Sleep Apnea Detection on a Wireless Multimodal Wearable Device Without Oxygen Flow Using a Mamba-based Deep Learning Approach

TL;DR

The paper tackles the need for scalable, non-intrusive sleep apnea assessment by deploying a Mamba-based deep learning model on the ANNE One multimodal wearable to predict AHI and characterize respiratory events without oxygen sensors. It demonstrates strong concordance with PSG at the recording level and provides detailed event- and duration-level analyses, supported by novel event-wise evaluation metrics including TAO. Architectural ablations show the necessity of the Context CNN and Distance MLP within the Mamba backbone, and feature studies highlight the complementary value of chest ECG/XYZ and limb PPG data. The work advances wearable-based OSA detection with robust performance across demographic groups and introduces a rigorous framework for evaluating apnea events beyond traditional segment-wise metrics, with implications for scalable, home-based sleep health screening.

Abstract

Objectives: We present and evaluate a Mamba-based deep-learning model for diagnosis and event-level characterization of sleep disordered breathing based on signals from the ANNE One, a non-intrusive dual-module wireless wearable system measuring chest electrocardiography, triaxial accelerometry, chest and finger temperature, and finger phototplethysmography. Methods: We obtained concurrent PSG and wearable sensor recordings from 384 adults attending a tertiary care sleep laboratory. Respiratory events in the PSG were manually annotated in accordance with AASM guidelines. Wearable sensor and PSG recordings were automatically aligned based on the ECG signal, alignment confirmed by visual inspection, and PSG-derived respiratory event labels were used to train and evaluate a deep sequential neural network based on the Mamba architecture. Results: In 57 recordings in our test set (mean age 56, mean AHI 10.8, 43.86\% female) the model-predicted AHI was highly correlated with that derived form the PSG labels (R=0.95, p=8.3e-30, men absolute error 2.83). This performance did not vary with age or sex. At a threshold of AHI$>$5, the model had a sensitivity of 0.96, specificity of 0.87, and kappa of 0.82, and at a threshold of AHI$>$15, the model had a sensitivity of 0.86, specificity of 0.98, and kappa of 0.85. At the level of 30-sec epochs, the model had a sensitivity of 0.93 and specificity of 0.95, with a kappa of 0.68 regarding whether any given epoch contained a respiratory event. Conclusions: Applied to data from the ANNE One, a Mamba-based deep learning model can accurately predict AHI and identify SDB at clinically relevant thresholds, achieves good epoch- and event-level identification of individual respiratory events, and shows promise at physiological characterization of these events including event type (central vs. other) and event duration.

Figures (14)

• Figure 1: The ANNE One's chest module (A), attached via an adhesive; the limb module (B), attached around a finger; and example raw signals with PSG-labelled apneas (C). Images by Sibel Health Anne-One-LimbAnne-One-Chest.
• Figure 2: Overview of the study pipeline.
• Figure 3: Overview of model architecture, with intermediate tensor shapes specified in parentheses. Hyperparameters were determined experimentally (see Section \ref{['model-params']} in the Appendix for a comprehensive print of model parameter values). The apnea logits (bottom-right) are used for training, while the unnormalized apnea distances are used for postprocessing. For binary apnea classification, $D_{\text{out}} = 1$.
• Figure 4: An example application of event-wise metrics.
• Figure 5: Scatterplot (left) and Bland-Altman plot (right) comparing the predicted and ground truth AHI across the test set. For the Bland-Altman plot, the solid line is the bias and the dotted lines in the Bland-Altman plot represent 0 as well as 1.96 standard deviations above and below the mean.