Early detection of heart diseases using a low-cost compact ECG sensor

• Published in: Multimedia tools and applications Vol. 80; no. 21-23; pp. 32615 - 32637
• Main Authors: Dixit, Shivam, Kala, Rahul
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2021; Springer Nature B.V
• Subjects: Computer Communication Networks; Computer Science; Data processing; Data Structures and Information Theory; Decision trees; Deep learning; Diagnosis; Expert systems; Heart diseases; Low cost; Machine learning; Multimedia Information Systems; Oversampling; Physicians; Sensors; Signs and symptoms; Special Purpose and Application-Based Systems; Voting; Windows (intervals); 1D convolution neural net; Machine learning; Random forest; Medical expert systems; ECG signal; Oversampling; Heart diseases; Gradient boosting
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-021-11083-9

Heart disease patients are continuously increasing. The patients face the problem of a delayed diagnosis as the subjects do not undergo routine tests and consult a doctor only after severe symptoms. Most medical expert systems are designed to aid the doctors in making wise decisions and only such data sets exist in the literature. We attack the problem of an early-stage diagnosis that can be done at the home by the subject himself on a routine basis, using a low cost and compact ECG sensor. Machine learning tools nowadays have become important for data processing and assistance in various fields including medicine. Attributed to an absence of data, we first developed our ECG dataset by collecting ECG signal data from 300 persons including 53 cardiac patients and 247 healthy persons, using a low-cost and compact ECG sensor. To detect the heart diseases from this data, classical methods (Random forest and Gradient boosting) and state of the art Deep Learning models (1D Convolution Neural Net) were used. A problem with machine learning in the specific context is a severe data imbalance, for which oversampling of minority data was used. Since the sensor is a low cost, noise can get added up. Hence, voting across multiple time windows is done to improve the results. After a healthy comparison between all classification methods with different techniques based on their test accuracy, 1D CNN with oversampling and using voting strategy comes out as the best classifiers with a 93% test accuracy.