This repository provides the code for TinyML algorithm deployed on low-end microcontrollers that detects life-threatening ventricular arrhythmias (VA) as part of the 2022 ACM/IEEE ICCAD TinyML Contest. Our approach won 1st place in the Flash Ocupation Track and 3rd place on the overall score. See the full list of submissions here.
First, let's set up environment!
- numpy
- argparse
- torch
- torchvision
- tqdm
Samples of waveforms that record the heart signals over time. Pictured images all have unique labels.
Data should be placed in the home directory with the path ./data/[label,filename] such as ./data/0,S27-SR-1.txt.
For inquires on getting access to data, please visit the Contest Website.
python train.py --tqdm_ # show progress bar
python train.py --ensemble # define multiple threshold points as ensembles
The factor parameter scales the standard deviation of the waveform to obtain the peak detection value.
It is defined by a simple equation peak_detection_value = waveform.std() * 2.0. Points in the waveform that are bigger than this value are recongized as peaks.
The average number of peaks for different labels are shown:
The 'threshold' parameter is a decision boundary that we use to classify whether or not the waveform is VA (positive) or non-VA (negative).
The VT and SR are representative labels that comprise ~84% of the total data. The separation result with threshold = 10 and factor = 2.0 is illustrated:
- Maximum and Minimum values of the waveform is plotted with each of the sample representing a unique subject. The high transparency of the samples mean less number of subjects for that bin and vice-versa. Algorithmically, the first quadrant classifies VA while third quadrant classifies non-VA. The fourth quadrant is prone for misclassification.
In order to maximize the f_beta score by tuning the factor and threshold parameters, we use bayesian search provided by Weights & Biases Sweeps.
We evaluate our peak-based approach against CNN (1D convolution) baselines and find that a simple decision boundary approach can actually provide better performance-efficiency tradeoffs.
