Unconstrained Sleep Apnea Detection With Conv-ViT Network: LoRA Tuning for Personalized Monitoring

Sleep apnea is a common sleep-related condition defined by repeated pauses in respiration during sleep, which increases the risk of severe health complications. Polysomnography (PSG) is the diagnostic gold standard; however, its limited accessibility and poor suitability for long-term monitoring necessitate alternative home-based techniques. This study introduces an unconstrained sleep apnea detection system based on a convolutional vision transformer (Conv-ViT) with personalization via low-rank adaptation (LoRA). Data were collected from 121 participants who underwent PSG using a polyvinylidene fluoride (PVDF) sensor placed underneath a mattress topper to capture unconstrained signals. The Conv-ViT model combines CNN for localized features extraction with ViT for global representation learning and was fine-tuned on PVDF signals via transfer learning from PSG-derived airflow data. For personalization, LoRA tuning was applied to reflect the individual physiological variability. The evaluation results showed that Conv-ViT outperformed standalone CNN and ViT models, achieving Cohen’s kappa (KAPPA) of 0.736 and accuracy of 0.903 on a test dataset of 61 participants. Personalization with LoRA improved the performance, increasing the mean participant-level KAPPA from 0.674 to 0.728 and the event-level KAPPA from 0.736 to 0.763. Furthermore, the model effectively classified apnea severity, achieving a KAPPA of 0.73 and an average accuracy of 0.93. These findings demonstrate that an end-to-end Conv-ViT model using unconstrained PVDF signals can reliably detect sleep apnea and assess its severity, offering scalable, long-term, and individualized home-based sleep apnea monitoring.