Dual-path continual learning with cross-scale attention for sparse physiological signal analysis

With rapid global population aging, chronic diseases have become a major public health challenge. Early detection and continuous analysis of physiological signals are essential for timely intervention, particularly in long-term monitoring scenarios. However, pathological events such as atrial fibrillation and epileptic seizures are often highly sparse over time, posing challenges for continuously adaptive models. Three key challenges arise in this setting. First, non-stationary physiological data streams require models to balance plasticity and stability. Second, insufficient modeling of both short-term dynamics and long-range temporal dependencies limits the representation of pathological patterns across multiple temporal scales. Third, continual model updates under sparse pathological supervision can lead to catastrophic forgetting of previously learned abnormal patterns. To address these challenges, we propose a continual learning (CL) framework for pathological event detection in physiological time series under evolving data distributions. The framework incorporates a dual-path memory system consisting of a working memory for rapid adaptation and a long-term memory for stable knowledge consolidation. The framework further employs a Cross-Scale Attention (CSA) mechanism to capture short-term dynamics and long-range temporal dependencies. An anomaly-prioritized replay strategy and a memory-consistency objective are further introduced to preserve rare pathological patterns during continual adaptation. We evaluate the proposed framework on three public datasets — MIT-BIH, CHB-MIT, and Siena — for atrial fibrillation detection and seizure detection. Experimental results show competitive and generally superior performance in accuracy and F1 score across the evaluated datasets, while BWT results demonstrate improved knowledge retention during continual learning. These results demonstrate the effectiveness of the proposed framework for detecting sparse pathological events under evolving physiological data distributions.