1207 Automated Detection of Sleep Opportunity Start and End from Polysomnography Signals

Abstract Introduction In clinical polysomnography (PSG), technicians routinely document lights off and lights on timings to define the sleep opportunity window, which serves as the reference for calculating metrics such as Total Dark Time (TDT) and Sleep Latency (SL). TDT is used for computing Sleep Efficiency when Time in Bed is not available. However, these metadata are often missing or inconsistently recorded, particularly in retrospective datasets or unattended studies. When such marks are unavailable, the boundaries of the sleep opportunity must be inferred retrospectively from physiological signals. This study investigates whether these boundaries can be estimated directly from PSG dynamics using an automated classification framework. Methods PSG recordings from 473 healthy participants and 244 individuals evaluated for suspected sleep-disordered breathing were analyzed. A decision-tree–based classifier originally developed for five-stage sleep scoring was extended to include a sixth label indicating epochs outside the sleep opportunity. Features derived from electroencephalography, electrooculography, and electromyography were computed for each 30-second epoch. The sleep opportunity start was defined as the first point at which N successive non-zero epochs, representing physiologically plausible sleep–wake stages, appeared. The end was determined analogously by scanning the record backward. The threshold N was selected by maximizing agreement with technician-recorded boundaries across cross-validation folds. Results The optimal threshold was obtained when fourteen successive non-zero epochs were required to define the start and end of the sleep opportunity window. Under this condition, the mean differences (root mean squared error) were 0.003 (0.509) hours for TDT and –0.044 (0.257) hours for SL.These findings indicate close correspondence between inferred and technician-defined boundaries. Conclusion The results demonstrate that a simple continuity-based criterion applied to sleep-stage predictions enables accurate estimation of sleep opportunity boundaries directly from PSG signals. This approach provides a practical solution when explicit lights off/on annotations are unavailable and supports reliable computation of sleep timing metrics within automated scoring workflows and large-scale retrospective analyses. Support (if any)