Hybrid-Pattern Recognition Modeling with Arrhythmia Simulation for Ubiquitous Health Management System (Preprint)

Abstract

Ubiquitous health management (UHM) services are getting used to working with intelligent medical technology. People expect to track their daily electrocardiogram (ECG) signals at home with AI-enabled computing for assessing arrhythmias, such as atrial premature atrial complex (APC), atrial fibrillation (AFib), ventricular premature complex (VPC), and ventricular tachycardia (VT), versus normal sinus rhythm (NSR).

We established a uHealth management system prototype “ECG4UHM” with coupled machine learning models to recognize hybrid arrhythmia ECG patterns via online AI-computation.

We coupled analytical modeling with machine-learning methods, such as multiple layer perceptron (MLP), random forest (RF), support vector machine (SVM), and naïve Bayes (NB), to recognize the hybrid patterns of four arrhythmia symptoms. The data pre-processing used Hilbert-Huang transform (HHT) with empirical mode decomposition to calculate ECG’s intrinsic mode functions (IMFs). The area centroids of the marginal Hilbert spectrum for the first three-order IMFs were suggested as the features. A strategy of simulative dataset modeling was conducted in the few-data resource analysis for comparison. The modeling was compiled to the modules “MMLCA” and “ECGHHT” for implementation. We engaged the MATLAB compiler and runtime server in the web-based ECG4UHM to build the recognition modules in the Java class library “HHTFeature” and drive the AI-computation to identify the arrhythmia symptoms, respectively.

The crucial datasets were extracted from the MIT-BIH arrhythmia open database in the modeling, which also derived the simulative datasets based on the features’ means and standard deviations. Four arrhythmia symptoms were studied, including the premature pattern (APC, VPC) and the fibril-rapid pattern (AFib, VT) against NSR. With 5-fold cross-validation, the RF model performed the best accuracy of 0.99 for the both patterns due to the crucial dataset. The test dataset evaluated the model and reached the area under the curve (AUC) of receiver operating characteristic about 0.99. The models for all hybrid patterns, excluding VPC versus AFib and VT, achieved an average accuracy of around 90%. With the prediction test, the respective AUCs of the NSR & APC versus the AFib, VPC, & VT were 0.94 and 0.93 for the RF and SVM on average. The average accuracy and the AUC of MLP, RF, and SVM models for APC-VT reached the value of 0.98 due to both crucial and simulative datasets. The feasible models for the hybrid patterns were installed in ECG4UHM to recognize four arrhythmia symptoms online for the UHM.

The self-developed system involved the AI-enabled modeling in recognizing hybrid patterns of the multiple arrhythmia symptoms in practice. The modeling can be extended to serve the UHM and home-isolated care in the future.