Combined Signal, Medical, and Transform Feature Set Based Heart Disease Classification Model Using Electrocardiogram Signal via IDCNN-LSTM Architecture: An IoT Scenario

This work proposes a new heart disease classification model utilizing electrocardiogram (ECG) signals under an IoT scenario. The input ECG signal is preprocessed by applying a low pass filter and the denoised signal is normalized by [Formula: see text]-score normalization process. Subsequently, the features namely, signal, medical, and transform features are derived from the preprocessed signal. Here, an improved stock well transform is proposed in the transform feature extraction using Shannon entropy to evaluate the standard deviation of the Gaussian window. Next, the weighted average fusion approach and enhanced feature level fusion are applied to fuse the extracted features. Finally, a hybrid model is proposed that integrates LSTM and improved DCNN, the final classification outcomes are produced based on the average of the intermediate results. The performance of the proposed work is compared over the conventional methods random forest (RF), bidirectional gated recurrent unit (Bi-GRU), and recurrent neural network (RNN) in terms of accuracy, specificity, sensitivity, etc. RF[Formula: see text][Formula: see text][Formula: see text]0.693, Bi-GRU[Formula: see text][Formula: see text][Formula: see text]0.651, RNN[Formula: see text][Formula: see text][Formula: see text]0.687, DCNN[Formula: see text][Formula: see text][Formula: see text]0.708, LSTM[Formula: see text][Formula: see text][Formula: see text]0.642, support vector machine-neural network (SVM-NN)[Formula: see text][Formula: see text][Formula: see text]0.678, and K-nearest neighbor (KNN)[Formula: see text][Formula: see text][Formula: see text]0.631, respectively, are some of the classic schemes that have seen lower F-measure values. In contrast, the improved deep convolutional neural network-long short-term memory (IDCNN-LSTM) scheme’s F-measure is 0.897.