CVD Detection by Deep ECG Learning

Presentation Slides

Step 0: Setup

pip install wfdb
pip install PyWavelets

If you haven't installed Keras and Tensorflow, please install them by yourself. Then run

mkdir result data

Step 1: Download Data

Open your python and run

import wfdb
wfdb.dl_database('mitdb', dl_dir='data')

Wait for a few minutes to download the data.

Step 2: Experiment Part I

In Part I, we compare the performance of different data preprocessing methods, including denoise, normalize, and augmentation. The detailed information is in the presentation slide. NLRAV and NSVFQ are two different kinds of labels of the diseases. Also see the slides for detailed information.

python preprocess.py <NLRAV/NSVFQ> <denoise/no> <normalize/no> <augment/no> <random_seed>

For example, if you want to try all the preprocessing methods for NLRAV setting, and use 1229 as the random seed for splitting train/valid/test data,

python preprocess.py NLRAV denoise normalize augment 1229

Then you'll get data_NLRAV.pk under the directory. Then run

python train.py <NLRAV/NSVFQ> <save_file_name>

You'll get save_file_name in /result. To easily run on different random seeds for multiple times, we save the result in the save_file_name, you can use

python check_result.py <save_file_name>

to count the average performance in the save_file_name.

Result:

NLRAV	SE	ACC	AUC	SP	vSE	vACC
default	0.9790	0.9881	0.9865	0.9940	0.9774	0.9882
denoise	0.9772	0.9882	0.9859	0.9947	0.9760	0.9882
normalize	0.9810	0.9885	0.9874	0.9938	0.9807	0.9887
denoise+normalize	0.9823	0.9891	0.9880	0.9937	0.9817	0.9896
augment	0.9806	0.9839	0.9843	0.9879	0.9804	0.9842
denoise+augment	0.9802	0.9842	0.9846	0.9889	0.9796	0.9844
normalize+augment	0.9844	0.9863	0.9871	0.9898	0.9841	0.9869
all	0.9853	0.9864	0.9876	0.9898	0.9846	0.9871

NSVFQ	SE	ACC	AUC	SP	vSE	vACC
default	0.9481	0.9836	0.9707	0.9934	0.9544	0.9840
denoise	0.9464	0.9834	0.9698	0.9932	0.9532	0.9841
normalize	0.9450	0.9843	0.9698	0.9945	0.9535	0.9853
denoise+normalize	0.9457	0.9835	0.9696	0.9936	0.9543	0.9847
augment	0.9591	0.9786	0.9722	0.9852	0.9631	0.9792
denoise+augment	0.9565	0.9777	0.9706	0.9848	0.9637	0.9782
normalize+augment	0.9572	0.9801	0.9723	0.9874	0.9643	0.9806
all	0.9572	0.9805	0.9725	0.9878	0.9639	0.9806

Step 3: Experiment Part II

In this part, we want to compare the performance of different deep learning models. Like previous step, run

python preprocess2.py <NLRAV/NSVFQ> <random_seed>

to get data2_<NLRAV/NSVFQ>.pk. Then run

python train2.py <NLRAV/NSVFQ> <model_type> <save_file_name>

model_type: 1D, 1D-small, 1D-large, LSTM, BiLSTM, Dense

Step 3-1: 2D model

Since the preprocessing of 2D is more complex, we split it from the previous step. Run

python preprocess_2D.py <NLRAV/NSVFQ> <random_seed>

to get data_2D_<NLRAV/NSVFQ>.pk . It takes a few minutes to finish this preprocessing. Then run

python train_2D.py <NLRAV/NSVFQ> <save_file_name>

Result:

NLRAV	SE	ACC	AUC	SP	vSE	vACC
1D	0.9869	0.9938	0.9922	0.9974	0.9900	0.9948
1D-small	0.9852	0.9929	0.9912	0.9972	0.9900	0.9939
1D-large	0.9878	0.9932	0.9922	0.9966	0.9908	0.9940
LSTM	0.9875	0.9936	0.9923	0.9971	0.9911	0.9944
BiLSTM	0.9878	0.9943	0.9926	0.9974	0.9919	0.9953
Dense	0.9828	0.9917	0.9895	0.9962	0.9862	0.9916
2D	0.9875	0.9938	0.9925	0.9975	0.9872	0.9938

NSVFQ	SE	ACC	AUC	SP	vSE	vACC
1D	0.9753	0.9917	0.9853	0.9954	0.9765	0.9916
1D-small	0.9736	0.9914	0.9846	0.9955	0.9738	0.9913
1D-large	0.9726	0.9905	0.9837	0.9949	0.9741	0.9902
LSTM	0.9677	0.9923	0.9824	0.9970	0.9742	0.9930
BiLSTM	0.9748	0.9930	0.9859	0.9970	0.9767	0.9932
Dense	0.9598	0.9908	0.9784	0.9969	0.9600	0.9907
2D	0.9648	0.9905	0.9804	0.9960	0.9683	0.9908