CoughNet

Ultra-Low Power Embedded AI for Respiratory Monitoring

Overview

CoughNet is an end-to-end Machine Learning project developed within a project based course I took at ETH Zürich called “Machine Learning on Microcontrollers”. I wanted to achieve continuous, 24/7 monitoring of cough frequency to assess the severity of respiratory diseases and the efficacy of antitussive therapies. By deploying high-performance inference directly on a microcontroller, I addressed the critical requirement for patient privacy and developed a proof-of-concept for ultra-low power healthcare monitoring in a wearable form factor.

Hardware-Software Codesing

The core innovation of CoughNet lies in its hardware-optimized architecture. While traditional audio classification relies on MFCC (Mel-frequency cepstral coefficients) for feature extraction, this process can consume over 2x more power than the actual neural network inference.

To solve this I used the MAX78000 AI microcontroller which features a dedicated hardware CNN accelerator, and designed an optimized 1D-CNN architecture that processes raw audio windows directly.

Target Device Architecture
Key Features:
  • Dual-core: Arm Cortex-M4 + RISC-V Coprocessor​
  • Convolutional Neural Network Accelerator​
  • 512KB Flash​ / 128KB SRAM
Method Comparison.

Dataset and Preprocessing

The model was trained on the COUGHVID crowdsourcing dataset from the EPFL Embedded Systems Laboratory, featuring approximately 30,000 samples. A pre-existing model was used to remove noisy or low-quality samples common in crowd-sourced data. Then, I applied downsampling to 16kHz to optimize memory usage while preserving key spectral information in the signal, segmentation to keep only the parts of the signal were there is a cough sound using standard signal processing, cropping into 1s overlapping windows, and finally augmentation by stretching, shifting and adding different background noises.

Additionally, some standard keyword classes were added to make the model a bit more complex than just a binary classification.

Preprocessing Pipeline

Neural Network Architecture

In my work I explored the trade-off between using a pure 1D-CNN (CoughNet1) and a hybrid 1D and 2D-CNN (CoughNet2) architecture. Both were trained using Quantization Aware Training (QAT) on 20 classes (Cough class + 18 Keywords + Unknown) for 100 epochs. The best model depends on the final application resource constraints. While both models reached similar accuracy after tuning, they differed in inference latency, size, and energy consumption per inference. The pure 1D-CNN was faster, used fewer operations per inference (lower power), but had a larger memory footprint.

Pure 1D-CNN Model
Hybrid 1D + 2D-CNN Model
Performance comparison of pure 1D-CNN and hybrid 1D+2D-CNN model

Model Evaluation & Results

Note: Training was done Locally on CPU and limited to 100 epochs, which keeps a lot of room for improvement. The best-performing model (Pure 1D-CNN with QAT and 20 classes) yielded the following metrics:

  • Accuracy: 88%
  • Precision (Cough): 98%
  • Sensitivity (Cough): 98%
  • F1-score: 98%
  • Size: 430KB
  • Latency: 1.8ms
  • Number of Operations: 8.4M ops

While cough detection results were excellent despite the limited training capabilities, the future work could potentially be on improving cough counting which is more tricky due to successive combined coughs and varying durations. A potential solution I would consider if I continue on this project would be to explore variable-sized temporal windows.

Repo