Abstract
Sleep apnea in children is a major health problem affecting one to five percent of children in the US. If not treated in a timely manner, it can lead to significant physical and mental health issues. Pediatric sleep apnea has different clinical causes and characteristics than adult apnea, yet it has been studied much less extensively. Consequently, while at-home testing tools and automated detection algorithms are widely available for adults, they are scarce for children. In this study, we address this gap by presenting a machine learning-based model for detecting apnea events from commonly collected sleep signals. Our method outperforms state-of-the-art techniques across two public datasets, as measured by F1-score and AUROC. Furthermore, we demonstrate that using just two signals easily collected at home—ECG and SpO2—can achieve highly competitive results, potentially increasing the accessibility of pediatric sleep apnea testing and improving the timeliness of treatment interventions.
Introduction
Obstructive sleep apnea (OSAHS) affects up to 5% of the pediatric population. The current clinical gold standard for diagnosis is Polysomnography (PSG), which requires an overnight stay in a specialized lab. For children, particularly those with special needs, this environment is often stressful, expensive, and inaccessible due to the extensive array of sensors required (EEG, EOG, ECG, SpO2, etc.).
While at-home testing has become a staple for adult diagnosis, reliable algorithmic tools for children remain limited. This research introduces a multi-modal transformer-based model designed to detect pediatric sleep apnea with state-of-the-art accuracy. Crucially, the study explores whether a minimal subset of sensors—specifically ECG and SpO2, which are easily captured by consumer wearables—can provide diagnostic utility comparable to full clinical PSG setups.
Methodology
The proposed architecture is a customized, encoder-only Transformer designed to handle the complexities of multi-modal physiological data. The pipeline is divided into five primary stages:
- Data Integration: The model combines raw PSG signals with demographic data from Electronic Health Records (EHR).
- Segmentation & Tokenization: A novel segmentor handles three distinct data formats: regular time series (e.g., EEG), irregular time series (e.g., R-R intervals from ECG), and tabular demographics. These are synchronized and re-sampled into uniform tokens.
- Transformer Encoders: The core of the model consists of $N$ stacked encoder modules. These utilize multi-head attention to capture complex, non-linear dependencies across different physiological modalities.
- Classification Head: A two-layer multi-layer perceptron (MLP) serves as the final classifier, predicting the probability of an apnea event within a 30-second epoch.
Datasets
The model was rigorously evaluated using two large-scale public pediatric datasets:
- NCH (Nationwide Children’s Hospital) Sleep Data Bank: A comprehensive dataset featuring nearly 4,000 sleep studies from 3,673 patients ranging from infancy to age 30.
- CHAT (Childhood Adenotonsillectomy Trial): A multi-center trial dataset containing 453 baseline sleep studies for children aged 5 to 9.
Results and Discussion
The transformer-based approach set new benchmarks for pediatric apnea detection, achieving an F1-score of 83.9% and an AUROC of 90.6% on the CHAT dataset. It consistently outperformed traditional deep learning baselines, such as CNNs and LSTMs, likely due to the transformer's superior ability to model long-range dependencies in physiological signals.
The "Power of Two"
A pivotal finding of this study is the efficacy of a minimalist sensor approach. Using only ECG and SpO2 signals, the model achieved an AUROC of 89.4% on CHAT and 88.4% on NCH. This performance is remarkably close to the results obtained using the full PSG suite (AUROC 90.0% and 90.4%, respectively). This suggests that high-fidelity diagnosis is possible without intrusive EEG sensors, which are typically required to detect arousals in clinical scoring.
Robustness and Latent Representation
To simulate real-world at-home conditions, the researchers introduced Gaussian noise into the clinical-grade data. The model demonstrated high resilience, maintaining stable performance even at low signal-to-noise ratios. Furthermore, t-SNE visualizations of the model's latent space revealed clear separation between "Normal" and "Apnea" epochs, confirming the transformer's ability to extract meaningful diagnostic features from raw signals.
Conclusion and Future Work
This study demonstrates that multi-modal transformers can significantly advance automated pediatric sleep apnea detection. By proving that ECG and SpO2 signals alone can provide clinical-grade insights, this work paves the way for accessible, at-home diagnostic tools using simple wearables.
However, limitations remain. While the study simulated noise, the data originated from clinical equipment; future validation must involve data collected directly from consumer-grade sensors. Additionally, improving the interpretability of the transformer's "black box" decisions through attention maps will be vital for clinical adoption. Prospective clinical trials are the next necessary step to translate these algorithmic successes into improved patient outcomes.
Summary updated