Fig. 1: Overview of ABED modeling approach and sleep disordered breathing detection results.

a Flow chart of ABED using Nasal flow, Naso-oral thermistor, Abdominal and Thoracic belts, SpO2, EEG, EOG, and EMG, with the latter three modalities being passed through MAD 45 to compute arousal and wake probability. Then, a 240 s window of the 7 input modalities is shown for a subject from the MrOS cohort, including annotations and predictions. Annotations are lines starting and ending with the same color, while boxes are predictions from ABED. b ABED architecture: Modalities are mixed into a single feature channel, amplified, and fed through 16 bottleneck blocks with residual connections 79. Two Bi-LSTM layers process temporal features, followed by two 1 × 60 kernel layers to determine 146 object class probabilities and positions. These are then passed to non-maximum suppression to remove overlapping predictions, to only focus on the most certain prediction. c Sleep apnea diagnosis confusion matrix on single-scored cohorts combined (MESA, MROS, WSC, CFS). The values represent the correctly diagnosed subjects as a fraction of the total number of true diagnoses in that category. d Confusion matrices of event predictions on all single-scored cohorts combined (MESA, MROS, WSC, CFS). The values represent the ratio of predicted events as a fraction of the total number of true events. e Pearson correlation plots between true and predicted apnea-hypopnea index (AHI) values. The blue line indicates the linear relationship, and the light blue hue shows the 95% confidence intervals. For correlation per cohort, please refer to Supplementary Table 1. f Latent space made using PHATE 46 zoomed to the area of apneas. A total of 200 PSGs were used to make the latent space, 50 from each of MESA, MrOS, WSC, and CFS. g, h F1-score distributions for ABED and individual experts on the Alliance (g; n = 58) and DREEM (h; n = 55) datasets. Each n represents independent biological measurements without replicas (one F1-score per unique subject). Data are shown as box plots: center line = median; box = 25th–75th percentiles (IQR); whiskers = 1.5 × IQR; and minima and maxima are suppressed. A star (★) above a box indicates a significant difference between ABED and the respective expert (two-sided t-test with Bonferroni correction, family-wise α = 0.05). PSG, polysomnography; EEG, electroencephalography; EOG, electrooculography; EMG, electromyography; MAD, multimodal arousal detector; OA, obstructive apnea; CA, central apnea; HYPO, hypopnea; IRE, irregular breathing event; Ch. mix, channel mixing; Ch. amp, channel amplification; ReLU, rectified linear unit; Bi-LSTM, bidirectional long short-term memory; Batch norm, batch normalization; Conv, convolutional layer; None, mild, moderate, severe, categories defined by AHI < 5, 5–15, 15–30, > 30 events h⁻¹; AHI, apnea–hypopnea index; ICC, intraclass correlation; ABED, Apneic Breathing Event Detector. Source data are provided as a Source Data file.