Data Analysis of Electrocardiogram-based Heart Failure Detection Using Machine Learning

The terminal stage of cardiovascular diseases is heart failure, which needs to be diagnosed at the earliest possible stage to have a better prognosis on the patient. Electrocardiograms are rich in cardiac functional information but manual interpretation is usually very subjective and highly inefficient. The proposed study utilized an ensemble learning model alongside convolutional neural networks to process the 12-lead electrocardiogram records and develop automatic heart failure detection models. The sample consisted of 21, 837 samples of electrocardiogram of mixed cohort of patients such that 2 655 were heart failure samples. The study introduced an end-to-end learning of raw electrocardiogram waveforms along with handcrafted features into a dual channel hybrid popular with a single architecture. The models performances of hybrid were shown to be at 91.8% accuracy, 89.4% sensitivity, 93.2% specificity, and 90.6% F1-score, which was 3-7 percentage points higher than the single one. Experiments on ablation proved that synergistic interaction between the features of waveforms and statistical features greatly increased the discriminative ability of the model. This study confirms that hybrid architectures are able to utilize all the multi-dimensional information contained by the electrocardiograms to the maximum and yield precise identification of heart failure which is highly accurate at assisting clinical diagnosis as an auxiliary diagnosis tool. The results form a basis in the development of intelligent cardiac monitoring systems with high levels of generalization in patients of various population.