RMSSC: A Robust Multimodal Framework for Sleep Stage Classification with Noisy Labels and Missing Modalities

Most existing sleep staging algorithms are developed under overly ideal conditions, assuming perfectly accurate labels, fully functioning sensors, and minimal inter-subject variability. Clinical practice, however, often presents a much more challenging scenario. In this study, we propose a novel Robust Multimodal Sleep Stage Classification (RMSSC) framework that explicitly addresses these challenges. RMSSC integrates: (i) a novel dual-head architecture for robust learning under label noise, (ii) a cross-modal transformer with learnable prompts to tolerate missing modalities, and (iii) an auxiliary subject classifier to enforce subject-invariant feature extraction. Evaluated on the ISRUC-S1 and ISRUC-S3 datasets, RMSSC achieves accuracies of 79.5% and 81.7% respectively, outperforming state-of-the-art methods. Extensive ablation studies validate the contribution of each component, demonstrating the framework’s robustness and generalisability in real-world sleep staging.