Bringing Clinical-Grade Sleep Scoring Home: How Wesper Calculates Total Sleep Time

Total Sleep Time (TST) is a key metric in diagnosing and managing sleep disorders such as obstructive sleep apnea (OSA). It reflects the amount of time a person is actually asleep during a sleep study, excluding periods of wakefulness. At Wesper, we use an FDA-cleared, AI-powered sleep scoring platform, to determine TST from the data collected by our wireless biosensors and pulse oximeter.

Wesper uses a third-party software-as-a-medical-device (SaMD) cleared by the FDA (K210034). ¹ It uses deep learning algorithms to automatically analyze physiological signals from sleep studies, including data collected via Type III and Type IV home sleep apnea tests (HSATs), as well as polysomnography (PSG). It assigns 30-second sleep stages in accordance with American Academy of Sleep Medicine (AASM) guidelines and calculates sleep metrics such as:

• Total Sleep Time (TST)
  
  

• Sleep efficiency
  
  

• Sleep onset latency
  
  

• Time in each sleep stage (N1, N2, N3, REM, Wake)
  
  

At Wesper, we integrate this algorithmic tool with our HSAT data to score sleep in real time, enabling the calculation of TST directly from the wireless biosensor outputs.

Our third-party sleep scoring platform calculates TST by summing the duration of all epochs scored as sleep (N1, N2, N3, and REM). The process includes:

1. Preprocessing signals (e.g., respiratory effort, pulse oximetry, heart rate, body position).
   
   

2. Segmenting the data into 30-second epochs.
   
   

3. Classifying each epoch as either wake, N1, N2, N3, or REM using a convolutional neural network trained on thousands of expert-scored sleep studies.
   
   

4. TST = sum of epochs scored as N1, N2, N3, and REM (in minutes or hours).
   
   

This method mirrors manual scoring conducted by RPSGTs but is automated and standardized across all Wesper studies.

Several validation studies have assessed the sleep scoring’s ability to accurately determine TST compared to manual scoring by human experts. ^2-3

Key Validation Results:

• Overall agreement with human scorers:
  
  

  • Third Party TST vs. manual TST: Pearson r = 0.95
    
    

  • Mean Absolute Error (MAE) for TST: ±20.2 minutes
    
    

  • Bias (Bland-Altman): Less than ±5 minutes on average, indicating minimal systematic error.
    
    

• Epoch-by-epoch agreement:
  
  

  • Overall sleep/wake epoch agreement exceeds 85%
    
    

  • Cohen’s Kappa for sleep stage classification: 0.78–0.82, indicating substantial agreement.
    
    

• FDA 510(k) Summary (K202292):
  
  

The sleep scoring SaMD demonstrated non-inferiority to human scoring, with TST showing no statistically significant difference from consensus scorer averages in multicenter testing.

In addition to Total Sleep Time (TST), the FDA-cleared third-party sleep scoring platform also calculates individual sleep stages. Sleep staging is determined using a deep learning system trained on over one million diagnostic PGSs paired with photoplethysmography (PPG) data. This approach allows the model to recognize stage-specific physiological patterns and classify 30-second epochs as Wake, N1, N2, N3, or REM.

• Physiological signals (PPG and oximetry) are preprocessed and segmented into 30-second epochs.
  
  
• Multiple deep neural network models and statistical signal processing techniques are applied.
  
  
• Epochs are automatically classified into sleep stages (Wake, N1/N2, N3, REM) based on PPG-derived features validated against gold-standard PSG scorers.
  
  
• The algorithm outputs both sleep stage labels and derived metrics such as stage duration, sleep efficiency, sleep latency, and wake after sleep onset.

Clinical validation compared algorithmic staging against PSG scored by a consensus panel of registered polysomnographic technologists. Results showed strong agreement across stages:

• Wake: Sensitivity 86.7%, Specificity 93.5%
  
  
• Light NREM (N1/N2): Sensitivity 80.9%, Specificity 85.5%
  
  
• Deep NREM (N3): Sensitivity 63.4%, Specificity 95.9%
  
  
• REM: Sensitivity 83.6%, Specificity 97.5%

Accurate estimation of total sleep time is critical for interpreting indices like the Apnea-Hypopnea Index (AHI) and hypoxic burden, which are normalized by sleep time. If TST is overestimated, OSA severity may be underdiagnosed. The high correlation and low error margin help ensure clinically reliable scoring of TST in home settings, enabling providers to make informed decisions based on Wesper reports.

Wesper ensures that TST and sleep stages are calculated with reliability comparable to manual PSG scoring. This approach brings hospital-grade sleep diagnostics into the home, reducing diagnostic delays and improving patient access to care.

1. EnsoData, Inc. 510(k) Premarket Notification: EnsoSleep (K202292). U.S. Food and Drug Administration. 2020 Nov 20. Available from: https://www.accessdata.fda.gov/cdrh_docs/pdf20/K202292.pdf

2. Patel R, Singh M, Pacheco D, Roy S, Staloch J, Halvorson S, et al. Automated sleep staging and scoring using a deep learning algorithm improves inter-scorer agreement and scoring efficiency. Sleep. 2022 Apr 11;45(4):zsab291. doi:10.1093/sleep/zsab291.