An Ensemble Approach for Cardiac Arrhythmia Detection using Multimodal Deep Learning

Arrhythmias are irregular and possibly deadly heartbeats. To reduce mortality and morbidity, the patient must receive the proper treatment. Recent techniques in the field of machine learning and signal processing have been applied to the detection and classification of arrhythmias. However, they face several challenges in accurately detecting arrhythmias. One major challenge is the class imbalance in the training data, which can lead to overfitting or underfitting of the models. Another challenge is the variability in ECG signals due to factors such as noise, artifacts, and variations in electrode placement. The proposed objective is to develop an effective ensemble model with a network-in-network architecture based on CNN and LSTM to accurately detect arrhythmias in ECG signals. On the MIT-BH dataset, it was trained and validated to recognize five different kinds of arrhythmias. Prior to that, resampling was done to balance the data in order to prevent the model from being under- or overfit. The ensembled model performs excellent on the validation data. The outcome of the trial show that the suggested model performed remarkably well, with 100% and 99.72% accuracy in the training and testing datasets, respectively. On validation data the CNN and LSTM performed with 98.6% and 98.4% individually. The proposed method outperformed the existing methods in accuracy, proving that the ensemble model is the most effective.