Scale-Space Derivative Encoding and Dilated Temporal Convolution with Attentive Statistics Pooling for Oversampling-Free Single-Lead ECG Arrhythmia Classification

Abstract

Automatic recognition of abnormal heartbeats from a single electrocardiogram lead is limited by two coupled difficulties: the clinically important beats are the rarest ones, and the classifier must be small enough to run on the ambulatory hardware that records them. This study presents a compact fully convolutional model for the four Association for the Advancement of Medical Instrumentation beat categories. Three elements define the method. Each beat is encoded as a Gaussian scale-space derivative stack, in which smoothed first and second derivatives at two spatial scales expose slope and curvature at both fine and coarse resolution rather than at the single implicit scale of a finite difference. A dilated residual temporal convolutional encoder with exponentially increasing dilation covers the entire beat without recurrence, so inference is fully parallel. Multi-head attentive statistics pooling then summarises the encoder output by weighted means and weighted standard deviations, which preserves the dispersion of the activations that ordinary attention pooling discards. Class imbalance is handled entirely in the representation, by a supervised contrastive regulariser added to an unweighted cross-entropy, so no synthetic beats are generated and no class is reweighted. On the MIT-BIH Arrhythmia Database the model reaches an accuracy of 99.36 percent and a macro-averaged F1 of 0.943 with 64,086 parameters. A patient-disjoint inter-patient evaluation quantifies how much of this performance survives when no beat from a test patient is seen during training. That evaluation also yields a negative result of practical value: prior-based logit adjustment, a standard remedy for long-tailed problems, degrades performance monotonically with its strength under both protocols on this task, so the reported model omits it.