Transcript Presentation

ECG data classification
with deep learning tools
Zhangyuan Wang
Motivation
• ECG data classification to assist health monitoring.
• E.g. in emergency room
• Challenge for current algorithm
• High false alarm rate
• Cannot tackle noisy data
Dataset
• MIT-BIH Arrhythmia Database
• 44 patients in total
• 30 mins of ECG data sampled at 360Hz for each patient
Dataset
• Input:
• Extract 200 points around the peak of each beat
• Label for beat
• following AAMI to 5 labels: N, S, V, F, Q
Dataset
• Acquire data
• WFDB App Toolbox Matlab version
• Store2hdf5 from caffe/matlab
• Preprocessing: median filter…
Method
• Run CNN on raw data
• Caffe
• Windows 10, GTX 765M
• CUDA 7.5
• Visual Studio 2013
Method
• CNN structure
•
•
•
•
Adopted from Mnist_demo_LeNet.prototxt
2*(conv+pooling+ReLu)+ip+ip+softmax
base_lr: 0.01momentum: 0.9
lr_policy: "inv“ gamma: 0.0001 power: 0.75
Method
• Train: augment data
• Use full training set vs part of training set
• 8/10 of the N type
• Add noise to abnormal type
• Test: report within class accuracy
•
•
•
•
•
Python wrapper
Native C code
Matlab wrapper
HDF5Output layer
Modify Caffe code
Modify caffe code
Result
• Overall accuracy of 92%
• Baseline 88%
Contribution
• Setup caffe on windows
• Modify code to output probability of each sample
• Prove the effectiveness of CNN
To Do
• Tune the network