The present study utilised a publicly available, preprocessed derivative of the PSG-Audio dataset, which was distributed through the Kaggle platform under the title “PSG-Audios – Apnea Audios” (Darquea, 2021). The original PSG-Audio dataset, as published by Korompili et al. (2021) and available on the Science Data Bank (ScienceDB), consists of simultaneous polysomnographic and respiratory audio recordings from adult patients undergoing clinical sleep studies. This constitutes a scored, clinically validated reference for research purposes. Nevertheless, it should be noted that full access to the raw recordings is contingent upon the conclusion of individual data-sharing agreements with the authors. The Kaggle-hosted derivative, created by Darquea (2021), provides preprocessed NumPy matrices (.npy format) extracted from the original recordings, thereby enabling the data to be downloaded at no cost for the purposes of algorithmic research, without the necessity of direct PSG equipment.

The derivative dataset comprises 383 files, which are organised within a single root directory (PSG-AUDIO) (Darquea, 2021). The files are divided into two clearly labelled categories: A total of 192 files were found to contain the specific suffix “_ap.npy”, denoting segments recorded during confirmed apnoea events. In contrast, 191 files were identified as “_nap.npy”, signifying normal respiratory patterns. Each.npy file is capable of storing a two-dimensional NumPy array of shape (N_subsignals, 160,000), where N_subsignals denotes the number of temporal segments extracted from a single overnight recording session, and 160,000 corresponds to the number of temporal samples per segment at a sampling frequency of 16 kHz, which is equivalent to 10-second windows. Following the implementation of the quality filter, the total number of individual sub-signals available for analysis was approximately 21,020.

It is imperative to acknowledge the several limitations of this dataset. Firstly, the dataset is moderately sized relative to large-scale clinical repositories. This has the effect of constraining the statistical power of subgroup analyses and may limit generalisation to populations not represented in the original PSG-Audio cohort (Darquea, 2021). Secondly, the demographic characteristics of the underlying participants (age distribution, sex, body mass index, comorbidities, and apnea severity as measured by AHI) are not fully disclosed in the Kaggle-hosted derivative, which prevents formal assessment of selection bias or covariate adjustment. Thirdly, the recordings originate from a single clinical centre under controlled laboratory conditions, which introduces the potential for a mismatch with real-world home recordings affected by ambient noise, variable microphone placement, and different recording hardware. Fourthly, the preprocessing applied by Darquea (2021) prior to public release, incorporating segmentation, potential filtering and feature extraction steps, is only partially documented. This introduces an element of irreducible uncertainty in the reproducibility chain. In the fifth instance, the distribution of classes is approximately balanced at the recording level (192 apnea vs. 191 normal), but the number of sub-signals per class differs due to variable session lengths. This results in a moderate imbalance at the sub-signal level (approximately 61% apnea, 39% normal). This imbalance was addressed through stratified partitioning.

Signal quality validation

Prior to the extraction of features, a systematic quality-control procedure was implemented with the objective of ensuring data integrity and reducing the influence of corrupted segments on model training. For each sub-signal, the presence of non-finite values (NaN and ± infinity) was initially assessed computationally. Sub-signals in which more than 10% of samples were non-finite were excluded from subsequent processing, as replacing a large proportion of values risks introducing an artificial structure. For sub-signals below this threshold, a two-stage sanitisation process was implemented. Isolated NaN or infinity values were replaced with zero when they appeared at signal boundaries or in isolated positions, or with the local temporal mean computed over a symmetric window of 512 samples otherwise. All replacement operations were meticulously documented, and the proportion of imputed samples per sub-signal was recorded to enable post-hoc sensitivity checks. This threshold-based exclusion criterion constitutes a methodological decision that aims to achieve a balance between the quantity of data collected and the quality of the data obtained. The impact of this criterion on class balance was meticulously monitored throughout the study to prevent the introduction of systematic bias.

Signal normalization

Each retained sub-signal underwent independent normalisation through z-score standardisation prior to feature extraction, as outlined in Equation (1).

The location of the focus is determined by the arithmetic mean of the sub-signal (μ), the standard deviation of the sub-signal (σ), and an additional constant, ε = 1 × 10⁻⁸, which is introduced to the denominator to avert division by zero in segments that are nearly silent. Normalisation was applied at the individual sub-signal level rather than globally to account for inter-session amplitude variability attributable to differences in microphone gain, patient-microphone distance, and respiratory effort. Subsequent to the process of z-score normalisation, the amplitude values were clipped to the interval (−10, 10) with a view to attenuating the influence of residual impulse artefacts that had survived the quality-control stage. The selection of this clipping bound was made based on empirical evidence, with the objective of preserving the physiological dynamic range of respiratory sounds whilst simultaneously removing extreme outliers. It was determined that values exceeding ±10 standard deviations were deemed to be physiologically implausible, in view of the signal characteristics exhibited by the dataset.

Respiratory audio recorded during sleep demonstrates significant short-term variability, even within a single polysomnographic recording session. It is evident that variations in airflow turbulence, breathing effort, and transient airway obstruction give rise to fluctuations in amplitude envelope, oscillatory structure, and quiescent intervals over time. This inherent non-stationarity is illustrated by the visualisation of several consecutive sub-signals extracted from the same overnight recording in Figure 1. The example demonstrates the significant variability of respiratory sound characteristics across neighbouring temporal segments. This emphasises the necessity of utilising statistical feature aggregation across frames for machine-learning analysis, as opposed to relying on raw waveform samples.

Spectral and temporal features were extracted from each normalised sub-signal to transform the raw audio waveform into a compact, discriminative representation suitable for a fully connected neural network. All feature extraction was performed using the librosa library (v0.10.x) with a sampling rate of 16,000 Hz, a Fast Fourier Transform (FFT) window size of 2,048 samples, and a hop length of 512 samples, unless otherwise specified. Four complementary feature families were computed:

The concatenation of all statistical aggregates yielded a final feature vector comprising approximately 280 scalar values per sub-signal. Subsequent to concatenation, the features were standardised by means of sklearn’s StandardScaler, which was fitted exclusively on the training subset. Any residual invalid values introduced during the extraction process were replaced with zero prior to fitting. The scaler parameters derived from the training data were then applied without modification to the validation and test subsets. This ensured that no information from the held-out data influenced the normalisation, thus implementing a critical safeguard against data leakage.

The complete set of sub-signals was then partitioned into three non-overlapping, mutually exclusive subsets using stratified random splitting. The purpose of this procedure was to preserve the apnea/non-apnea class ratio within each subset. The stratification process was executed at the sub-signal level. The partition followed an 80/20 primary split (training pool vs. held-out test set), with a further internal 80/20 split of the training pool into training and validation subsets. The resulting approximate proportions were 64% for training, 16% for validation, and 20% for testing (see Table 1).

The held-out test set was isolated prior to any model development activity and was used exclusively for final performance evaluation. It is evident that no hyperparameter selection, threshold adjustment, or architectural decision was informed by test-set results, thereby precluding the presence of optimistic bias in the reported performance estimates. The process of stratification was instrumental in ensuring that the class imbalance that was observed at the sub-signal level was distributed proportionally across all three subsets. This was done with a view to preventing differential class representation from confounding the dynamics of training or the metrics by which it was evaluated.

A potential source of bias inherent to sub-signal-level splitting, as opposed to recording-level or patient-level splitting, is that sub-signals derived from the same overnight session may share correlated acoustic characteristics (e.g. ambient noise, microphone gain, patient-specific breathing patterns). The implementation of strict patient-level data isolation was rendered unfeasible due to the absence of patient identifiers within the Kaggle-hosted derivative. This has been identified as a known limitation, with the potential to result in performance estimates that are more optimistic than those observed when treating a fully independent patient cohort.

Subsequent to the extraction of features and the aggregation of statistics, the resultant feature vectors are to be processed by a classification model that can differentiate between respiratory patterns associated with apnoea events and those associated with normal breathing segments. The overall analytical workflow, which encompasses audio preprocessing, time–frequency feature extraction, statistical aggregation, and neural network classification, is schematically summarised in Figure 2. The architecture under consideration here illustrates the transformation of raw respiratory audio into a compact feature representation, and its subsequent processing by a dense neural network designed for efficient binary apnoea detection.

The proposed classifier is a fully connected (dense) feedforward neural network comprising three hidden layers with 256, 128, and 64 neurons, respectively, followed by a single output neuron. The present architecture was selected deliberately over alternatives such as convolutional or recurrent networks for two reasons: The input to the network consists of pre-aggregated scalar feature vectors rather than raw temporal sequences or two-dimensional spectrograms. This renders spatial or temporal inductive biases unnecessary. Furthermore, a dense architecture imposes lower computational overhead, thus supporting the goal of developing a lightweight model suitable for deployment on portable or low-resource devices.

Each hidden layer applies the Rectified Linear Unit (ReLU) activation function, defined as ReLU(x) = max(0, x), which mitigates the vanishing-gradient problem and promotes sparse activation patterns. Batch Normalisation is applied after each hidden layer before the activation, re-centring and re-scaling layer inputs to accelerate convergence and reduce sensitivity to weight initialisation. Dropout regularisation is applied after each hidden layer with rates of 0.3, 0.3, and 0.2 for layers 1 through 3 respectively; at each training step, the corresponding fraction of neurons is randomly deactivated, discouraging co-adaptation and reducing overfitting to training-set idiosyncrasies. The output neuron uses a sigmoid activation function, σ(z) = 1/ (1 + e^−z), producing a scalar probability p ∈ (0, 1) representing the likelihood that the input sub-signal contains an apnea event. A fixed decision threshold of 0.5 was applied to convert predicted probabilities into binary class labels; this threshold was set a priori and was not optimised on the test set.

The model was trained using the Adam optimiser (Wang et al., 2021) with an initial learning rate of 1 × 10⁻³ and default exponential decay rates β₁ = 0.9, β₂ = 0.999, and ε = 1 × 10⁻⁷. Adam was chosen for its adaptive per-parameter learning rate updates, which generally achieve faster convergence than standard stochastic gradient descent on heterogeneous feature spaces such as the concatenated spectral-temporal vectors used here.

The training objective was binary cross-entropy loss, defined in Equation (2):

Where: N is the batch size, y_i ∈ {0, 1} is the true binary label (0 = no apnea; 1 = apnea), and ŷ_i ∈ (0, 1) is the model’s predicted probability. Binary cross-entropy strongly penalises confident but incorrect predictions, encouraging well-calibrated probability outputs. No class-weighting was applied to the loss function given that the class imbalance at the sub-signal level was moderate (~61/39) and stratified partitioning ensured consistent representation across subsets.

The training process was executed over a maximum of 200 epochs, with a mini-batch size of 32. The early stopping method was employed with a patience of 15 epochs, and the validation loss was monitored at the conclusion of each epoch. Training was terminated when no improvement in validation loss was observed for 15 consecutive epochs, and the model weights corresponding to the epoch with the lowest validation loss were restored. This procedure is designed to prevent overfitting without the need for manual epoch selection. A comprehensive overview of all hyperparameters is provided in Table 2.

The performance of the model was evaluated using the following standard binary classification metrics:

- Accuracy ( Equation 3):

- Precision ( Equation 4):

- Recall ( Equation 5):

- F1-score ( Equation 6):

- Area Under the ROC Curve (AUC-ROC, Equation 7):

Accuracy is defined as the proportion of correct predictions made, while precision is the proportion of predicted positive cases that are actually positive. The recall function evaluates the model’s capacity to accurately identify apnoea events, while the F1-score provides a balanced assessment by combining precision and recall through their harmonic mean. In conclusion, the area under the ROC curve (AUC-ROC) is indicative of the model’s discriminative capacity across all potential decision thresholds.

In the interest of full reproducibility, the complete Python implementation, including data loading, quality control, feature extraction, model training, and evaluation scripts, is publicly accessible at: https://github.com/SebasTS15/Sleep-Apnea-Detection-Model (Torres-Sánchez, 2025). The repository contains a requirements file that details all the library versions that have been utilised. The preprocessed dataset is available at: https://www.kaggle.com/datasets/bryandarquea/psg-audio-apnea-audios (Darquea, 2021). Prior to data splitting and model initialisation, random seeds were fixed at the TensorFlow, NumPy, and Python levels to ensure deterministic behaviour; the seed values utilised are documented in the Kaggle repository. It is evident that the present study did not make use of any questionnaire instruments or patient-administered surveys. Consequently, the relevant questionnaire validation procedures are not applicable.

To evaluate the optimisation behaviour and generalisation stability of the proposed neural network, the evolution of loss and classification accuracy was monitored across all training epochs for both the training and validation partitions. Examining these curves provides insight into convergence dynamics, potential overfitting and the effectiveness of regularisation strategies, such as dropout, batch normalisation and early stopping. The full training history is presented in Figure 3. The model was trained for 175 epochs before early stopping was triggered, indicating that the validation loss had not improved for 15 consecutive epochs beyond epoch 175, at which point the weights from the optimal epoch were restored for evaluation.

A pronounced and monotonic decrease in training loss was observed during the first 50 epochs, reflecting the model’s rapid identification of the dominant spectral-temporal patterns associated with apnoea and normal breathing. Beyond epoch 50, the loss rate declined progressively more slowly, stabilising in the range 0.28–0.32 by the final epochs. Validation loss tracked training loss closely throughout without diverging at any stage of training. The absence of a crossover — where validation loss begins to rise while training loss continues to fall — indicates that the combination of dropout regularisation, batch normalisation and early stopping was effective in preventing overfitting to the training partition.

The model demonstrated a marked improvement in accuracy, with values exceeding 80% in both the training and validation sets during the initial iterations. This improvement was particularly evident, with both curves reaching values consistently above 85% by the midpoint of training. The training and validation accuracy curves demonstrated consistent alignment throughout the study, with no persistent discrepancy exceeding approximately 2 percentage points, further substantiating the absence of overfitting. It is suggested by these convergence characteristics that the feature representation, which comprises aggregated spectral and temporal descriptors extracted from respiratory audio, provides a stable and learnable signal for binary apnoea classification.

While point estimates, such as accuracy, precision, and recall, offer valuable summaries of classification performance, they are contingent upon a fixed decision threshold and consequently capture only a fraction of the model’s diagnostic behaviour. A more comprehensive assessment can be obtained by examining the Receiver Operating Characteristic (ROC) curve, which evaluates classifier performance across all possible discrimination thresholds. The ROC curve obtained for the held-out test set is presented in Figure 4.

The ROC curve deviates significantly from the random-classifier diagonal, attaining a true positive rate exceeding 0.80 while maintaining a false positive rate of approximately 0.10. This outcome signifies the model’s capacity to accurately discern the majority of apnea events, concurrently generating a comparatively low rate of false alarms.

Whilst global metrics are capable of summarising the performance of classifiers in an overall sense, they are not able to reveal the distribution of prediction outcomes between the two diagnostic classes. A confusion matrix is a tool that provides a detailed breakdown of classification results by separating true positives, true negatives, false positives, and false negatives. This representation facilitates a more precise examination of the model’s diagnostic behaviour and potential error patterns. The confusion matrix derived from the test dataset is illustrated in Figure 5.

The precision is moderately lower for the Apnea class (86.82%) than for recall (91.21%), indicating that a proportion of non-apnea sub-signals are incorrectly flagged as apnea events. The 1,788 false positives recorded in the confusion matrix represent segments whose acoustic characteristics, such as intermediate amplitude, irregular oscillatory patterns, or brief signal perturbations, fall close to the decision boundary. In contrast, the 1,135 false negatives are attributable to apnoea sub-signals that the model assigned a predicted probability below 0.5. This is likely due to the acoustic interruption being subtle or partially obscured by noise.

In addition to binary classification outcomes, the distribution of predicted probabilities can offer further insight into model confidence and calibration. By examining the distribution of predicted probabilities for each ground-truth class, it becomes possible to assess whether the model produces well-separated confidence scores or generates uncertain predictions near the decision boundary. As demonstrated in Figure 6, the probability distributions of the model outputs for the held-out test set are presented.

Sub-signals belonging to the Apnea class are strongly concentrated at predicted probability values approaching 1.0, with the modal bin located above 0.90. Sub-signals belonging to the No Apnea class are predominantly clustered at predicted probabilities near 0.0, with the modal bin below 0.10. The two distributions demonstrate a clear separation around the decision threshold of 0.5, with comparatively few sub-signals from either class falling in the interval (0.3, 0.7). The bimodal, well-separated pattern is indicative of effective calibration within the standard threshold, suggesting that the model assigns high-confidence predictions to the majority of sub-signals. This contrasts with the production of uniformly uncertain outputs concentrated near 0.5.

The residual overlap between the two distributions in the intermediate probability range (0.3, 0.7) is responsible for the occurrence of false positives and false negatives. The intermediate-confidence cases are associated with sub-signals, characterised by the absence of sufficiently clear spectral or energy cues in the acoustic signal, which hinders the unambiguous distinction between apnoea and normal breathing.

In order to provide a more complete assessment of the data, representative respiratory sub-signals were examined directly in the waveform domain. A visual inspection of the given examples enables a qualitative comparison to be made between correctly and incorrectly classified signals. This provides insight into the acoustic patterns that may have an influence on the model’s predictions. As demonstrated in Figure 7, illustrative examples are provided in order to elucidate the four potential prediction outcomes: true positive, true negative, false positive and false negative.

Sub-signals that are correctly classified as Apnea (true positives; predicted probability = 0.97 in the representative example) exhibit a characteristic pattern of signal interruption: a region of markedly reduced amplitude occupying a substantial fraction of the 10-second window, flanked by lower-amplitude irregular oscillations, consistent with the cessation of airflow and the associated loss of respiratory sound energy. Sub-signals that have been correctly classified as No Apnea (true negatives; predicted probability = 0.00) display regular, sustained oscillatory activity across the full window duration. These oscillations are stable and relatively high in amplitude, reflecting uninterrupted respiratory airflow.

The two misclassified categories offer a more insightful perspective on the model’s failure modes. The example of a false positive result (where the predicted apnea probability was 0.51 and the true label was ‘No Apnea’) displays a waveform that is superficially similar to an apnea segment, with a transient reduction in amplitude in the middle of the window. This may be indicative of a brief arousal, a swallowing event, or a movement artefact, rather than a true apnea event. The false negative example (predicted apnea probability = 0.11; true label = Apnea) exhibits a waveform with partially preserved amplitude fluctuations throughout the window, suggesting either a hypopnea-level event in which airflow is reduced but not entirely absent, or an apnea event captured at the margin of the 10-second segment boundary. The two misclassification patterns under consideration both involve sub-signals with predicted probabilities that approximate the 0.5 threshold. This confirms the hypothesis that errors are concentrated in the intermediate-confidence region identified above, rather than being distributed uniformly across the score range (see Figure 7).

It is recommended that future research efforts concentrate on enhancing the robustness and clinical applicability of audio-based sleep apnea detection models (Tzeng et al., 2025). A promising avenue for future research involves the expansion of training datasets to encompass recordings obtained under a range of environmental conditions, sensor configurations, and patient populations. Such variability would assist in enhancing the generalisation of the model and reducing sensitivity to noise or recording artefacts. Moreover, the incorporation of multimodal physiological signals, including but not limited to airflow, oxygen saturation, and respiratory effort, has the potential to enhance the accuracy of detection. This integration enables the extraction of complementary physiological information that may not be fully captured by audio signals alone (Wu et al., 2025).

The potential for enhanced performance can also be attributed to methodological advancements. Future research could involve the exploration of alternative neural architectures, such as convolutional or transformer-based models, which have the capacity to learn more complex temporal representations from respiratory signals (Sturekova et al., 2025; Wang & Strodthoff, 2025). Concurrently, research into explainable artificial intelligence techniques has the potential to enhance the interpretability of model predictions, thereby fostering greater clinical trust and facilitating integration into diagnostic workflows. In conclusion, it is imperative to undertake prospective validation using clinically annotated datasets and real-world monitoring scenarios in order to assess the translational potential of automated respiratory sound analysis as a scalable tool for sleep apnoea screening and monitoring.

All data necessary to reproduce the findings reported in this article are either publicly accessible through open repositories or have been deposited in a long-term, citable data archive. It is important to note that no proprietary, restricted, or personally identifiable data were generated or utilised in the course of this study. The raw data supporting the results of this study — comprising the complete set of pre-processed NumPy sub-signal arrays (.npy files) used for training, validation, and evaluation, together with the associated binary class labels, the held-out test-set predicted probability scores, and the per-epoch training and validation loss and accuracy logs — have been deposited in the Mendeley Data Repository and are publicly available at the following persistent identifier:

“Sleep Apnea Detection from Respiratory Sounds: Dataset Documentation”, Mendeley Data, V1, https://doi.org/10.17632/wrf77nvb8z.1 (Tunubala, et al., 2026).

Data are available under the terms of the Creative Commons Attribution 4.0 International license (CC-BY 4.0).

The pre-processed PSG-Audio derivative dataset from which the sub-signals in this study were drawn is available on the Kaggle platform. The present authors did not modify this dataset prior to the quality-control and feature-extraction stages. The dataset is cited as reference and can be accessed at:

Darquea, B. et al. (2021). PSG-Audio – Apnea Audios [Dataset]. Kaggle. https://www.kaggle.com/datasets/bryandarquea/psg-audio-apnea-audios

The PSG-Audio dataset from which the Kaggle derivative was produced was published by Korompili et al. and is hosted on the Science Data Bank (ScienceDB). This dataset comprises the original scored polysomnographic recordings and is cited as the primary clinical source. Access to the full raw recordings is contingent upon adherence to the data-sharing conditions stipulated by the original authors; interested researchers are advised to contact the corresponding author directly. The dataset’s publication metadata and access information are available at:

Korompili, G. et al. (2021). PSG-Audio: A scored polysomnography dataset with simultaneous audio recordings for sleep apnea studies. Scientific Data, 8. https://doi.org/10.1038/s41597-021-00977-w

The complete source code for all preprocessing, feature extraction, model training, evaluation, and figure-generation procedures is available to the public in the following GitHub repository, cited as reference:

Torres-Sánchez, S. (2025). Sleep Apnea Detection Model [Software]. GitHub. https://github.com/SebasTS15/Sleep-Apnea-Detection-Model

The authors gratefully acknowledge the financial and institutional support provided by the Unidad Central del Valle del Cauca (UCEVA), Colombia and the Universidad de San Buenaventura, Cali, Colombia. This work was developed within the framework of the research project “Construcción de un prototipo de experimentación en tecnologías de ciudades inteligentes que permita validar soluciones en movilidad, ambiente, salud, conectividad y gestión urbana.”

•  Alarcón AS, Madrid NM, Seepold R, et al.: Obstructive Sleep Apnea Event Detection Using Explainable Deep Learning Models for a Portable Monitor. Front. Neurosci. 2023; 17: 1155900. Publisher Full Text
•  Attia S, Oksenberg A, Levy J, et al.: Clinical Validation of Artificial Intelligence Algorithms for the Diagnosis of Adult Obstructive Sleep Apnea and Sleep Staging From Oximetry and Photoplethysmography—SleepAI. J. Sleep Res. 2025.
•  Barroso-García V, Fernández-Poyatos M, Sahelices B, et al.: Prediction of the Sleep Apnea Severity Using 2D-Convolutional Neural Networks and Respiratory Effort Signals. Diagnostics (Basel). 2023; 13(20). Reference Source
•  Chaw HT, Kamolphiwong T, Kamolphiwong S, et al.: ZleepNet: A Deep Convolutional Neural Network Model for Predicting Sleep Apnea Using SpO2 Signal. Applied Computational Intelligence and Soft Computing. 2023; 2023(1): 1–12. Publisher Full Text
•  Chen J-W, Liu C-M, Wang C-Y, et al.: A deep neural network-based model for OSA severity classification using unsegmented peripheral oxygen saturation signals. Eng. Appl. Artif. Intell. 2023; 122: 106161. Publisher Full Text
•  Crowson MG, Gipson KS, Kadosh OK, et al.: Paediatric sleep apnea event prediction using nasal air pressure and machine learning. J. Sleep Res. 2023; 32(4).
•  Darquea B: PSG-Audio – Apnea Audios. (Dataset). Kaggle. Science Data Bank; 2021. Reference Source
•  Ganguly B, Dasgupta R, Dey D: A residual deep learning framework for sleep apnea diagnosis from single lead electrocardiogram signals: An explainable artificial intelligence approach. Eng. Appl. Artif. Intell. 2025; 148: 110481. Publisher Full Text
•  Ghafourian MS, Noori A, Taghipour S, et al.: Multimodal apnea detection: advancements through convolutional neural networks and STFT analysis of EEG, ECG, and nasal signals to tackle key challenges in innovation. Comput. Methods Biomech. Biomed. Engin. 2025; 1–11. Publisher Full Text Reference Source
•  Gosselin N, Baril AA, Osorio RS, et al.: Obstructive Sleep Apnea and the Risk of Cognitive Decline in Older Adults. Am. J. Respir. Crit. Care Med. 2019; 199(2): 142–148. Publisher Full Text Reference Source
•  Javeed A, Berglund JS, Dallora AL, et al.: Predictive Power of XGBoost_BiLSTM Model: A Machine-Learning Approach for Accurate Sleep Apnea Detection Using Electronic Health Data. Int. J. Comput. Intell. Syst. 2023; 16(1). Publisher Full Text
•  Kim Y, Kim M, Shin J, et al.: ApneaWhisper: Transformer-Based Audio Segmentation for Sleep Apnea Detection. Nature and Science of Sleep. 2025; 17: 2455–2468. PubMed Abstract | Publisher Full Text | Free Full Text
•  Korompili G, Amfilochiou A, Kokkalas L, et al.: PSG-Audio: A scored polysomnography dataset with simultaneous audio recordings for sleep apnea studies. Sci. Data. 2021; 8(1).
•  Kumar CB, Bhongade A, Panigrahi BK, et al.: Exploring Siamese-Based Self-Supervised Learning for Sleep Apnea Detection. Comput. Intell. 2025; 41(3): e70080. Publisher Full Text
•  Le VL, Kim D, Cho E, et al.: Real-Time Detection of Sleep Apnea Based on Breathing Sounds and Prediction Reinforcement Using Home Noises: Algorithm Development and Validation. J. Med. Internet Res. 2023; 25: e44818. Publisher Full Text
•  Li F, Xu Y, Chen J, et al.: A Deep Learning Model Developed for Sleep Apnea Detection: A Multi-Center Study. Biomedical Signal Processing and Control. 2023; 85: 104689. Publisher Full Text
•  Liu K, Geng S, Shen P, et al.: Development and Application of a Machine Learning-Based Predictive Model for Obstructive Sleep Apnea Screening. Frontiers in Big Data. 2024; 7: 1353469. PubMed Abstract | Publisher Full Text | Free Full Text
•  Mortazavi E, Tarvirdizadeh B, Alipour K, et al.: Deep learning approaches for assessing pediatric sleep apnea severity through SpO2 signals. Sci. Rep. 2024; 14(1): 22696. PubMed Abstract | Publisher Full Text | Free Full Text
•  Nandakumar R, Arunachalam R, Pugalenthi R, et al.: Automatic model of sleep apnea detection using optimized weighted fusion process of hybrid convolution (1D/2D) efficient attention network from EEG signals. EURASIP Journal on Advances in Signal Processing. 2025; 2025(1). Publisher Full Text
•  Osa-Sanchez A, Ramos-Martinez-de-Soria J, Mendez-Zorrilla A, et al.: Wearable Sensors and Artificial Intelligence for Sleep Apnea Detection: A Systematic Review. J. Med. Syst. 2025; 49(1): 66. PubMed Abstract | Publisher Full Text | Free Full Text
•  Padovano D, Martinez-Rodrigo A, Pastor JM, et al.: Deep Learning and Recurrence Information Analysis for the Automatic Detection of Obstructive Sleep Apnea. Appl. Sci. 2025; 15(1). Publisher Full Text
•  Quironsalud: ¿Qué consecuencias tiene la apnea del sueño?|Hospital Universitario Quirónsalud Madrid.2025. Reference Source
•  Saha S, Fattah SA, Saquib M: StApneaNet: A Deep Learning-Based Automatic Sleep Stage Adaptive Apnea Detection Network Using Single Channel EEG Signal. IEEE Access. 2024; 12: 198250–198261. Publisher Full Text
•  Serrano S, Patanè L, Serghini O, et al.: Detection and Classification of Obstructive Sleep Apnea Using Audio Spectrogram Analysis. Electronics. 2024; 13(13): 2567. Publisher Full Text
•  Siyahjani F, Kalenyk K, Garcia-Molina G, et al.: Validating a smart bed against polysomnography for sleep apnea detection. Sci. Rep. 2025; 15(1): 23607. PubMed Abstract | Publisher Full Text | Free Full Text
•  Sturekova J, Kamencay P, Sykora P, et al.: A Comparison of Convolutional Neural Network Transfer Learning Regression Models for Remote Photoplethysmography Signal Estimation. AI. 2025; 6(2): 24. Publisher Full Text
•  Torres-Sánchez S: Sleep Apnea Detection Model – Source Code Repository.2025.
•  Tunubala JS, Quitian GV, Paredes CM, et al.: Sleep Apnea Detection from Respiratory Sounds: Dataset Documentation. Mendeley Data. 2026; V1. Publisher Full Text
•  Tzeng J-T, Li J-L, Chen H-Y, et al.: Improving the Robustness and Clinical Applicability of Automatic Respiratory Sound Classification Using Deep Learning–Based Audio Enhancement: Algorithm Development and Validation. JMIR AI. 2025; 4: e67239. Publisher Full Text
•  Ullah N, Mahmood T, Kim SG, et al.: DCDA-Net: Dual-convolutional dual-attention network for obstructive sleep apnea diagnosis from single-lead electrocardiograms. Eng. Appl. Artif. Intell. 2023; 123: 106451. Publisher Full Text
•  Ullah N, Sultan H, Hong JS, et al.: Convolutional self-attention with adaptive channel-attention network for obstructive sleep apnea detection using limited training data. Eng. Appl. Artif. Intell. 2025; 156: 111154. Publisher Full Text
•  Varshini GV, Hussain M, Lee JW: Sleep Apnea and Rapid Eye Movement Detection using ResNet-50 and Gradient Boost. Int. J. Adv. Comput. Sci. Appl. 2024; 15. Reference Source
•  Wang B, Tang X, Ai H, et al.: GA–Adam hybrid optimization for text classification. Inf. Process. Manag. 2021; 58(6): 102708. Publisher Full Text
•  Wang B, Tang X, Ai H, et al.: Obstructive sleep apnea detection based on sleep sounds via deep learning. Nature and Science of Sleep. 2022; 14: 2033–2045. PubMed Abstract | Publisher Full Text | Free Full Text
•  Wang E, Koprinska I, Jeffries B: Sleep Apnea Prediction Using Deep Learning. IEEE J. Biomed. Health Inform. 2023; 27(11): 5644–5654. Publisher Full Text
•  Wang T, Strodthoff N: S4Sleep: Elucidating the design space of deep-learning-based sleep stage classification models. Comput. Biol. Med. 2025; 187: 109735. PubMed Abstract | Publisher Full Text
•  Wu Y, Mi Q, Gao T: A Comprehensive Review of Multimodal Emotion Recognition: Techniques, Challenges, and Future Directions. Biomimetics. 2025; 10(7): 418. PubMed Abstract | Publisher Full Text | Free Full Text
•  Zhang Y, Shi Y, Su Y, et al.: Detection and severity assessment of obstructive sleep apnea according to deep learning of single-lead electrocardiogram signals. J. Sleep Res. 2025; 34(1).