Design and Implementation of a Machine Learning-Based System for Atrial Fibrillation Diagnosis

This paper proposes a multi-class, highly scalable machine learning algorithm for atrial fibrillation (AF) detection and develops an intelligent diagnostic service plat-form. The AF detection algorithm is divided into three parts: pathological feature classification, ECG spectrum map classification using CNN, and an integrated stacking model. First, the time domain features, time-frequency domain features, and phase space features are extracted. The feature sets are enhanced by linear predictive coefficients, and the feature subsets are input into the random forest classification model. Then the original single-lead ECG data is converted into a single-channel ECG spectrum map by short-time Fourier transform and logarithm normalization. The feature of spectrum map is extracted by CNN. Moreover, the model is integrated based on the stacking learning method, which can reach an accuracy rate of 88.28% and F1-Score of 87.58%. Finally, an intelligent ECG storage and diagnosis platform is designed and developed to integrate the above algorithms.