Automated Sleep Staging Performance Across Two Experimental Sleep Manipulation Protocols in Adolescents

Automated sleep staging algorithms improve the scalability of polysomnography analysis in adolescent research but are primarily validated in undisturbed, single-night sleep recordings. It remains unclear whether these algorithms maintain performance when sleep architecture is experimentally manipulated. This study evaluated the validity of two popular automated sleep staging approaches across two distinct sleep manipulation paradigms in adolescents: a sleep restriction protocol (N=43, mean age [SD]=13.37[1.06] years; four nights including two sleep extension [10hr] and two restriction [4hr] nights) and a slow-wave sleep deprivation (SWSD) protocol (N=18, mean age [SD]=17.17[1.79] years; baseline, SWSD, and recovery nights). Automated sleep staging was performed using the POPulation-level Sleep Stager (POPS), the Yet Another Spindle Algorithm (YASA), and combined POPS+YASA. Manual polysomnography scoring by experienced sleep technicians served as the reference standard. Agreement was assessed using Cohen's κ, and stage-specific performance was evaluated using F1 scores and confusion matrices. Mixed-effects models tested differences in overall and stage-specific agreement across nights. All staging approaches demonstrated moderate-to-substantial agreement with manual scoring across manipulated sleep conditions (κ range=0.53-0.76). In the sleep restriction study, agreement was lower in restricted versus extended sleep (P<0.05), and agreement declined across restriction nights, with poorer Wake, N1, and N2 classification, while N3 remained preserved. In the SWSD study, overall agreement was lower on the SWSD night, with poorer N3 classification while other stages were preserved. Automated sleep staging performance may be less consistent in adolescents when sleep is curtailed or fragmented; future studies should develop and validate these algorithms under multiple sleep conditions.