Deep Learning Transforms Pediatric Sleep Apnea Diagnosis

By Muhammad OsamaReviewed by Susha Cheriyedath, M.Sc.Oct 7 2024

An article recently published in the journal Scientific Reports introduced a novel method for assessing the severity of pediatric sleep apnea. This method uses deep learning (DL) models to analyze blood oxygen saturation (SpO₂) signals. The researchers aimed to offer a less invasive and more child-friendly alternative to the traditional polysomnography (PSG) test, which can be distressing or uncomfortable for children.

Pediatric Sleep Apnea and Diagnostic Challenges

Pediatric Sleep Apnea-Hypopnea (SAH) is a serious sleep disorder in children, characterized by breathing disruptions during sleep. These disruptions can reduce airflow (hypopnea) or complete airflow cessation (apnea). Such issues can negatively impact sleep quality, leading to daytime drowsiness, difficulty concentrating, and long-term health problems, including cognitive, behavioral, and cardiovascular issues.

The conventional diagnostic technique for pediatric SAH is the overnight PSG test, which monitors various physiological signals. Although PSG is effective, it is complex, costly, and uncomfortable for children, highlighting the need for simple diagnostic methods.

Deep Learning for SpO₂ Signal Analysis

In this paper, the authors explored DL techniques to analyze SpO₂ signals. These signals can be easily recorded with pulse oximeters. SpO₂ indicates the oxygen content in blood hemoglobin and is convenient for portable monitoring.

Previous studies have utilized machine learning (ML) methods to analyze airflow (AF) and SpO₂ signals for detecting pediatric obstructive sleep apnea (OSA). However, DL algorithms offer advantages over traditional ML methods. They automatically extract complex features from raw data, which improves diagnostic accuracy and robustness.

The study employed two advanced DL architectures, including residual network (ResNet), and attention-augmented hybrid convolutional neural network bidirectional gated recurrent unit (CNN-BiGRU) model. These models were employed to process SpO₂ signals to estimate the Apnea-Hypopnea Index (AHI) in pediatric subjects.

Methodology and Model Development

The researchers utilized the Childhood Adenotonsillectomy Trial (CHAT) dataset, which included 844 SpO₂ signals from pediatric subjects aged 5 to 9.9. The data was divided into training (60 %), validation (10 %), and testing (30 %) sets. Furthermore, a threefold cross-validation approach was employed to ensure model robustness.

The ResNet-based model was adapted to a one-dimensional (1D) format suitable for analyzing SpO₂ signals. In contrast, the CNN-BiGRU-Attention model combined convolutional neural networks (CNNs) with bidirectional gated recurrent units (BiGRUs) and an attention mechanism to improve feature extraction and temporal processing.

The methodology involved several key stages. First, signal segmentation and preprocessing were performed. Then, SpO₂ signal segments were labeled, followed by model training and optimization, and finally, AHI estimation. The SpO₂ signals were re-sampled to a unified rate of 1 Hz and divided into non-overlapping 20-minute segments.

Motion and zero-level artifacts were addressed, and a 3-second moving average filter was applied to smooth the signals. The labeling algorithm identified desaturation events associated with apneic episodes. The models were trained using the Adam optimizer and Huber loss function.

Key Findings and Insights

The comparative analysis revealed that the CNN-BiGRU-Attention model outperformed the ResNet model, achieving an average accuracy of 75.95 % and a kappa score of 0.63, compared to the ResNet model's accuracy of 72.9 % and kappa score of 0.57. Additionally, the CNN-BiGRU-Attention model demonstrated superior performance in estimating the AHI and categorizing the severity levels of SAH.

Both models were evaluated using standard AHI thresholds of 1, 5, and 10 events per hour during each test fold, confirming their accuracy in detecting pediatric SAH. Various regression metrics, including mean absolute error (MAE), root mean squared error (RMSE), and R-squared (R²), were used to assess the models' proficiency in AHI estimation.

The CNN-BiGRU-Attention model exhibited better agreement between actual and estimated AHI values, with narrower limits of agreement (LoA) and reduced variability compared to the ResNet model. Moreover, confusion matrices and diagnostic metrics such as sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were used to further evaluate the models' diagnostic capabilities.

Potential Applications

This research has significant implications for diagnosing and managing pediatric SAH. The proposed technique is a non-invasive and child-friendly alternative to the conventional PSG test. This could improve accessibility and comfort for children undergoing SAH diagnosis.

Additionally, this method can be integrated into portable monitoring devices, making SAH diagnosis more accessible and less distressing. It may also reduce the costs and complexities associated with traditional diagnostic methods. By accurately estimating the AHI and categorizing SAH severity levels using SpO₂ signals, this approach could lead to more efficient and effective diagnostic processes.

Conclusion and Future Scope

In conclusion, the novel approach demonstrated effectiveness in diagnosing pediatric SAH. By utilizing DL techniques to analyze SpO₂ signals, this study presented a promising alternative to the conventional PSG test. The superior performance of the CNN-BiGRU-Attention model underscores the potential of DL models to enhance diagnostic accuracy and reliability.

Future research could focus on further refining these models and exploring their application in broader clinical settings to improve the diagnosis and management of pediatric SAH. Overall, this approach has the potential to significantly enhance early detection and treatment, ultimately benefiting the health and well-being of children.

Journal Reference

Mortazavi, E., Tarvirdizadeh, B., Alipour, K. et al. Deep learning approaches for assessing pediatric sleep apnea severity through SpO2 signals. Sci Rep 14, 22696 (2024). DOI: 10.1038/s41598-024-67729-9, https://link.springer.com/article/10.1038/s41598-024-67729-9

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.