SleepStageML

SleepStageML is an AI-powered software device designed to automatically score sleep stages from polysomnography recordings by analyzing EEG signals using deep learning algorithms. Its efficacy is supported by clinical validation demonstrating non-inferior performance to expert consensus and a predicate device, with a predetermined plan for future algorithm enhancements.

SleepStageML is an artificial intelligence and machine learning (AI/ML) enabled software-only medical device designed to assist qualified clinicians in evaluating sleep quality for patients aged 18 and older. The device analyzes level 1 polysomnography (PSG) recordings, specifically utilizing only the electroencephalography (EEG) signals, to automatically score sleep stages: Wake (W), non-REM 1 (N1), non-REM 2 (N2), non-REM 3 (N3), and REM (R). Clinicians provide PSG recordings in European Data Format (EDF), and SleepStageML processes these using a deep learning algorithm based on convolutional neural networks, returning an EDF+ file containing the original PSG data along with the software-generated sleep stage annotations. These outputs are intended to be comparable to American Association of Sleep Medicine (AASM) guidelines and are always subject to review by a qualified clinician. The clinical performance of SleepStageML was established through a retrospective validation study involving 100 diverse patient PSG recordings. These recordings were independently scored by three registered PSG technologists (RPSGTs), with a 2/3 majority consensus forming the ground truth. SleepStageML's automated scores were compared against this consensus, demonstrating non-inferiority to a predicate device (Sleep Profiler K153412) in terms of positive, negative, and overall agreement for each sleep stage, achieving an overall multi-stage agreement of 84.02%. The device also includes a Predetermined Change Control Plan (PCCP) that outlines specific modifications allowed for the machine learning model, signal preprocessing, probability postprocessing, and signal quality checks, aiming to improve sleep staging performance while ensuring continued safety and effectiveness through rigorous re-validation.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does SleepStageML's performance vary across different EEG montages or in the presence of common EEG artifacts not explicitly filtered by the signal quality check?; Given that SleepStageML only uses EEG signals, what is the potential impact on staging accuracy for sleep stages where EOG and EMG are traditionally considered important discriminators (e.g., REM sleep, differentiating N1 from wakefulness)?; The clinical validation used a 2/3 RPSGT consensus. What was the inter-rater reliability among the three RPSGTs, and how does SleepStageML's performance compare to individual expert scorers, not just the consensus?; Regarding the Predetermined Change Control Plan (PCCP), what specific neural network architecture limitations (beyond 'model size and type') are in place for the 'Update of Machine Learning Model' to ensure changes remain within a defined scope?; How does the device handle PSG recordings with significant signal loss or poor quality in some, but not all, required EEG channels, and does the 'signal quality check' provide granular feedback on which channels are problematic?