Difference Poincaré Image Feature Based Persistent Atrial Fibrillation Classification Using Short-Term Electrocardiograms

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia with substantial health consequences. Persistent atrial fibrillation (PersAF), an earlier stage of AF, persists over a week and may spontaneously return to a normal rhythm but untreated cases can escalate to chronic AF. Due to growing intricacy, there is a rising demand for automated detection of PersAF. This study introduces a novel approach for PersAF classification using bidirectional long-short-term memory (BiLSTM) that harnesses both time series and image features extracted from electrocardiogram signals. The minimum redundancy maximum relevance algorithm was executed to optimize the feature set, enhancing the model's effectiveness. The model was evaluated via both 5-fold cross-validation and blindfold validation, utilizing three publicly available datasets. Training_set_I achieved 98.87±0.78 % accuracy, 99.14±0.70% sensitivity, and an F1 score of 0.992±0.005, while Training_set_II demonstrated a detection accuracy of 96.35±3.09% with 91.00±8.39% sensitivity and 92.00±6.90% F1 score. The model retained its performance when subjected to independent, unseen data with 94.91% accuracy, 98.57% sensitivity, and 96.10% F1 score. The tremendous performance of the model with a small number of features makes it a promising tool for wearable healthcare devices in the clinical applications of PersAF classification.