Dynamic compression transformer multimodal for robust and efficient detection of REM sleep behavior disorder

Abstract Background REM Sleep Behavior Disorder (RBD) is a parasomnia characterized by the loss of normal muscle atonia during REM sleep and is regarded as an early biomarker of neurodegenerative diseases such as Parkinson’s disease. Accurate and efficient RBD detection remains challenging, particularly for real-time clinical and wearable-oriented applications. Aim A Dynamic Compression Transformer Multimodal (DCTM) framework is introduced for automated REM sleep behavior disorder (RBD) and RSWA detection by explicitly integrating entropy-driven adaptive knowledge distillation and single-channel–aware REM-focused modeling. Unlike prior approaches based on static multimodal fusion, DCTM is designed to dynamically regulate modality contribution and inference complexity through adaptive temperature scaling and entropy-weighted gating mechanisms. Methods Multimodal polysomnography signals are processed through signal-quality–aware gating, modality-specific encoders, and a cross-modal Transformer with Mixture-of-Experts fusion. To improve efficiency and robustness, dynamic compression strategies, including entropy-based token pruning, attention-head pruning, and mixed-precision quantization, are applied during inference. Progressive inference is further used to enable conditional early termination when confident predictions can be obtained from reliable modalities. Results DCTM was developed using three PSG datasets with distinct analytic roles. Final RBD/RSWA-specific performance was computed on the label-eligible CAP RBD evaluation split, where AUC = 0.95, Cohen’s kappa = 0.85, and F1-score = 0.947 were achieved. Under standardized inference settings, latency of 28 ms, 3.02 GFLOPs, 360 MB memory usage, and 0.045 mWh energy consumption were obtained. Detailed preprocessing, gating, progressive inference, and comparative analyses are provided in the supplementary tables and Supplementary Figure S1a–S1g. Conclusion A favorable accuracy-efficiency trade-off and preliminary evidence of SQI-guided behavioral adaptation were observed under retrospective and controlled-degradation settings. These findings support the potential value of signal-quality-aware adaptive inference for automated RSWA/RBD analysis, while prospective multicenter validation and hardware-level implementation studies remain necessary before any clinical or wearable use can be claimed.