Evaluation of sound-based sleep stage prediction in shared sleeping settings

Sound-based AI models for sleep staging face challenges in shared sleeping environments due to acoustic interference from bed partners. This study aimed to evaluate the performance of a sound-based model in two-person polysomnography (PSG) scenarios, with independently recorded sound data for each participant.

Eighty-eight participants (37 males, 51 females) were recruited, including 74 from mixed-gender pairs and 14 from all-female pairs. Bed partners underwent simultaneous PSG in a shared room, with sound recorded separately for each participant using MEMS microphones placed 1.2 m from the bed, oriented toward the closest participant. Sleep staging was classified into 4-stage (wake, REM, light NREM, deep NREM), 3-stage (wake, REM, NREM), and 2-stage (wake, sleep) categories. Macro F1 scores were used to evaluate the model's performance.

The model achieved mean macro F1 scores of 0.590, 0.665, and 0.741 and Cohen's kappa of 0.470, 0.525, and 0.499 for 4-stage, 3-stage, and 2-stage classifications, respectively, across subjects. Performance for 4-stage classification varied by group composition, with macro F1 score and Cohen's kappa of 0.585 and 0.458 for mixed-gender pairs and 0.616 and 0.529 for all-female pairs. Subgroup analyses revealed higher macro F1 scores in males (0.683) compared to females (0.523) and in individuals with higher BMI (0.674), higher AHI (0.659), and higher sleep efficiency (0.641).

This study demonstrates the ability of a sound-based model to predict sleep stages effectively in shared sleeping environments, overcoming the interference challenges from bed partners. Future research will aim to refine the model for broader demographic applicability.