Compressed Phonocardiogram Classification on STM32 boards w/o Neural-ART acceleration

In this work three deep learning models for sequential series were studied, optimized and evaluated (CNN, GRU and Hybrid CNN-GRU). Specifically, they were designed for phonocardiogram (PCG) classification in normal or abnormal recordings. Such classification was also assessed under varying compression using the compressive sensing approch from the ModAU project ModAu: Modernized Auscultation.

After this evaluation, the CNN was deployed on the STM32H747I-DISCO board and on the Nucleo-N657X0-Q board exploiting also the novel Neural-ART accelerator.

Table of Contents

• Compressed Phonocardiogram Classification on STM32 boards w/o Neural-ART acceleration
  • Requirements
  • Directories and files
  • Results
    • Citation

Requirements

• Pytorch 2.8.0 (+ all the packages in the requirements.txt file)
• Cuda 12.6 for parallelization
• STM32CubeMX 6.14
• STM32CubeIDE 1.19
• STM32CubeProgrammer 2.19

Directories and files

• models/ - Contains the trained models for the three models and different compression ratio.
• src/ - Code used to optimize, train, evaluate and the test models before board deployment.
• STM32Cube_projects/ - STM32CubeIDE project files obtained from STM32CubeMX to deploy the models on the two mentioned boards. There is also a .npy file containing representative data from the test set to be used for profiling board execution.

Results

🎯 All the results obtained from the tests are explained in the following paper: Ultra-Efficient Compressed Phonocardiogram Classification on a Custom Embedded Neural Accelerator.

Citation

@ARTICLE{11263842,
  author={Ragusa, Domenico and Baeyens, Rens and Pau, Danilo and Marenzi, Elisa and Steckel, Jan and Daems, Walter and Leporati, Francesco and Torti, Emanuele},
  journal={IEEE Internet of Things Journal}, 
  title={Ultraefficient Compressed Phonocardiogram Classification on a Custom Embedded Neural Accelerator}, 
  year={2026},
  volume={13},
  number={4},
  pages={6485-6494},
  keywords={Phonocardiography;Accuracy;Recording;Artificial intelligence;Real-time systems;Feature extraction;Heart;Monitoring;Power demand;Long short term memory;Cardiovascular diseases (CVDs);compressive sensing (CS);edge artificial intelligence (AI);neural accelerator;phonocardiogram (PCG)},
  doi={10.1109/JIOT.2025.3635785}
}