Prototyping an End-to-End Multi-Modal Tiny-CNN for Cardiovascular Sensor Patches

TL;DR

The paper tackles on-device cardiovascular monitoring by introducing an end-to-end, multi-modal CNN that fuses synchronized ECG and PCG signals for binary anomaly detection. It demonstrates competitive accuracy on Physionet 2016 training-a while achieving three orders of magnitude fewer parameters and FLOPs than state-of-the-art, enabling efficient edge inference. Through rigorous 10×5-fold cross-validation and hardware experiments on microcontrollers and sensor devices, it shows that on-device inference with a small CNN and an NPU can be more energy-efficient than continuous data transmission. The work highlights the benefits of ECG+PCG fusion, lightweight 1D bottleneck blocks, and quantified energy savings, while acknowledging dataset limitations and calling for larger, diverse datasets and robustness enhancements.

Abstract

The vast majority of cardiovascular diseases may be preventable if early signs and risk factors are detected. Cardiovascular monitoring with body-worn sensor devices like sensor patches allows for the detection of such signs while preserving the freedom and comfort of patients. However, the analysis of the sensor data must be robust, reliable, efficient, and highly accurate. Deep learning methods can automate data interpretation, reducing the workload of clinicians. In this work, we analyze the feasibility of applying deep learning models to the classification of synchronized electrocardiogram (ECG) and phonocardiogram (PCG) recordings on resource-constrained medical edge devices. We propose a convolutional neural network with early fusion of data to solve a binary classification problem. We train and validate our model on the synchronized ECG and PCG recordings from the Physionet Challenge 2016 dataset. Our approach reduces memory footprint and compute cost by three orders of magnitude compared to the state-of-the-art while maintaining competitive accuracy. We demonstrate the applicability of our proposed model on medical edge devices by analyzing energy consumption on a microcontroller and an experimental sensor device setup, confirming that on-device inference can be more energy-efficient than continuous data streaming.

Figures (3)

• Figure 1: An overview of the system pipeline ibrahim2024end. After the ECG and PCG signals are measured the data is interpolated, downsampled and normalized. After that windows are extracted. The windows of ECG and PCG signals are concatenated and fed into our model which classifies them into normal or abnormal. A majority voting over multiple windows gives the final classification. In case of a cardiovascular anomaly the data is transmitted to a doctor.
• Figure 2: A schematic and a picture of the hardware components in the setup used for our second experiment described in Section \ref{['sect:eff-A']}. (a) The measured ECG signal is transmitted from the SNCE800 to the NHS52S04. From there, it is pre-processed with a high-pass, low-pass and notch filter to remove noise before it gets transmitted via BLE to a hypothetical doctor’s device which can for example display the ECG data. For simplicity’s sake, we skip any intermediary devices or connections between the NHS52S04 and the doctor’s device that would be necessary in real life. (b) Red: Custom board with the NHS52S04 chip. Blue: SNCE800 chip. Green: Three ECG electrode cables.
• Figure 3: Comparison of mean power consumption of ML and BLE communication on our experimental sensor device setup.