Exploring CNN and XAI-based Approaches for Accountable MI Detection in the Context of IoT-enabled Emergency Communication Systems

Abstract

The ageing European population and the expected increasing number of medical emergencies put pressure on the medical sector and existing emergency infrastructures, which calls for new innovative digital solutions. In parallel, the increasing utilization of the Internet of Things (IoT) has enabled the collection of real-Time data, allowing for the autonomous detection of acute medical emergencies. In this context, this paper presents two distinct machine learning (ML) models that leverage electrocardiogram (ECG) sensor data to autonomously detect Myocardial Infarctions (MI), a leading cause of emergencies. These models are intended to be integrated into an IoT-enabled next-generation emergency communications system (NG112) capable of detecting emergencies, initiating emergency calls (eCalls), and providing relevant information to emergency call takers, which reduces response time. To realize this, two disparate models working on fundamentally different data structures are proposed and compared: A one-dimensional convolutional neural network (CNN) operating on the raw ECG signals and a GoogLeNet-based model trained on ECG images. The PTB-XL dataset is used to evaluate the proposed models, and the results indicate the 1D CNN exhibits a favourable trade-off between precision and recall for the eCall use case. Finally, the paper also discusses applying eXplainable AI (XAI) methods to achieve explainability for the ML models, paving the way for an accountable and reliable implementation in safety-critical systems.