The College of Administration and Economics at the University of Baghdad discussed , a master’s thesis in field of Statistics by the student (Baqir Ahmed Hussein) and tagged with (ECG signal classification using a hybrid model based on the Hilbert-Huang transform and convolutional neural network) , Under supervision of (Assist. Prof. Dr.Rabab Abdul-Ridha ).
This study sought to classify ECG signals into five main classes (N, S, V, Q, F) using a hybrid model combining the Hilbert-Huang transform (HHT) and a one-dimensional convolutional neural network (1D-CNN). Its performance was then compared to a single 1D-CNN model based on a raw time signal. The HHT was chosen for its ability to analyze nonlinear and instability signals by decomposing them into intrinsic intrinsic patterns (IMFs) and extracting instantaneous frequency-energy features that represent the dynamic behavior of the signal. The applied aspect of the study relied on building six classification models, including one single-stage model (1D-CNN) and five hybrid models using different features extracted from the Hilbert-Huang transform: First_E, Last_E, Mean_E, Median_E, and Mean_f. The performance of the models was evaluated using a set of statistical performance indicators, including overall accuracy, specific accuracy, sensitivity (recall), F1 score, ambiguity matrices, and ROC-AUC curves, to ensure a comprehensive evaluation under conditions of data class imbalance.
The results showed the superiority of the 1D-CNN model based on the raw signal, achieving the highest classification performance among all studied models. Its test accuracy reached 98.00%, and its F1 score reached approximately 97.89%, clearly outperforming all hybrid models based on Hilbert-Huang transform features. The results also indicated that the best performance among the hybrid models was achieved using the Mean_E + 1D-CNN model, with a test accuracy of 92.63% and an F1 score of 91.69%. This model performed well for the normal (N) and (Q) categories, and moderately for the (V) category. However, the detection of rare pathological (S) and (F) categories remained weak, with sensitivity values remaining low compared to the normal category.
Therefore, the 1D-CNN model can be considered a valuable model for classifying ECG signals, while the Mean_E + 1D-CNN model can be used as a suitable alternative when energy-based physical interpretation is required without significantly impacting performance. This message provides a scientific basis upon which future studies can be built to develop more efficient and accurate hybrid models, and to support smart cardiac diagnostic applications and continuous health monitoring systems.
