Paradigm Shift from Feature-Based Machine Learning to End-to-End Deep Residual Neural Networks for Pediatric Age Classification from 12-Lead ECG

ECG criteria vary significantly by age in early years of childhood. Accurate knowledge of age is vital to select appropriate ECG criteria for the patient. Two machine learning and deep learning models were trained on three different inputs from a dataset of 151, 725 recordings of 12-lead ECGs. The feature-based machine learning model utilizing multi-layer perceptron (MLP) was trained with handcrafted features. The end-to-end residual network (ResNet) model received two inputs: a 10-second raw ECG (RawECG) signal and a 1.2-second average representative beat (RepBeat) signal. The ResNet was trained with four data augmentation techniques using RandAugment. The feature-based MLP achieved an F1 score of 0.72 for pediatric detection (age < 16 years) and an average F1 score of 0.65 for classifying pediatric groups (neonate, infant, child, and adolescent) and adult. The end-to-end ResNet achieved an F1 score of 0.88 for pediatric detection from RawECG input and an F1 score of 0.82 from RepBeat input. The model also achieved an average F1 score of 0.78 for classifying pediatric groups and adult from RawECG input, and an F1 score of 0.74 from RepBeat input. The proposed end-to-end ResNet achieves the highest F1 score and outperforms the feature-based MLP, making it the best choice for pediatric age classification.