We developed and validated PANDA, a deep-learning model for pediatric arousal detection in polysomnography. PANDA uses a U-Net-style encoder-decoder architecture trained on 10-channel PSG signals with 17.5 min contextual windows and 2 Hz arousal outputs. The model was trained on 9150 PSGs from 7604 subjects, validated on 2455 PSGs from 2000 subjects, and evaluated on a held-out Boston Children’s Hospital test set of 3804 recordings from 3000 subjects. To reduce reference-label noise, we constructed a rigorously adjudicated 200-PSG platinum set and quantified inter-rater reliability before and after adjudication. Performance was assessed using sample-wise agreement, event-wise arousal-index agreement, and external validation on CHAT and PATS. On routine labels, PANDA achieved mean subject-wise Cohen’s κ = 0.45 compared with κ = 0.26 for CAISR. Agreement increased to κ = 0.87 on platinum labels, consistent with reduced scoring variability after expert adjudication. PANDA also showed closer event-wise agreement with reference arousal index and lower bias than CAISR across internal and external cohorts. External validation yielded κ = 0.46 on CHAT and κ = 0.38 on PATS, improving to κ = 0.57 and κ = 0.50 after cohort-specific calibration. PANDA enables scalable pediatric arousal detection with improved consistency and reduced manual scoring time.
Gupta et al. (Tue,) studied this question.