AI Sleep Analysis Now Explains Its Decisions Like a Doctor

Sleep disorders affect millions worldwide, yet diagnosing them relies on a labor-intensive process where trained technicians visually inspect overnight brainwave recordings. While artificial intelligence has matched human accuracy in automating this task, its adoption in clinics has been stalled by a critical flaw: these AI systems operate as black boxes, offering no explanation for their decisions. Clinicians cannot verify if the AI's reasoning aligns with medical guidelines, undermining trust and safety. A new study introduces SleepVLM, a vision-language model that not only stages sleep with expert-level accuracy but also generates detailed, rule-based explanations for every decision, directly addressing the transparency gap that has hindered clinical use.

SleepVLM achieves this by combining visual analysis of polysomnography (PSG) waveform images with natural language generation, all grounded in the American Academy of Sleep Medicine (AASM) scoring rules. The researchers designed a two-phase training pipeline to teach the model both how to perceive key brainwave features and how to reason about them using clinical standards. In the first phase, waveform-perceptual pre-training, the model learned to interpret PSG images by predicting per-second spectral band powers and amplitudes, sharpening its ability to recognize details like alpha rhythms or slow waves. In the second phase, rule-grounded supervised fine-tuning, the model was trained on a mix of fine and coarse annotations, with AASM rules injected into the system prompt, enabling it to generate sleep stage labels along with applicable rule identifiers and structured rationales. This approach mimics a sleep technologist's workflow, where visual inspection leads to rule application and a written justification.

Demonstrate that SleepVLM performs on par with state-of-the-art s while adding unprecedented explainability. On a held-out test set (MASS-SS1, 53 subjects), SleepVLM achieved a Cohen's kappa score of 0.767, overlapping with the best signal-based baseline LPSGM (0.763) and image-based SleepXViT (0.771). On an external clinical test set (ZUAMHCS, 100 subjects), it maintained robust performance with a kappa of 0.743, ranking second only to LPSGM (0.750) and showing better cross-domain stability than SleepXViT, which dropped to 0.694. Confusion matrices revealed that stages like Wakefulness, N2, and REM sleep were well-classified, with recalls above 0.86 on ZUAMHCS, while N1 remained challenging, reflecting its inherent ambiguity even among human scorers. Beyond accuracy, expert evaluation of the model's reasoning quality scored it above 4.0 out of 5.0 across all dimensions—Factual Accuracy, Evidence Comprehensiveness, and Logical Coherence—on both test sets, indicating that the rationales are clinically plausible and detailed.

This breakthrough has significant for sleep medicine and beyond. By providing auditable, rule-cited explanations, SleepVLM bridges the gap between AI performance and clinical trust, allowing doctors to review and verify automated staging decisions rather than relying on opaque predictions. The model's ability to generate rationales that describe channel-specific observations, cite AASM rules, and use exclusionary logic mirrors the diagnostic language clinicians use, enhancing its utility as a decision-support tool. Moreover, the researchers made the model practical for real-world deployment through 4-bit quantization, which reduced its size by 55% to 3.2 GB and doubled inference speed with minimal performance loss, enabling use on consumer-grade hardware. They also released MASS-EX, a novel dataset with expert annotations, to foster further research in interpretable sleep staging.

However, the study acknowledges several limitations. The reasoning quality was evaluated by a single sleep technologist, lacking multi-rater reliability metrics, and the training data spanned limited centers, necessitating broader validation across diverse populations and recording environments. The model operationalized only 15 adult AASM rules, excluding pediatric and infant criteria, and it sometimes made errors, such as perceptual hallucinations or misapplying rules at ambiguous stage boundaries, particularly for N1. Additionally, rendering PSG signals as fixed-resolution images may lose fine detail, and using only three consecutive epochs restricts temporal context. Future work should address these by incorporating multi-center data, extending rule sets, and exploring hybrid signal-image inputs to improve robustness and applicability.