Improving Minority Class Detection in ECG Arrhythmia Classification through Class-Specific Cascade Thresholding

The proper identification of cardiac arrhythmias based on electrocardiogram (ECG) is essential in order to provide prompt clinical responses and follow-ups with the patient. Although machine learning and deep learning models based upon automated classification systems have reached high average accuracy, they still do not identify minority classes of arrhythmia like fusion (F) and supraventricular (S) beats. Such infrequent but clinically important classes tend to be early warning signs of serious cardiac incidents. The current paper suggests a cascade based thresholding approach to the classification based on a class-selective feature, which is the combination of lightweight Light GBM classifier and regular morphological and temporal ECG features. Amplitude peak-to-peak, kurtosis, skewness, and RR-interval features were extracted on MIT-BIH Arrhythmia Database and per-class threshold optimization to grid search and two-stage cascade decision rule was adopted. The proposed framework is sensitive to minority classes unlike global thresholding and is sensitive to dominant classes. The results of the experimental evaluation with 5-fold cross-validation obtained a macro-F1 score of 0.4654, macro ROC-AUC of 0.8836, and macro PR-AUC of 0.5041, which is a 19.4 point higher than baseline thresholding. Class-wise F1 increased to 0.7198 in ventricular (V) beat and 0.0668 in fusion (F) beat with a total accuracy of 94.3. The clinical relevance of amplitude and RR-based temporal signatures was shown by feature importance analysis, whereas the calibration curve and PR curve showed that superior separability and stability of predicted probabilities were achieved. It offers a computationally small and interpretable answer to real-time arrhythmia detection to bridge the divide between algorithmic and clinical decision support.