Channel-Imposed Fusion: A Simple yet Effective Method for Medical Time Series Classification

ICLR 2026 Conference Submission824 Authors

Keywords: Medical Time Series Classification, Channel Imposed Fusion (CIF), Data-centric

TL;DR: We propose a data-centric framework for medical time series using Channel-Imposed Fusion (CIF) and HM-BiTCN. CIF encodes physiological priors for noise suppression, and combined with HM-BiTCN, achieves SOTA results beyond Transformers.

Abstract: Medical time series (MedTS) such as EEG and ECG are critical for clinical diagnosis, yet existing deep learning approaches often struggle with two key challenges: the misalignment between domain-specific physiological knowledge and generic architectures, and the inherent low signal-to-noise ratio (SNR) of MedTS. To address these limitations, we shift from a conventional model-centric paradigm toward a data-centric perspective grounded in physiological principles. We propose Channel-Imposed Fusion (CIF), a method that explicitly encodes causal inter-channel relationships by linearly combining signals under domain-informed constraints, thereby enabling interpretable signal enhancement and noise suppression. To further demonstrate the effectiveness of data-centric design, we develop a simple yet powerful model, Hidden-layer Mixed Bidirectional Temporal Convolutional Network (HM-BiTCN), which, when combined with CIF, consistently outperforms Transformer-based approaches on multiple MedTS benchmarks and achieves new state-of-the-art performance on general time series classification datasets. Moreover, CIF is architecture-agnostic and can be seamlessly integrated into mainstream models such as Transformers, enhancing their adaptability to medical scenarios. Our work highlights the necessity of rethinking MedTS classification from a data-centric perspective and establishes a transferable framework for bridging physiological priors with modern deep learning architectures.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 824