CausalTCC: causal temporal contrastive learning for automated Alzheimer's disease biomarker discovery with bio-electrical signals

OBJECTIVE: Learning robust representations from scarce labeled bio-electrical time-series data remains a critical challenge in clinical diagnosis. While contrastive learning has shown promise, existing approaches often overlook the intrinsic causal dynamics inherent in physiological signals, leading to over-smoothed representations. This study presents CausalTCC, an end-to-end framework for causal temporal contrastive learning of bio-electrical signals toward Alzheimer's disease (AD) biomarker discovery. APPROACH: CausalTCC is developed through three modules: (1) asymmetrical augmentation: distinct weak and strong augmentation strategies generate diverse views while respecting physiological characteristics; (2) causal temporal contrasting: a Transformer-based autoregressive backbone with causal masking integrates intra-view causal loss to capture intrinsic temporal dependencies; and (3) causal contextual contrasting: a symmetric InfoNCE loss leverages instance-level discrimination to learn domain-invariant causal representations, reducing reliance on labeled examples. MAIN RESULTS: Extensive experiments compared CausalTCC to six state-of-the-art counterparts (e.g., EEGNet and EEG-SSL) on four datasets (HAR, AD_A, AD_FTD, and Brain_Lat): (1) CausalTCC achieves the best average F1-scores of 60.1%, 71.0%, and 76.9% under 1%, 5%, and 10% labeled data, outperforming the second-best method by up to 7.2%; (2) under extreme data scarcity (1% labels), it demonstrates substantial improvements on AD_A (Acc: 70.0%, F1: 67.9%); and (3) the causal self-supervision makes CausalTCC far superior to conventional supervised methods. SIGNIFICANCE: Overall, CausalTCC presents a physiologically grounded and relatively parameter-efficient framework that maintains competitive inference efficiency while balancing model complexity and predictive performance for EEG-based clinical decision support.