WikiJournal Preprints/Utilizing Blood Oxygen Saturation and Heart Rate for AI-Driven Apnea/Hypopnea Episode Detection with Brain.js

Authors: Catur Ari Setianto^1,2, Zamroni Afif^1,2, Badrul Munir^1,2, Neila Raisa^1,2, Fadhillah Randy Widiawan³, Faisal Mohammad Rifqi Aqil³, Indrian Novita³, Melina Suhartono³, Muhammad Nayif Alan Hamada³, Nurlinah Amalia³, Shelby Amrus Ernanda^1,2,a, Nada Yuliandha^1,2

¹Neurology Study Program, Faculty of Medicine, Brawijaya University, Indonesia

²Neurology Department, Saiful Anwar General Hospital, Indonesia

³Faculty of Medicine, Brawijaya University, Indonesia

^aCorresponding Author: shelby.ernanda@gmail.com

Abstract

Apnea and hypopnea episodes, characterized by irregular breathing patterns during sleep, are associated with significant health risks if left undetected. This study presents an AI-driven application for identifying apnea and hypopnea episodes based on blood oxygen saturation (SpO₂) and heart rate data, utilizing the Brain.js library. We developed a neural network model trained on a dataset with labeled episodes and evaluated its diagnostic performance. The application achieved an overall accuracy of 81%, with a sensitivity of 68%, specificity of 94%, Positive Predictive Value (PPV) of 91.9%, and Negative Predictive Value (NPV) of 74.6%. These results underscore the app's high specificity and predictive precision, supporting its potential as a reliable tool for screening apnea/hypopnea events in various settings. Future work is needed to integrate this detection software with appropriate wearables to conduct real time screening.

Keywords: Diagnostic, Obstructive Sleep Apnea, machine learning, brain.js, Saturation

Plain language summary

The introduction of health sensors in wearable devices has enabled real-time health monitoring for many individuals. One such sensor measures blood oxygen saturation (SpO₂) and heart rate, which provide information on the oxygen level in the blood and the heart's ability to pump blood throughout the body. Abnormal changes in blood oxygen saturation and heart rate patterns can indicate potential health issues.

Sleep apnea is a medical condition characterized by interrupted breathing during sleep, often accompanied by loud snoring. Many people dismiss it as a non-serious issue, but sleep apnea can lead to worsened sleep quality and impaired brain function. The hope is that health sensors in wearables will be able to detect signs of sleep apnea, even in individuals who may not exhibit obvious symptoms or recognize its impact on their health.

Introduction

Sleep apnea is a common but underdiagnosed disorder characterized by repeated interruptions in breathing during sleep, which can lead to adverse health outcomes, including cardiovascular disease, hypertension, and reduced quality of life. The two main types of episodes in sleep apnea are apnea, a complete cessation of airflow for 10 seconds or more, and hypopnea, a partial reduction in breathing resulting in decreased oxygen levels . Both conditions are associated with significant physiological stress, particularly in individuals with pre-existing health conditions, making early detection critical for effective management and intervention.

Traditional diagnostic methods for apnea and hypopnea, such as polysomnography (PSG), are costly, require specialized equipment, and are limited to clinical settings, which restricts accessibility for many patients. Consequently, there is a growing demand for affordable, portable, and non-invasive alternatives for detecting these episodes . Advances in wearable technology and machine learning provide a promising approach to achieve such solutions, allowing for continuous monitoring of vital parameters like blood oxygen saturation (SpO₂) and heart rate, both of which are relevant biomarkers for identifying apnea and hypopnea episodes .

Recent development showed that machine learning models can effectively analyze physiological data and identify patterns associated with respiratory irregularities, enabling accurate classification of apnea and hypopnea episodes. This study aims to build upon these advancements by developing a lightweight, deep learning-based application using Brain.js, a JavaScript-based neural network library. The choice of Brain.js allows for seamless integration into web and mobile platforms, making the solution accessible to a broader population and more adaptable to various devices. In this study, we present a novel application that uses SpO₂ and heart rate data to detect apnea and hypopnea episodes.

Methods

Data Collection

To train and evaluate the model, we collected physiological data consisting of blood oxygen saturation (SpO₂) and heart rate from two individuals with diagnosed sleep apnea and healthy controls. The data was obtained retrospectively using Polysomnography Record. The dataset includes time-series measurements from wearable sensors that monitor SpO₂ and heart rate continuously during sleep with intervals  of 3 seconds. The data was labeled to indicate the presence or absence of apnea and hypopnea episodes based on clinical diagnoses. The dataset was divided into two categories: one group consisting of data of apnea/hypopnea episodes, and the other group consisting of data during no episode of respiratory irregularities.

Data Preprocessing

The collected raw data consisted of continuous time-series recordings of SpO₂ and heart rate. The data was preprocessed to be formatted as comma separated values. For time-series analysis, data was segmented into 3 seconds windows, each representing a snapshot of SpO₂ and heart rate levels. Each window was labeled as either containing an apnea/hypopnea episode or not based on the clinical diagnosis. The resulting dataset was split into a training set (80%) and a test set (20%) to evaluate the performance of the model. Total of 200 data points of apnea/hypopnea episode and 200 data points of normal respiratory episodes were used as training data, and 100 were used as test data.

Model Development

The machine learning model used in this study was based on a neural network architecture implemented with Brain.js, a JavaScript-based deep learning library. We chose Brain.js due to its simplicity, flexibility, and compatibility with both web and mobile platforms, making it scalable for future development. The training options are presented in table 1.

Table 1. Training Options setting for Brain.js in this study.

Training Options	Value
activation	sigmoid
layer	9
iterations	20000
errorThresh	0.005
log	false
logPeriod	10
leakyReluAlpha	0.01
learningRate	0.3
momentum	0.1
CallbackPeriod	10
timeout	Infinity
beta1	0.9
beta2	0.999
epsilon	1e-8

Model Evaluation

To evaluate the performance of the trained model, we used the test set, which consisted of data points not seen during the training phase. The evaluation metrics included:

• Accuracy: The proportion of correctly classified instances (both positive and negative).
• Sensitivity (Recall): The proportion of true positives (correctly identified apnea/hypopnea episodes) among all actual positive instances.
• Specificity: The proportion of true negatives (correctly identified non-apnea/hypopnea episodes) among all actual negative instances.
• Positive Predictive Value (PPV): The proportion of true positives among all predicted positive instances.
• Negative Predictive Value (NPV): The proportion of true negatives among all predicted negative instances.

These metrics were used to assess the model's ability to detect apnea/hypopnea episodes and its reliability in distinguishing between abnormal and normal breathing patterns.

Implementation

The application was implemented as a web-based tool, with the neural network running in the browser using JavaScript. This study was intended as a pilot study, thus the data should necessarily be inputted manually. No integration with real time sensors were used in this study.

Results

Model Performance

The performance of the neural network model in detecting apnea and hypopnea episodes was evaluated using a test set comprising 20% of the total dataset. The model achieved an overall accuracy of 81%, indicating that 81% of all instances were correctly classified as either containing or not containing an apnea/hypopnea episode. The sensitivity of the model was found to be 68%, meaning that 68% of actual apnea/hypopnea episodes were correctly identified by the model. The specificity of the model was notably high at 94%, reflecting the model’s strong ability to correctly identify normal breathing episodes (true negatives) and avoid false positives. This suggests that the model is particularly reliable in distinguishing between normal and abnormal breathing patterns. The Positive Predictive Value (PPV) was 91.9%, which indicates that when the model predicts an apnea or hypopnea episode, there is a 91.9% probability that the prediction is correct. The Negative Predictive Value (NPV) was 74.6%, which shows that when the model predicts no apnea/hypopnea episode, it is correct 74.6% of the time. The AUC of this neural network is 0.81. Figure 1 showing the ROC Curve for the diagnostic performance of this neural network. which indicates that the model has good discriminatory power for identifying apnea/hypopnea episodes.

Figure 1. ROC Curve for the diagnostic performance of Brain.js for classification of Apnea/Hypopnea Episodes and Normal Episodes.

Apps User Interface

The app opens to a clean, focused interface designed for straightforward data entry and analysis. At the top of the screen, there’s a text input area where users can paste or manually enter SpO2 and heart rate readings, with each line representing a 3-second measurement window. The design of this section makes it clear what’s expected, as it resembles a simple spreadsheet with headers like “Time (3-sec intervals),” “SpO2 (%),” and “Heart Rate (bpm),” guiding users in organizing their data correctly. Once data is entered, a prominent "Analyze Data" button sits just below, inviting users to initiate the deep learning analysis. This button stands out, ensuring users can easily find it when they’re ready to process the information. When the analysis is triggered, the app runs through the data to detect potential apnea or hypopnea episodes based on the provided SpO2 and heart rate values. As results appear, the screen transitions to reveal how many apnea/hypopnea episodes were detected within the data set, presented in an intuitive, easily readable format. There’s also a graph that visually displays the series of the given SpO2 and Heart Rate data. Figure 2 presents the app user interface.

Figure 2. User interface of the App built over Brain.js.

Discussion

Several studies have focused on the use of artificial intelligence (AI) for the detection of obstructive sleep apnea (OSA). Khandoker et. al., (2009) aimed to identify individual apnea and hypopnea events using wavelet-based features of ECG signals . Hayano et. al., (2010) developed an algorithm for screening OSA based on cyclic variation of heart rate . Almazaydeh et. al., (2012) utilized a neural network system for OSA detection through SpO2 signal features . Hassan, (2015) proposed an algorithm using single-lead ECG for automatic screening of OSA, showing superior accuracy. Additionally, Hassan, (2016) and Hassan et. al., (2017) introduced computer-aided OSA detection systems using different parameters and boosting techniques . Mendonca et. al., (2018) highlighted the need for alternative OSA detection approaches outside of sleep laboratories, utilizing sensors and algorithms for automatic analysis . Thorey et. al., (2019) compared the performance of AI-based sleep event detection to human scorers for OSA diagnosis. Vaquerizo-Villar et. al., (2022) employed a deep-learning model based on convolutional neural networks for classifying apnea-hypopnea events from oximetry signals . Abd-Alrazaq et. al., (2024) conducted a systematic review and meta-analysis on wearable AI systems for sleep apnea detection, emphasizing the importance of identifying OSA type and severity .

The primary objective of this study was to develop and evaluate an AI-based application for detecting apnea and hypopnea episodes using blood oxygen saturation (SpO₂) and heart rate data. The results indicate that the neural network model trained on this data performed well in detecting apnea/hypopnea episodes, with an accuracy of 81%, specificity of 94%, and a Positive Predictive Value (PPV) of 91.9%. These metrics suggest that the model is highly effective at identifying normal breathing patterns and is reliable when it predicts the presence of an apnea/hypopnea episode. These findings showed that AI-based diagnostic methods for obstructive sleep apnea (OSA) are showing strong potential, especially when compared to traditional methods in terms of sensitivity and specificity. Our results are comparable to other previous studies. Although polysomnography (PSG) remains the gold standard, AI models, particularly those leveraging convolutional neural networks (CNNs) and support vector machines (SVMs), have achieved competitive performance metrics that position them as valuable tools in OSA screening. In studies involving CNNs, sensitivity has been particularly high. One study reported sensitivity rates up to 93.8% for identifying subjects with OSA and 88.45% for segment identification, marking significant progress in AI's capacity for precise detection . SVM models have also demonstrated strong results; one recent study highlighted that an SVM model achieved a sensitivity of 93% and a specificity of 80% in detecting severe OSA risk, surpassing traditional logistic regression methods . Home monitoring devices integrated with AI have similarly shown promise; large-scale studies reveal that multi-night home monitoring can yield sensitivities above 95% and specificities around 82.5% for severe OSA detection . Unfortunately, our model’s sensitivity of 68% and Negative Predictive Value (NPV) of 74.6% indicate that while the model can effectively identify many true positives, it still misses a substantial number of apnea/hypopnea events. This suggests that the model may not be sufficiently sensitive to detect all occurrences of apnea and hypopnea, particularly in cases where the changes in SpO₂ and heart rate are subtle or occur over brief periods.

When compared to PSG, AI-driven methods offer a more accessible and efficient alternative. PSG is not only expensive but also time-intensive, and its sensitivity for single-night studies often falls into the low 70s, making it less effective for broad-scale, preliminary screening . In contrast, AI methods are increasingly recognized as effective and affordable screening options, with the potential to reduce both diagnostic costs and patient wait times . While PSG will likely remain essential for complex cases and comprehensive evaluations, AI is rapidly becoming a complementary asset in the early detection of OSA, making screening more accessible and potentially transformative for early intervention and patient outcomes.

Limitation

Despite its strengths, the current model has notable limitations. The most significant issue is the sensitivity of 68%, which indicates that the model misses about one-third of the apnea/hypopnea episodes. This limitation may arise from several factors, including the quality and granularity of the data, the choice of features (SpO₂ and heart rate), and the neural network's ability to generalize from the training data. Apnea and hypopnea episodes can vary in severity, duration, and frequency, and the model may not fully capture these variations with the current feature set. Additionally, the NPV of 74.6% suggests that the model does not always correctly predict the absence of an episode. This issue could be mitigated by refining the model’s decision thresholds or incorporating additional features, such as respiratory rate.

Future Directions

The current study provides a strong foundation for the development of AI-based tools for sleep apnea detection. However, there are several avenues for improving the model’s performance. Future work could focus on:

• Feature Enhancement: Incorporating additional physiological parameters, such as respiratory rate or movement data, could improve the model’s ability to detect subtle changes in breathing patterns and enhance sensitivity. The inclusion of data from other wearable sensors (e.g., accelerometers) could help differentiate between apnea-related events and other sleep disturbances.
• Model Optimization: Exploring more advanced machine learning techniques, such as deep convolutional neural networks (CNNs) or recurrent neural networks (RNNs), may allow the model to better capture temporal patterns in the data, improving sensitivity and reducing false negatives.
• Larger and More Diverse Datasets: A more extensive dataset, including data from diverse populations with varying degrees of sleep apnea severity, would improve the generalizability of the model. Including data from different demographics could also help tailor the app for a broader user base, ensuring that it remains effective across a range of individuals with different health profiles.
• Real-Time Performance Enhancements: Optimizing the app for real-time processing is essential for practical use in clinical and home environments. This may involve refining the model’s computational efficiency to allow for faster classification while maintaining accuracy.
• Clinical Validation: Further clinical validation of the model, especially through larger cohort studies and comparison with established diagnostic methods like PSG, will be critical for establishing the model's reliability and effectiveness in real-world settings. In addition, longitudinal studies could assess whether the model can help identify individuals at risk for developing severe sleep apnea, facilitating earlier intervention and better outcomes.

Conclusion

In conclusion, the AI-powered application developed in this study demonstrates promising potential for detecting apnea and hypopnea episodes based on SpO₂ and heart rate data. While the model achieves high specificity and a strong PPV, its sensitivity and NPV suggest areas for improvement. The application represents an important step toward affordable, accessible, and real-time monitoring of sleep-disordered breathing, with potential implications for improving patient outcomes and reducing the burden on healthcare systems. Continued improvements in the model and its deployment will make it a valuable tool for sleep apnea screening and management in diverse settings.

References

Abd-alrazaq, Alaa, Hania Aslam, Rawan AlSaad, Mohammed Alsahli, Arfan Ahmed, Rafat Damseh, Sarah Aziz, and Javaid Sheikh. 2024. “Detection of Sleep Apnea Using Wearable AI: Systematic Review and Meta-Analysis.” Journal of Medical Internet Research 26:e58187. doi: 10.2196/58187.

Almazaydeh, Laiali, Miad Faezipour, and Khaled Elleithy. 2012. “A Neural Network System for Detection of Obstructive Sleep Apnea Through SpO2 Signal Features.” International Journal of Advanced Computer Science and Applications 3(5). doi: 10.14569/ijacsa.2012.030502.

Al-Ratrout, Serein, and Abdulnasir Hossen. 2022. “Convolution Neural Network for Identification of Obstructive Sleep Apnea.”

Hassan, Ahnaf Rashik. 2015. “Automatic Screening of Obstructive Sleep Apnea from Single-Lead Electrocardiogram.” in 2015 International Conference on Electrical Engineering and Information Communication Technology (ICEEICT). IEEE.

Hassan, Ahnaf Rashik. 2016. “Computer-Aided Obstructive Sleep Apnea Detection Using Normal Inverse Gaussian Parameters and Adaptive Boosting.” Biomedical Signal Processing and Control 29:22–30. doi: 10.1016/j.bspc.2016.05.009.

Hassan, Ahnaf Rashik, and Md. Aynal Haque. 2017. “An Expert System for Automated Identification of Obstructive Sleep Apnea from Single-Lead ECG Using Random under Sampling Boosting.” Neurocomputing 235:122–30. doi: 10.1016/j.neucom.2016.12.062.

Hayano, Junichiro, Eiichi Watanabe, Yuji Saito, Fumihiko Sasaki, Keisaku Fujimoto, Tetsuo Nomiyama, Kiyohiro Kawai, Itsuo Kodama, and Hiroki Sakakibara. 2011. “Screening for Obstructive Sleep Apnea by Cyclic Variation of Heart Rate.” Circulation: Arrhythmia and Electrophysiology 4(1):64–72. doi: 10.1161/circep.110.958009.

Khandoker, A. H., J. Gubbi, and M. Palaniswami. 2009. “Automated Scoring of Obstructive Sleep Apnea and Hypopnea Events Using Short-Term Electrocardiogram Recordings.” IEEE Transactions on Information Technology in Biomedicine 13(6):1057–67. doi: 10.1109/titb.2009.2031639.

Kushida, Clete A., Andrew Cotton-Clay, Laura Fava, V. Easwar, Arthur Kinsolving, and Philippe Kahn. 2023. “Number of Nights to Achieve High Sensitivity/Specificity for Detecting OSA Using a Large US Sample by Home Under-Mattress Devices.” Sleep. doi: 10.1093/sleep/zsad077.0502.

Maniaci, Antonino, Paolo Marco Riela, Giannicola Iannella, Jerome R. Lechien, Ignazio La Mantia, Marco de Vincentiis, Giovanni Cammaroto, Christian Calvo-Henriquez, Milena Di Luca, Carlos Miguel Chiesa Estomba, Alberto Maria Saibene, Isabella Pollicina, Giovanna Stilo, Paola Di Mauro, Angelo Cannavicci, Rodolfo Lugo, Giuseppe Magliulo, Antonio Greco, Annalisa Pace, Giuseppe Meccariello, Salvatore Cocuzza, and Claudio Vicini. 2023. “Machine Learning Identification of Obstructive Sleep Apnea Severity through the Patient Clinical Features: A Retrospective Study.” Reproductive and Developmental Biology. doi: 10.3390/life13030702.

Mendonca, Fabio, Sheikh Shanawaz Mostafa, Antonio G. Ravelo-Garcia, Fernando Morgado-Dias, and Thomas Penzel. 2019. “A Review of Obstructive Sleep Apnea Detection Approaches.” IEEE Journal of Biomedical and Health Informatics 23(2):825–37. doi: 10.1109/jbhi.2018.2823265.

Samadi, Behnam, Shahram Samadi, Mehrshad Samadi, Sepehr Samadi, Mehrdad Samadi, and Mahdi Mohammadi. 2024. “Systematic Review of Detecting Sleep Apnea Using Artificial Intelligence: An Insight to Convolutional Neural Network Method.” Archives of Neuroscience. doi: 10.5812/ans-144058.

Thorey, Valentin, Albert Bou Hernandez, Pierrick J. Arnal, and Emmanuel H. During. 2019. “AI vs Humans for the Diagnosis of Sleep Apnea.” Pp. 1596–1600 in 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE.

Vaquerizo-Villar, Fernando, Daniel Álvarez, Gonzalo C. Gutiérrez-Tobal, C. A. Arroyo-Domingo, F. del Campo, and Roberto Hornero. 2022. “Deep-Learning Model Based on Convolutional Neural Networks to Classify Apnea–Hypopnea Events from the Oximetry Signal.” Pp. 255–64 in Advances in the Diagnosis and Treatment of Sleep Apnea. Springer International Publishing.

Wang, Yao, Siyu Ji, Tianshun Yang, Xiaohong Wang, Huiquan Wang, and Xiaoyun Zhao. 2021. “An Efficient Method to Detect Sleep Hypopnea- Apnea Events Based on EEG Signals.” IEEE Access. doi: 10.1109/ACCESS.2020.3038486.

Additional Information

Acknowledgements

We appreciate members of the neurology departement and sleep clinic at Saiful Anwar Hospital for making the study posibble

Competing interests

All authors declare that there is no conflict of interest.

Authors' contribution

CAS: Supervision, Conceptualization, Writing – Review & Editing

ZA: Supervision, Resources

BM: Validation

NR: Validation

FRW: Investigation, Project Administration, Software

FMRA: Writing – Review & Editing, Project Administration

IN: Project Administration, Funding Acquisition

MS: Project Administration, Funding Acquisition

MNAH: Investigation, Project Administration, Software

NA: Project Administration, Funding Acquisition

SAE: Writing – Original Draft, Visualization, Data Curation, Software, Methodology

NY: Data Curation, Formal Analysis

Ethics statement

Data collection was performed according to the ethical approval that was registered in Saiful Anwar General Hospital Research Ethics Committee with registration number of 420 / 280 / K.3 / 102.7 / 2024.