Neurological mHealth Applications in a Federated Learning Enabled Mobile Edge Computing Framework

Abstract

Real-time mobile health (mHealth) applications are expected to proliferate over next generation networks with the recent advances in wireless communications and wearable sensing. As the reliability of mHealth applications is governed by the limited energy and computing resources of the user equipment (UE), mobile edge computing (MEC) has been introduced to provide UEs with proximal high-end computing power. Along with MEC, enabling real-time mHealth applications requires low latency machine learning (ML) intervention to provide high quality inference and predictions. However, sending health data over the network towards centralized cloud servers incurs elevated delays while exposing the data to privacy threats. This necessitates a shift towards edge machine learning (Edge ML) to provide model training and prediction tools at the network edge. In this thesis, we incorporate Edge ML and federated learning (FL) to enable real-time neurological mHealth monitoring applications. In particular, we benefit from the MEC and ML capabilities at both the UE side and MEC server side to propose a novel energy efficient distributed Edge ML framework to reliably satisfy the strict latency requirements of a real-time electroencephalogram (EEG)-based epileptic seizure prediction application. Specifically, we formulate an optimization problem to distribute the feature extraction and classification processes between the UE and MEC server in response to the wireless channel variations and computing resources allocation, while determining the optimal UE uplink transmit power. We demonstrate the effectiveness of the proposed approach using a real-time epileptic seizure prediction case study. We also propose a distributed feature extraction mechanism between the UE and MEC server to obtain the features used for epileptic seizure prediction at the UE. We model the distributed feature extraction mechanism as a joint energy-latency optimization problem and develop an effective solution approach that captures performance trade-offs in terms of reduced UE energy and latency. Moreover, we utilize a lightweight FL scheme from the literature that is based on K-means clustering in order to train the epileptic seizure prediction ML model. Finally, we present a comprehensive study for an immersive multisensory virtual reality (VR) system over a Tactile Internet (TI)-enabled next generation 6G network, where we couple Internet of Things (IoT) neurological sensing with the support engine (SE) of the IEEE 1918.1 architecture to initiate an adaptive feedback aiming to increase the overall perception and reach new levels of immersive experience. We highlight design alternatives and potential enablers of this system and utilize the quality of physical experience (QoPE) as the main criterion in addressing the latency-throughput quality of service (QoS) and the multisensory perception.