RR-Normalized Attention on Neural ODE for ECG Beat Classification toward Physical AI Healthcare

Physical-AI healthcare requires reliable perception of physiological signals in real-world human-device interactions, where the electrocardiogram (ECG) is a key modality for recognizing cardiac state. Lightweight ECG beat classification is therefore important for continuous arrhythmia monitoring on wearable and edge platforms. Neural Ordinary Differential Equations (NODE) model ECG beat dynamics with few parameters, but their behavior under the clinically realistic inter-patient paradigm, where training and test patients are disjoint, remains insufficiently studied. This paper empirically investigates attention mechanisms on a NODE-based encoder using the MIT-BIH Arrhythmia database, focusing on the interpatient setting and the AAMI classes N, S, and V. We progressively augment a baseline NODE with temporal attention and an RR-normalized attention module that scales aggregated feature responses using the beat-wise RR interval at the early layer. Attention improves macro-average accuracy from 92.67% to 94.73%, while RR normalization further increases accuracy, sensitivity, and specificity to 95.46%, 91.82%, and 95.64%, respectively, with the largest gains for minority S and V beats. These results show that RR-guided attention provides a lightweight approach for improving inter-patient generalization in real-time physical-AI healthcare.