for Robust Healthcare Signal Imputation

Abstract

Federated learning is an emerging paradigm for collaborative model training in healthcare applications where sensitive patient data cannot be centralized due to privacy concerns; however, missing values in physiological signals remain a major challenge, often leading to degraded model accuracy and unstable convergence. To address this issue, this work presents an AI based preprocessing framework that combines advanced imputation methods, including K Nearest Neighbors (KNN), Autoencoder reconstruction, and Generative Adversarial Networks (GAN), with classical baselines such as mean and median imputation. Building on this preprocessing stage, this study proposes a novel R squared weighted aggregation strategy that dynamically scales clients' contributions based on their prediction quality. The experiments were conducted on BIDMC physiological datasets with missingness rates ranging from 0.1 to 0.5 across multiple federated clients. Results show that weighted aggregation converges to lower RMSE and MAE than classical FedAvg, especially in the early rounds where high client heterogeneity exists. The proposed method adapts to client quality, thereby providing robust convergence and stability, resulting in balanced performance. These results demonstrate that combining AI based preprocessing with R squared weighted aggregation effectively improves the reliability of federated learning in healthcare IoT systems. The proposed approach presents a scalable and feasible solution for real world scenarios where data incompleteness and client heterogeneity are inevitable.