Title: SleepVLM: Enhancing Trust in Automated Sleep Staging with Explainable, Rule-Based Vision-Language Models

Although automated sleep staging has reached accuracy levels comparable to human experts, its integration into clinical settings is often impeded by the "black box" nature of these systems, which lack auditable decision-making processes. To address this challenge, we present SleepVLM, a novel vision-language model (VLM) designed to perform sleep staging using multi-channel polysomnography (PSG) waveform images while simultaneously providing clear, clinician-friendly explanations. These rationales are strictly grounded in the scoring criteria established by the American Academy of Sleep Medicine (AASM).

The development of SleepVLM involved two key technical stages: waveform-perceptual pre-training and supervised fine-tuning anchored in medical rules. In terms of performance, the model demonstrated robust capabilities, achieving a Cohen’s kappa coefficient of 0.767 on the held-out MASS-SS1 test set and 0.743 on the external ZUAMHCS cohort. These results place SleepVLM on par with current state-of-the-art methods.

Beyond raw accuracy, the model’s interpretability was rigorously assessed. Two qualified sleep technologists independently evaluated the quality of the model’s reasoning across both datasets. The system received mean scores ranging from 3.75 to 3.96 on a 5-point scale, reflecting high ratings for factual accuracy, the comprehensiveness of evidence provided, and logical coherence. By combining competitive predictive performance with transparent, rule-based justifications, SleepVLM aims to enhance the reliability and auditability of automated sleep staging within clinical workflows. Furthermore, to support ongoing research in interpretable sleep medicine, we have publicly released MASS-EX, a new dataset annotated by experts.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC