Embedded Sleep Apnea Detection (TinyML on STM32)

ezgif-537466593956e125.mp4

A real-time medical device prototype that uses Deep Learning (TinyML) to detect sleep apnea events from raw ECG signals directly on a microcontroller. This project demonstrates the end-to-end pipeline from training a 1D-CNN in Python to deploying optimized C code on bare-metal hardware.

Project Overview

• Goal: Classify 60-second ECG windows as "Apnea" or "Normal" in real-time.

• Hardware: STM32 Nucleo-F446RE (ARM Cortex-M4 @ 180MHz).

• Latency: $<$ 130 ms per inference.

• Model Size: ~13KB parameters < 1KB Stack usage).

• Performance: < 89% Confidence on Apnea events (post-optimization).

System Architecture

• Input: PhysioNet ECG Data (Processed in Python).

• Model: Trained .h5 Model converted via STM32Cube.AI.

• Firmware: Generated C Code running on STM32 Firmware.

• Output: Real-Time Inference triggering an LED Indicator.

Technical Approach

1. The Model (Python & TensorFlow)

I designed a lightweight 1D Convolutional Neural Network (CNN) optimized for embedded deployment.

• Input: 6000 samples (60 seconds @ 100Hz).

• Layers: 3x Conv1D blocks with MaxPooling, followed by GlobalAveragePooling and a Dense Sigmoid output.

• Dataset: PhysioNet Apnea-ECG Database (v1.0.0).

2. The Challenge: "The Shy Model"

Problem: Initial training yielded 99% accuracy on paper but failed in hardware testing. The model was "shy," predicting a low probability 27% for both Apnea and Normal inputs.

Root Cause: Severe class imbalance (the dataset had far more healthy minutes than apnea minutes). The model minimized error by safely guessing "Normal" constantly.

Solution:

1. Class Weighting: Implemented sklearn.class_weight during training to heavily penalize missed Apnea events.

2. Smart Scanning: Developed a custom "Scanner Script" (apnea_helper.py) to mine the dataset for high-confidence "Gold Standard" windows for Hardware-in-the-Loop validation.

3. Embedded Deployment (STM32)

• Used STM32Cube.AI (X-CUBE-AI) to quantize and convert the Keras model into optimized C code.

• Integrated the model into a bare-metal C application (main.c).

• Hardware Logic:

  • Input: Serialized ECG windows injected via firmware (HIL Testing).

  • Output: If Probability $&gt;$ 0.5, trigger LD2 (Green LED).

Repository Structure

.
|-- Core/
|   |-- Src/
|   |   |-- main.c           # Main application loop (Inference logic)
|   |   `-- apnea_ai.c       # Wrapper for X-CUBE-AI functions
|   `-- Inc/
|       `-- one_window.h     # Header file containing Test Data
|-- Python_Training/
|   |-- train_apnea_small.py # Training script with Class Weighting
|   |-- apnea_helper.py      # Scanner script
|   `-- requirements.txt     # Python dependencies
|-- SleepApneaDetection.ioc  # STM32CubeMX Configuration
`-- README.md

Results

Performance Metrics (On-Device)

Metric	Value
Inference Time	128.75 ms
CPU Load	~12%
Flash Usage	~177 KB
Stack Usage	912 Bytes

Functional Verification

UART logs demonstrating the discrimination capability of the retrained model:

Injecting APNEA Data... Apnea Prob: 0.8939  [LED ON]
Injecting NORMAL Data... Apnea Prob: 0.1956  [LED OFF]

Author & Links

• Steven Reynoso

• LinkedIn: linkedin.com/in/stevenreynoso123

• GitHub: github.com/StevenReynoso

License: MIT License. Dataset provided by PhysioNet (ODC-BY 1.0).