Optimizing and Adapting Neural Network Models for Healthcare and Federated Learning

QC 20260427

Abstract

Neural Networks (NNs) have demonstrated considerable capabilities across a diverse set of fields, including Natural Language Processing (NLP), image classification, and regression. However, deploying increasingly large Deep Learning (DL) models introduces practical challenges, especially in the healthcare domain: high memory and computational costs, and inefficient training when data is noisy or corrupted. In this thesis, we design, evaluate, and analyze novel Artificial Intelligence (AI) models for Federated Learning (FL) and Electrocardiogram (ECG)-related tasks, whose common rationale is the incorporation of additional domain knowledge into the model structure, thus optimizing the training pipeline and reducing model complexity.

As the first contribution, we analyze the distributed case with multiple low-powered devices in a federated scenario. Cross-device FL is a branch of Machine Learning (ML) where multiple participants train a common global model without sharing data in a centralized location. In this thesis, a novel technique named Coded Federated Dropout (CFD) is proposed to carefully split the global model into sub-models, thus increasing communication efficiency and reducing the burden on the devices. We showcase our results for an example image classification task.

As the second contribution, we consider the anomaly detection task on ECG recordings and show that including prior knowledge in NNs models drastically reduces model size, inference time, and storage resources for multiple state-of-the-art NNs. In particular, we focus on Autoencoders (AEs), a subclass of NNs, which is suitable for anomaly detection. We propose a novel approach, called FMM-Head, which incorporates basic knowledge of the ECG waveform shape into an AE. The evaluation shows that we improve the Area Under The ROC Curve (AUROC) of baseline models while guaranteeing under-100 ms inference time, thus enabling real-time monitoring of ECG recordings from hospitalized patients. 

As the third contribution, we propose a Graph Neural Network (GNN)-based model that exploits the spatial relationships between ECG leads to reduce the number of electrodes required for ECG Imaging (ECGI), while achieving higher cardiac activation map reconstruction accuracy compared to the baseline. This work may thus contribute to making practical ECGI more feasible, as its clinical adoption is often constrained by the cost of the electrode vest and the burden of placing a large number of sensors.

As the fourth contribution, we conduct the first comprehensive retrospective multicentre DL study for cardiac arrest aetiology classification from early post-Return of Spontaneous Circulation (ROSC) ECGs, evaluating multiple DL and foundation models across two datasets. We experimentally demonstrate that model performance decreases with the time elapsed since ROSC, and that the best-performing models focus on physiologically relevant ECG segments, as evidenced by saliency map analysis. 

Finally, several directions for future work are presented. The use of prototypical networks for cardiac arrest aetiology classification from ECGs may provide example-based explanations, helping clinicians better interpret the predictions of AI models when noisy and disturbed inputs are involved, as is the case for post-ROSC ECGs. Additionally, this thesis advocates for further investigation into techniques that integrate ECG anomaly detection into distributed and federated settings.

Link to DiVA