Spectrogram-Driven Convolutional Neural Network for Real-Time Non-invasive Hyperglycaemia Detection in Paediatric Type-1 Diabetes via Wearable Sensors

Abstract

Real-time detection of glycaemic events is crucial in the effective management of type 1 diabetes, particularly in paediatric patients. Recent advances in wearable sensors and machine learning have allowed for the inference of glycaemic events based on non-invasive physiological signals such as electrocardiogram (ECG). However, existing approaches have limitations due to the limited number of ECG features analysed and their applicability to real-life conditions. To overcome these limitations, we propose a spectrogram-driven deep learning methodology for real-time glycaemic event detection. We calculated beat-level spectrograms using Short Time Fourier Transform (STFT) on ECG beats extracted from continuous signals using our deep learning ECG segmentation tool. Subject-specific multi-layer 2D convolutional neural networks were trained on the spectrograms. We evaluated our methodology on an original dataset comprising continuous ECG and interstitial glucose data collected from children with type-1 diabetes over several days in real-life conditions. Our approach achieved an average personalised hyperglycaemia detection accuracy of 96.9%.