ADAPTIVE MULTI-OBJECTIVE FEDERATED LEARNING FRAMEWORK FOR ENERGY-EFFICIENT DISTRIBUTED HEALTHCARE IOT SYSTEMS

Abstract

The rise of the Healthcare Internet of Things (H-IoT) has led to continuous monitoring of patients with remote sensors, innovative medical appliances, and edge computing systems. Although the idea of Federated Learning (FL) has shown to be a good means for training models without revealing sensitive data, ordinary FD methods usually need arbitrary client participation, which leads to great costs of communications, energy consumption, training time, and less adaptability of the systems. This is especially true for battery-powered IoMT systems with changing networks. The present paper represents the idea of Adaptive Multi-Objective Federated Learning (AMO-FL) that makes possible the efficient use of resources in the distributed environment of the healthcare IoT. The suggested method offers a mechanism that selects clients for a given federated training iteration resulting in balancing many goals including prediction accuracy, energy use, communication speed, quality of data, and reliability of devices. The suggested technique includes a composite metric, taking into account all aspects pf performance of the clients participating in training thus enabling the possibility to choose the best possible clients considering resource use and effectiveness. The system ensures privacy of the patients as the healthcare data remains at the local devices while only the information about the model is communicated. To test the validity of the proposed method the experiments are to be held using relevant health care datasets in comparison with well-known federated learning methods including FedAvg, FedProx, SCAFFOLD, and Random Client Selection. The method is assessed through the use of machine learning performance metrics (Accuracy, Precision, Recall, F1-score, and AUC) as well as system-level performance metrics (energy consumption, communication cost, training duration, convergence speed). The expected results show that the suggested method ensures high prediction performance while being highly economical in energy spent and communication level and therefore it can be used in advanced systems for healthcare applications.