Towards Hardware Supported Domain Generalization in DNN-Based Edge Computing Devices for Health Monitoring

TL;DR

This work tackles domain shift in ECG data for on-device health monitoring by introducing correction layers (CLs) as a minimal, hardware-friendly DG method. A pre-trained DNN is augmented with a single trainable CL while the rest of the network remains frozen, enabling DG with low computational and memory overhead during training. Empirical results on AF classification show average $F_1$-score gains over 20% on generalized domains, while training compute decreases by more than $2.5 imes$ and memory usage by over $3 imes$, demonstrated on 22 nm CMOS ECG accelerators. The study highlights effective algorithm-hardware co-optimization for robust, privacy-preserving edge health monitoring with minimal changes to existing accelerators.

Abstract

Deep neural network (DNN) models have shown remarkable success in many real-world scenarios, such as object detection and classification. Unfortunately, these models are not yet widely adopted in health monitoring due to exceptionally high requirements for model robustness and deployment in highly resource-constrained devices. In particular, the acquisition of biosignals, such as electrocardiogram (ECG), is subject to large variations between training and deployment, necessitating domain generalization (DG) for robust classification quality across sensors and patients. The continuous monitoring of ECG also requires the execution of DNN models in convenient wearable devices, which is achieved by specialized ECG accelerators with small form factor and ultra-low power consumption. However, combining DG capabilities with ECG accelerators remains a challenge. This article provides a comprehensive overview of ECG accelerators and DG methods and discusses the implication of the combination of both domains, such that multi-domain ECG monitoring is enabled with emerging algorithm-hardware co-optimized systems. Within this context, an approach based on correction layers is proposed to deploy DG capabilities on the edge. Here, the DNN fine-tuning for unknown domains is limited to a single layer, while the remaining DNN model remains unmodified. Thus, computational complexity (CC) for DG is reduced with minimal memory overhead compared to conventional fine-tuning of the whole DNN model. The DNN model-dependent CC is reduced by more than 2.5x compared to DNN fine-tuning at an average increase of F1 score by more than 20% on the generalized target domain. In summary, this article provides a novel perspective on robust DNN classification on the edge for health monitoring applications.

Figures (8)

• Figure 1: Problem Overview
• Figure 2: An overview on the sources of domain shift in the application of ECG monitoring and the challenges of ECG data acquisition in general. Many databases differ in both measurement setup and labeling schemes making it challenging to combine them for DG.
• Figure 3: Sketch of state-of-the-art DG algorithms for ECG classification
• Figure 4: Insertion of Correction Layer into CNN from Loh2022. The training and validation of DG capability is performed in two stages. The first stage trains the CNN without CL on SD. The second stage trains the CL only on a subset of TD, while the remainder of TD is used for validation. Both stages are performed for different combinations of patients in SD and are cross-validated for statistically robust QoS evaluation.
• Figure 5: Feature distribution of a fully convolutional DNN for IN and CTL. Features are mapped to the 2D plane using principle component analysis (PCA).