ECG Classification

The code contains the implementation of a method for the automatic classification of electrocardiograms (ECG) based on the combination of multiple Support Vector Machines (SVMs). The method relies on the time intervals between consequent beats and their morphology for the ECG characterisation. Different descriptors based on wavelets, local binary patterns (LBP), higher order statistics (HOS) and several amplitude values were employed.

For a detailed explanation refer to the paper: http://www.sciencedirect.com/science/article/pii/S1746809418301976

If you use this code for your publications, please cite it as:

@article{MONDEJARGUERRA201941,
author = {Mond{\'{e}}jar-Guerra, V and Novo, J and Rouco, J and Penedo, M G and Ortega, M},
doi = {https://doi.org/10.1016/j.bspc.2018.08.007},
issn = {1746-8094},
journal = {Biomedical Signal Processing and Control},
pages = {41--48},
title = {{Heartbeat classification fusing temporal and morphological information of ECGs via ensemble of classifiers}},
volume = {47},
year = {2019}
}

Requirements

Python implementation is the most updated version of the repository. Matlab implementation is independent. Both implementations are tested under Ubuntu 16.04.

Python

• Numpy
• Scikit learn
• Matplotlib (Optional)

Matlab

Performed using Matlab 2016b 64 bits

• LibSVM

Implementation for TensorFlow is in early stage and will not be maintained by the author.

Steps (How to run)

1. Download the dataset:

   • a) Download via Kaggle:

     The raw signals files (.csv) and annotations files can be downloaded from kaggle.com/mondejar/mitbih-database

   • b) Download via WFDB:

     https://www.physionet.org/faq.shtml#downloading-databases

     Using the comand rsync you can check the datasets availability:

     rsync physionet.org::
     

     The terminal will show all the available datasets:

     physionet      	PhysioNet web site, volume 1 (about 23 GB)
     physionet-small	PhysioNet web site, excluding databases (about 5 GB)
     ...
     ...
     umwdb          	Unconstrained and Metronomic Walking Database (1 MB)
     vfdb           	MIT-BIH Malignant Ventricular Ectopy Database (33 MB)
     

     Then select the desired dataset as:

     rsync -Cavz physionet.org::mitdb /home/mondejar/dataset/ECG/mitdb
     

     rsync -Cavz physionet.org::incartdb /home/mondejar/dataset/ECG/incartdb
     

     Finally to convert the data as plain text files use convert_wfdb_data_2_csv.py. One file with the raw data and one file for annotations ground truth.

     Also check the repo WFDB_utils_and_others for more info about WFDB database conversion and the original site from Physionet_tools.

2. Run:

   Run the file run_train_SVM.py and adapt the desired configuration to call train_SVM.py file. This call method will train the SVM model using the training set and evaluates the model on a different test set.

   Check and adjust the path dirs on train_SVM.py file.

3. Combining multiples classifiers:

   Run the file basic_fusion.py to combine the decisions of previously trained SVM models.

Methodology

The data is splited following the inter-patient scheme proposed by Chazal et al., i.e the training and eval set not contain any patient in common.

This code classifies the signal at beat-level following the class labeling of the AAMI recomendation.

1 Preprocess:

First, the baseline of the signal is substracted. Additionally, some noise removal can be done.

Two median filters are applied for this purpose, of 200-ms and 600-ms. Note that this values depend on the frecuency sampling of the signal.

    from scipy.signal import medfilt
    ...
    
    # median_filter1D
    baseline = medfilt(MLII, 71) 
    baseline = medfilt(baseline, 215) 

The signal resulting from the second filter operation contains the baseline wanderings and can be subtracted from the original signal.

    # Remove Baseline
    for i in range(0, len(MLII)):
        MLII[i] = MLII[i] - baseline[i]

2 Segmentation: Beat Detection

In this work the annotations of the MIT-BIH arrhyhtmia was used in order to detect the R-peak positions. However, in practise they can be detected using the following software (see Software references: Beat Detection).

3 Feature Descriptor

In order to describe the beats for classification purpose, we employ the following features:

1. Morphological: for this features a window of [-90, 90] was centred along the R-peak:

   0. RAW-Signal (180): is the most simplier descriptor. Just employ the amplitude values from the signal delimited by the window.

   1. Wavelets (23): The wavelet transforms have the capability to allow information extraction from both frequency and time domains, which make them suitable for ECG description. The signal is decomposed using wave_decomposition function using family db1 and 3 levels.

       import pywt
       ...
   
       db1 = pywt.Wavelet('db1')
       coeffs = pywt.wavedec(beat, db1, level=3)
       wavel = coeffs[0]

   2. HOS (10): extracted from 3-4th order cumulant, skewness and kurtosis.

       import scipy.stats
       ...
       
       n_intervals = 6
       lag = int(round( (winL + winR )/ n_intervals))
       ...
       # For each beat 
       for i in range(0, n_intervals-1):
           pose = (lag * (i+1))
           interval = beat[(pose -(lag/2) ):(pose + (lag/2))]
           # Skewness  
           hos_b[i] = scipy.stats.skew(interval, 0, True)
   
           # Kurtosis
           hos_b[5+i] = scipy.stats.kurtosis(interval, 0, False, True)

   3. U-LBP 1D (59) 1D version of the popular LBP descriptor. Using the uniform patterns with neighbours = 8

       import numpy as np
       ...
   
       hist_u_lbp = np.zeros(59, dtype=float)
   
       for i in range(neigh/2, len(signal) - neigh/2):
           pattern = np.zeros(neigh)
           ind = 0
           for n in range(-neigh/2,0) + range(1,neigh/2+1):
               if signal[i] > signal[i+n]:
                   pattern[ind] = 1          
               ind += 1
           # Convert pattern to id-int 0-255 (for neigh =8)
           pattern_id = int("".join(str(c) for c in pattern.astype(int)), 2)
   
           # Convert id to uniform LBP id 0-57 (uniform LBP)  58: (non uniform LBP)
           if pattern_id in uniform_pattern_list:
               pattern_uniform_id = int(np.argwhere(uniform_pattern_list == pattern_id))
           else:
               pattern_uniform_id = 58 # Non uniforms patternsuse
   
           hist_u_lbp[pattern_uniform_id] += 1.0

   4. My Descriptor (4): computed from the Euclidean distance of the R-peak and four points extracted from the 4 following intervals:
      • max([0, 40])
      • min([75, 85])
      • min([95, 105])
      • max([150, 180])

       import operator
       ...
   
       R_pos = int((winL + winR) / 2)
   
       R_value = beat[R_pos]
       my_morph = np.zeros((4))
       y_values = np.zeros(4)
       x_values = np.zeros(4)
       # Obtain (max/min) values and index from the intervals
       [x_values[0], y_values[0]] = max(enumerate(beat[0:40]), key=operator.itemgetter(1))
       [x_values[1], y_values[1]] = min(enumerate(beat[75:85]), key=operator.itemgetter(1))
       [x_values[2], y_values[2]] = min(enumerate(beat[95:105]), key=operator.itemgetter(1))
       [x_values[3], y_values[3]] = max(enumerate(beat[150:180]), key=operator.itemgetter(1))
       
       x_values[1] = x_values[1] + 75
       x_values[2] = x_values[2] + 95
       x_values[3] = x_values[3] + 150
       
       # Norm data before compute distance
       x_max = max(x_values)
       y_max = max(np.append(y_values, R_value))
       x_min = min(x_values)
       y_min = min(np.append(y_values, R_value))
       
       R_pos = (R_pos - x_min) / (x_max - x_min)
       R_value = (R_value - y_min) / (y_max - y_min)
                   
       for n in range(0,4):
           x_values[n] = (x_values[n] - x_min) / (x_max - x_min)
           y_values[n] = (y_values[n] - y_min) / (y_max - y_min)
           x_diff = (R_pos - x_values[n]) 
           y_diff = R_value - y_values[n]
           my_morph[n] =  np.linalg.norm([x_diff, y_diff])

2. Interval RR (4): intervals computed from the time between consequent beats. There are the most common feature employed for ECG classification.

   1. pre_RR
   2. post_RR
   3. local_RR
   4. global_RR

3. Normalized RR (4): RR interval normalized by the division with the AVG value from each patient.

   1. pre_RR / AVG(pre_RR)
   2. post_RR / AVG(post_RR)
   3. local_RR / AVG(local_Python (Scikit-learn)
   4. global_RR / AVG(global_RR)

NOTE: Beats having a R–R interval smaller than 150 ms or higher than 2 s most probably involve segmentation errors and are discarded. "Weighted Conditional Random Fields for Supervised Interpatient Heartbeat Classification"*

4 Normalization of the features

Before train the models. All the input data was standardized with z-score, i.e., the values of each dimension are divided by its standard desviation and substracted by its mean.

    import sklearn
    from sklearn.externals import joblib
    from sklearn.preprocessing import StandardScaler
    from sklearn import svm
    ...

    scaler = StandardScaler()
    scaler.fit(tr_features)
    tr_features_scaled = scaler.transform(tr_features)

    # scaled: zero mean unit variance ( z-score )
    eval_features_scaled = scaler.transform(eval_features)

5 Training and Test

In scikit-learn the multiclass SVM support is handled according to a one-vs-one scheme.

Since the MIT-BIH database presents high imbalanced data, several weights equal to the ratio between the two classes of each model were employed to compensate this differences.

The Radial Basis Function (RBF) kernel was employed.

    class_weights = {}
    for c in range(4):
        class_weights.update({c:len(tr_labels) / float(np.count_nonzero(tr_labels == c))})


    svm_model = svm.SVC(C=C_value, kernel='rbf', degree=3, gamma='auto', 
                    coef0=0.0, shrinking=True, probability=use_probability, tol=0.001, 
                    cache_size=200, class_weight=class_weights, verbose=False, 
                    max_iter=-1, decision_function_shape=multi_mode, random_state=None)

    svm_model.fit(tr_features_scaled, tr_labels) 

For evaluating the model, the jk index Mar et. al) were employed as performance measure

    decision_ovo        = svm_model.decision_function(eval_features_scaled)
    predict_ovo, counter    = ovo_voting_exp(decision_ovo, 4)

    perf_measures = compute_AAMI_performance_measures(predict_ovo, labels)

6 Combining Ensemble of SVM

Several basic combination rules can be employed to combine the decision from different SVM model configurations in a single prediction (see basic_fusion.py)

7 Comparison with state-of-the-art on MITBIH database:

Classifier	Acc.	Sens.	jk index
Our Ensemble of SVMs	0.945	0.703	0.773
Zhang et al.	0.883	0.868	0.663
Out Single SVM	0.884	0.696	0.640
Mar et al.	0.899	0.802	0.649
Chazal et al.	0.862	0.832	0.612

About datasets:

https://physionet.org/cgi-bin/atm/ATM

MIT-Arrythmia Database

360HZ

48 Samples of 30 minutes, 2 leads 47 Patients:

• 100 series: 23 samples
• 200 series: 25 samples. Contains uncommon but clinically important arrhythmias

Symbol	Meaning
· or N	Normal beat
L	Left bundle branch block beat
R	Right bundle branch block beat
A	Atrial premature beat
a	Aberrated atrial premature beat
J	Nodal (junctional) premature beat
S	Supraventricular premature beat
V	Premature ventricular contraction
F	Fusion of ventricular and normal beat
[	Start of ventricular flutter/fibrillation
!	Ventricular flutter wave
]	End of ventricular flutter/fibrillation
e	Atrial escape beat
j	Nodal (junctional) escape beat
E	Ventricular escape beat
/	Paced beat
f	Fusion of paced and normal beat
x	Non-conducted P-wave (blocked APB)
Q	Unclassifiable beat

beats and rhythms

	Rhythm annotations appear below the level used for beat annotations
(AB	Atrial bigeminy
(AFIB	Atrial fibrillation
(AFL	Atrial flutter
(B	Ventricular bigeminy
(BII	2° heart block
(IVR	Idioventricular rhythm
(N	Normal sinus rhythm
(NOD	Nodal (A-V junctional) rhythm
(P	Paced rhythm
(PREX	Pre-excitation (WPW)
(SBR	Sinus bradycardia
(SVTA	Supraventricular tachyarrhythmia
(T	Ventricular trigeminy
(VFL	Ventricular flutter
(VT	Ventricular tachycardia

AAMI recomendation for MIT

There are 15 recommended classes for arrhythmia that are classified into 5 superclasses:

SuperClass
N (Normal)	N	L	R
SVEB (Supraventricular ectopic beat)	A	a	J	S	e	j
VEB (Ventricular ectopic beat)	V	E
F (Fusion beat)	F
Q (Unknown beat)	P	/	f	u

Inter-patient train/test split (Chazal et al):

DS_1 Train: 101, 106, 108, 109, 112, 114, 115, 116, 118, 119, 122, 124, 201, 203, 205, 207, 208, 209, 215, 220, 223, 230

Class	N	SVEB	VEB	F	Q
instances	45842	944	3788	414	0

DS_2 Test: = 100, 103, 105, 111, 113, 117, 121, 123, 200, 202, 210, 212, 213, 214, 219, 221, 222, 228, 231, 232, 233, 234

Class	N	SVEB	VEB	F	Q
instances	44743	1837	3447	388	8

INCART Database

https://www.physionet.org/pn3/incartdb/

257HZ

75 records of 30 minutes, 12 leads [-4000, 4000]

Gains varying from 250 to 1100 analog-to-digital converter units per millivolt. Gains for each record are specified in its .hea file.

The reference annotation files contain over 175,000 beat annotations in all.

The original records were collected from patients undergoing tests for coronary artery disease (17 men and 15 women, aged 18-80; mean age: 58). None of the patients had pacemakers; most had ventricular ectopic beats. In selecting records to be included in the database, preference was given to subjects with ECGs consistent with ischemia, coronary artery disease, conduction abnormalities, and arrhythmias;observations of those selected included:

Software references: Beat Detection

1. Pan Tompkins

   third_party/Pan_Tompkins_ECG_v7/pan_tompkin.m

2. ecgpuwave Also gives QRS onset, ofset, T-wave and P-wave NOTE:The beats whose Q and S points were not detected are considered as outliers and automatically rejected from our datasets.

3. ex_ecg_sigprocessing

4. osea

License

The code of this repository is available under GNU GPLv3 license.