HEARTBEAT CLASSIFICATION FOR WEARABLES

Entry for Hackoff v3.0 Hackathon in the Siemens Healthineers Challenge - ECG Classification of Heart Disorders. The same project is also applicable for the Open track Data Science and AI .

Team: Gray Matter

• Arpit Agarwal
• Tamoghno Bhattacharya
• Varun Bhardwaj
• Smitesh Hadape
• Hardik Bhati

Problem Statement

We were told to propose and implement an intelligent real time heartbeat classification algorithm based on ECG graph data.
The goal of the study was to design a model that is able to classify cardiac arrhythmia (17 diagnostic classes encompassing “normal sinus rhythm”, “pacemaker rhythm” and 15 other rhythm disorders) effectively from analysis of long-duration (10-s) ECG signal fragments.

We implemented a modified version of the paper "Arrhythmia detection using deep convolutional neural network with long duration ECG signals".

Dataset

The dataset was acquired from here. ECG signals were obtained from the PhysioNet service from the MIT-BIH Arrhythmia database.

• The ECG signals were from 45 patients: 19 female (age: 23-89) and 26 male (age: 32-89).
• The ECG signals contained 17 classes: normal sinus rhythm, pacemaker rhythm, and 15 types of cardiac dysfunctions (for each of which at least 10 signal fragments were collected).
• All ECG signals were recorded at a sampling frequency of 360 [Hz] and a gain of 200 [adu / mV].
• For the analysis, 1000, 10-second (3600 samples) fragments of the ECG signal (not overlapping) were randomly selected.
• Only signals derived from one lead, the MLII, were used.
• Original data is in mat format (Matlab).

Approach

Preprocessing (Normalization)

• Firstly, we wrote a function to convert .mat files to .csv files and combined them to form a dataset which was used for further purpose.
• Data was preprocessed which included rescaling the ECG data values between [-1,1] so as to have a better and a faster classification.

Classification Model

• We tried our hands on a variety of methods to find which one was best suited for the given task and found that 1-D CNN model worked best.
• Since, the number of classes were more (17) and the dataset was comparatively smaller, we had to use only training and testing sets.
• The proposed 1D CNN Classification model consists of 6-7 convolutional layers having dropout after each of them along with ‘relu’ activation.
• After flattening, the last layer consists of a ‘softmax’ layer.
• Many changes had to be done by brute-force inorder to get a high accuracy model.

Deployment

• Our model can be used in clinical scenarios along with the use in Wearables.
• The patient ECG can be aquired and sent through the mobile phone to the cloud where our developed model is trained and kept.
• Result can be validated by ECG beats and the message regarding the defect/disease will be received on mobile device as well as Wearables.

Result

At the end of 20 epochs for the 17-classes the training and validation stages attained accuracy rates of 98% and 93.6%, respectively. The result could be sometimes abnormal as the dataset is small and many of the classes have just 10-11 fragments, so training over them couldn’t provide apt results. Also, we tried using LSTM, RCNN, but 1-D CNN was the simplest and optimized compared to all other models.

References

https://www.frontiersin.org/articles/10.3389/fphys.2020.569050/full