This service uses machine learning (ML) to analyze polysomnography (PSG) data in children. Computed statistical, non-linear, and spectral features are used for sleep staging, arousal, and respiratory scoring. Retraining is flexible and models may be configured to use some or all available features, enabling clinicians to swap between higher performance models where all data are available, or limited-channel models where data are missing or poor quality.

Peer-reviewed article containing out-of-sample performance data coming soon.

Sleep Staging

Sleep staging uses any combination of uploaded EEG, EOG, and EMG data. Thirty-second segments (epochs) of the recording are classified as Stage Wake, N1, N2, N3, or REM. Log probability smoothing is then used to stabilize scored stages and transitions. Scoring data are used to compute sleep statistics including sleep duration, latency, efficiency, and WASO.

Functional Brain Age (FBA) uses uploaded EEG data. Predictions are made during periods of contiguous segments of Stage N2 to produce a point estimate of FBA and estimates of uncertainty.

Arousal Scoring

Arousal scoring uses any combination of uploaded EEG, EOG, and EMG. Following respiratory scoring, arousals are categorized as respiratory if associated with a respiratory event, or as spontaneous otherwise.

Respiratory Scoring

Respiratory scoring uses any combination of uploaded respiratory effort band, nasal and oronasal flow data. Trained model outputs are refined using pulse oximetry data. Scoring data are used to compute respiratory statistics including the apnea-hypopnea index (AHI), central AHI, and obstructive AHI.

Limb Movement Scoring

Limb movement and periodic limb movement scoring uses a specialized algorithm if left and/or right anterior tibialis EMG are provided. This algorithm applies the most current American Academy of Sleep Medicine scoring rules, and does not use machine learning.

Online Retraining

This service offers a streamlined pipeline for training and validating models using a bank of pre-computed features and datasets. Newly trained models are immediately available for use after saving, enabling rapid prototyping and rollout of updated models. All models produce a set of calibrated probabilities for each stage of sleep.

Model training uses a set of custom solvers. Gradient Boosting Machines are second-order Gradient Boosting Decision Trees trained using a histogram-based solver. Linear Regressors use Quadratic Programming. Multinomial Regressors and Neural Networks use L-BFGS-B. Support Vector Machines are trained using L-BFGS-B for multiclass models, while solver options for Support Vector Clustering include Sequential Minimal Optimization and Quadratic Programming. Solver performance is comparable to commonly used machine learning libraries.