Evaluation of Automated Total Sleep Time Estimation Using an FDA-Cleared Sleep Scoring Algorithm Integrated with Wesper Lab

Study Objectives: Total Sleep Time (TST) is a critical metric for the diagnosis and severity classification of sleep-disordered breathing, as indices such as the apnea–hypopnea index (AHI) and hypoxic burden are normalized by sleep duration. Accurate estimation of TST is therefore essential, particularly in home sleep apnea testing (HSAT) environments where electroencephalography is not available. This evaluation summarizes the methodology and validation performance of an FDA-cleared, third-party automated sleep scoring software used to derive TST from physiologic signals collected during sleep studies and integrated with HSAT data [1].

Methods: This prospective clinical validation study enrolled adult patients undergoing routine attended in-laboratory diagnostic PSG testing. A total of 158 adults were included in the final analysis, of whom 109 (69.0%) met criteria for obstructive sleep apnea (AHI ≥ 5 events/h) and 49 (31.0%) served as normative controls. The cohort was demographically and clinically representative of adults referred for sleep evaluation, spanning ages 18–89 years (Females: 55.7%, Males: 44.3%). Simultaneous recordings were obtained using FDA-cleared PSG systems as the reference standard and a single-channel photoplethysmography (PPG) device for algorithmic analysis. Sleep stages and respiratory events from PSG were scored by a majority scoring panel consisting of three registered polysomnographic technologists (RPSGTs), with consensus defined as agreement by at least two of three scorers, and were reviewed by a board-certified sleep physician.

Automated sleep staging and TST estimation were performed using an FDA-cleared third-party software. Physiologic signals were preprocessed, segmented into 30-second epochs, and classified as Wake, N1, N2, N3, or REM in accordance with American Academy of Sleep Medicine (AASM) guidelines. Total Sleep Time was calculated as the cumulative duration of all epochs classified as N1, N2, N3, or REM.

Results: Across validation datasets, automated TST demonstrated strong agreement with manual PSG scoring. Pearson correlation between algorithm-derived TST and manually scored TST was r = 0.95. Mean absolute error for TST was approximately ±20.2 minutes, with Bland–Altman analysis demonstrating a mean bias of less than ±5 minutes, indicating minimal systematic error.

Epoch-by-epoch sleep/wake agreement exceeded 85%. Sleep stage classification demonstrated substantial agreement, with Cohen’s kappa values ranging from 0.78 to 0.82. Stage-specific performance showed high specificity across all stages, including REM and deep NREM sleep. In multicenter testing submitted as part of FDA clearance (K210034), the SaMD demonstrated non-inferiority to human scoring, with no statistically significant differences between algorithm-derived and consensus-scored TST.

In addition to TST, the automated scoring software provides sleep stage classification for Wake, N1, N2, N3, and REM. Sleep staging is performed using deep neural networks trained on large datasets of PSG recordings paired with physiologic signals, enabling recognition of stage-specific patterns from PPG-derived features. Validation against consensus PSG scoring demonstrated strong performance across stages, including sensitivity of 86.7% and specificity of 93.5% for Wake, sensitivity of 80.9% and specificity of 85.5% for light NREM sleep, sensitivity of 63.4% and specificity of 95.9% for deep NREM sleep, and sensitivity of 83.6% and specificity of 97.5% for REM sleep.

Conclusions: Automated TST estimation using an FDA-cleared sleep scoring algorithm demonstrates accuracy comparable to manual PSG scoring, with strong agreement, low systematic bias, and substantial epoch-level concordance. When integrated with home sleep apnea testing data, this approach enables reliable calculation of sleep duration and derived indices, supporting clinically meaningful interpretation of sleep studies performed in home-based settings.

References: 

1. EnsoData, Inc. Clinical Validation of EnsoSleep PPG AI Scoring: White Paper. Madison, WI: EnsoData, Inc.; 2024.