Exploring the feasibility of real-time on-device ECG biometric classification using quantized neural networks

Figures

• 

  Figure 1. Three-dimensional render of the self-developed single-lead ECG Holter device designed for continuous wearable monitoring. The system consists of three interconnected modules: (1) the main digital/analog module housing the STM32H7 microcontroller, analog front-end, and wireless communication interface; (2) the ground reference electrode providing common-mode noise rejection and patient isolation; and (3) the right-arm (RA) and left-arm (LA) sensing electrodes responsible for ECG signal acquisition. All components are connected through low-profile shielded cables to ensure minimal motion artifacts and user comfort. The modular design enables both high-quality signal acquisition and efficient on-device neural network inference for real-time biometric classification.

• 

  Figure 2. Training and validation loss over 50 epochs for the PTB dataset. The model’s validation loss steadily declined during early epochs, reaching a minimum of 0.0289 at epoch 35.

• 

  Figure 3. Visualization of convolutional layer outputs. The original ECG signal (top) is gradually transformed as it moves through the network. Early layers keep high-frequency components, while deeper layers capture broader waveform patterns. (a) Original signal. (b) Convolved signals at Layer 1. (c) Convolved signals at Layer 2. (d) Convolved signals at Layer 3. (e) Convolved signals at Layer 4.

• 

  Figure 4. The architecture of the proposed CNN, consisting of four convolutional layers followed by a fully connected classifier. The model gradually transforms ECG signals into feature representations suitable for classification.

• 

  Figure 5. Operation complexity breakdown of the quantized ECG model on STM32H7. This figure presents a layer-wise analysis of computational complexity, comparing the number of Multiply-Accumulate (MAC) operations (blue bars) with the actual computational cost (c_macc contribution, red bars), which accounts for memory access and processing overhead.

• 

  Figure 6. Embedded neural network inference timeline, showing execution stages in both milliseconds (ms) and CPU cycle counts. The processing pipeline is divided into five stages: (1) downsampling, (2) standardization, (3) first inference (signal quality assessment), (4) weights reload, and (5) second inference (biometric classification). The total execution time of approximately 1.35 s ensures real-time processing for 5-s ECG data blocks.