More than half of the studies (n=18) used proprietary datasets collected within research facilities, with dataset size ranging from fewer than 50 sleep records (16, 17) to several hundreds (18–21). While proprietary datasets allow tailored data collection and expert labeling, their limited public availability poses challenges for comparison and benchmarking.

In contrast, public datasets provide valuable alternatives. Widely-used ones include the Apnea-ECG Database (AED) (22), the St.Vincent’s University Hospital/University College Dublin Sleep Apnea Database (UCD) (23), and more recently, the OSASUD dataset (24). However, the relatively small sizes of these datasets (often fewer than 100 records) limit their applicability primarily to epoch-wise, rather than subject-wise screening (8, 9). The Sleep Heart Health Study (SHHS) (25), with over 5,000 recordings and high-resolution labels, is ideal for deep learning models. Other publicly available datasets, such as the Wisconsin Sleep Cohort dataset (n=2,570) (26), Cleveland Family Study (CFS) dataset (n=2,284) (27), Osteoporotic Fractures in Men Study (MROS) dataset (n=3,753) (28) and the Multi-Ethnic Study of Atherosclerosis (MESA) dataset (n=2,002) (29), offer valuable data but have been less utilized in sleep apnea research so far.

Physiological signals such as SpO₂ are prone to movement contamination and thus require pre-processing to remove noises. However, our analysis revealed that many studies (n=11) proceed with raw SpO₂ without pre-processing, and there is a lack of standardized protocols for filtering noise or assessing signal quality.

SpO₂ are typically calibrated within a range of 70% to 100% saturation, with an accuracy of ±2% to ±4% (30). Consequently, readings below 70% may be inaccurate, prompting some studies to use thresholds of 70% or 65% to remove unreliable readings (3, 31). A lower threshold of 50% has also been used in several studies to account for physiological limitations and equipment errors (9, 12, 17, 20, 32). In addition to cut-off thresholds, some studies applied further noise reduction techniques, such as removing data points where consecutive SpO₂ values differ by more than a predefined value (e.g., 4%) (3, 12, 20, 33).

Few studies (n=3) explored optimal SpO₂ signal bands. One study identified the apnea-related band as 0.014–0.033 Hz (33), while (21) argued that the shape of SpO₂ signal is similar to a sinusoid with 0.02 Hz frequency and therefore used a 0.02 Hz IIR Butter-worth low-pass filter to suppress and smooth the SpO₂ signal. Another study employed a complex Wavelet filter to eliminate noise from muscle movements (4). Furthermore, Stuban and Niwayama (34) demonstrated that lowering the low-pass filter frequency to a value closer to the fundamental frequency of the PPG signal reduced noise without compromising measurement accuracy. An additional 10 dB of signal-to-noise ratio (SNR) is recommended for accurate SpO₂ measurement (35). However, no study provided a detailed justification for these methods, nor did they analyzed how these techniques influenced signal quality and subsequent classification performance.

The most common pre-processing practice involves using a threshold between 50% and 100% to eliminate hardware errors. However, to determine the optimal filter settings, future studies are needed to conduct a comprehensive benchmarking of all possible filter options.

Two primary approaches dominate feature extraction methods. The first approach relies on manual feature extraction, leveraging the researchers’ domain expertise. This approach has been applied in 19 studies, where hand-crafted features were derived to ensure model interpretability and applicability across various shallow learning models. While this method can be time-consuming, it remains valuable for its transparency and ease of understanding. A comprehensive list of SpO₂ features is detailed in Xie and Minn (9), Gutiérrez-Tobal et al. (36), Levy et al. (37).

To optimize features for model construction, several techniques have been employed to select features with strong discriminating power. These include forward stepwise logistic regression (18), recursive feature elimination (38), fast correlation-based filter (36), maximum relevance minimum redundancy (11), and heat-map (39). One commonly used feature is the Oxygen Desaturation Index (ODI), due to its strong correlation with the Apnea-Hypopnea Index (AHI) which is a standard measurement of sleep apnea severity (7, 11, 20, 33, 39–42). However, ODI requires a minimum sampling rate of 1 Hz, which limits its applicability in datasets from lower-frequency devices like smartwatches. Other features such as entropies (7, 12, 40), Lempel-Ziv complexity (LZ) (7, 33, 43), and demographic features (e.g., age, gender, neck circumference, body mass index (BMI)) (41, 44, 45) have also gained prominence. These features have been ranked highly in importance compared to others.

Deep learning (DL) has recently emerged as a novel feature extraction paradigm. DL models, such as convolutional neural networks (CNNs) (46, 47) and long short-term memory (LSTM) networks (8), are capable of automatically extracting features, potentially uncovering patterns that manual feature extraction may overlook. For instance, a study by Lyden et al. (48) demonstrated that CNN and LSTM models achieved high performance in epoch-wise classification, with accuracy, sensitivity, and precision exceeding 90%, even when working with reduced signal sampling rates. This highlights the effectiveness of DL in apnea screening. However, the trade-off is that the interpretability and explainability of these extracted features remain an ongoing challenge as visualizing how the variables are interconnected and weighted within the network is virtually impossible (49).

Manually crafting features is a viable approach to enhancing the explainability and interpretability of machine learning methods, particularly when aiming for clinical acceptance. Unlike features derived from deep learning models, which often suffer from the “black box” problem that limits transparency, handcrafted features are well-defined, easy to visualize, and straightforward to interpret. Feature-ranking techniques, such as SHAP or Grad-CAM, can further facilitate a deeper understanding of how machine learning algorithms work. Ensuring that methods are explainable and transparent not only improves their transition to real-world applications but also enables targeted interventions when the model produces incorrect predictions.

All studies in this review employ supervised learning for sleep apnea screening. As shown in Figure 1, the studies were categorized into three main problem formulations, epoch-wise screening, AHI regression, and subject-wise screening.

The studies reviewed reveal a combination of classification techniques for the screening of sleep apnea. Table 1 shows various shallow and deep learning models have been applied. However, it is difficult to conclude which method is the most optimal due to the lack of standardized evaluation metrics. For classification problems, evaluation metrics include accuracy (ACC), sensitivity (SE), specificity (SPEC), F1 score, Kappa coefficient, Matthews correlation coefficient (MCC), and area under the curve (AUC). On the other hand, regression problems require a distinct set of evaluation metrics, such as correlation coefficient, intraclass correlation coefficient (ICC), Bland-Altman plots, root mean squared error (RMSE), mean absolute error (MAE), and R-squared. There has yet to be a consensus as to which metrics are the best for evaluating model performance.

Our analysis identified two major research gaps in AI-empowered sleep apnea screening using SpO2 measurements. The first gap pertains to the limitations in the quantity and diversity of datasets. Specifically, there is a scarcity of large-scale, open-access datasets. Large volumes of training data are essential for developing robust AI models, and sharing open datasets is increasingly crucial for advancing research. However, we found that most studies used proprietary datasets that are not publicly available. Despite our focus on wearable deceives for apnea detection, we identified only four datasets collected with such devices (19, 36, 53, 56), none of which are openly accessible. In addition, most of the open sleep datasets were collected in Western countries and predominantly included data from Caucasian individuals. The lack of data from diverse populations, especially Black, Hispanic, and Asian groups, poses a challenge to developing generalizable AI models across different demographic groups (11, 39).

Furthermore, in the context of home-based or self-tracking, the user range has expanded beyond patients to include healthy, young individuals in everyday life. The innovation of portable and wearable sensors has highlighted the urgent need for datasets derived from smartwatches, smart rings, and similar devices. The closer these datasets reflect real-world conditions, the more effectively machine learning algorithms can be applied in practice. Another promising trend is the use of longitudinal data, which provides a more detailed and stable representation of health status over time. In the near future, sleep apnea data from healthy young individuals, collected through consumer devices over extended periods, will be essential for advancing research and improving screening models.

The second major gap is the need for standardization in data collection, signal preprocessing, and model benchmarking. This lack of standardization creates challenges for reproducibility, comparability and generalizability. Data collected across different studies vary significantly due to differences in the type of devices used, the protocols followed (e.g., timing, duration, sensor placement), and the environmental conditions under which data are gathered. Signal preprocessing is another area where standardization is absent. Steps such as noise filtering and artifact removal are often performed differently across studies, leading to inconsistencies in the data quality. In addition, variations in datasets used for model training and testing, coupled with differences in evaluation metrics, hinder direct comparisons between models (19). Many studies focus solely on accuracy (e.g., (11, 31, 47)), which can be misleading, especially in the presence of unbalanced data. The generalizability of models is frequently overlooked; only one study extensively investigated model performance on different datasets other than the training set (37). In contrast, several studies reported decreased model performance when applied to new datasets (21, 31, 46, 47). To facilitate meaningful cross-model comparisons, sharing source code is recommended; however, only a few studies have done so (21, 37, 45).

To address the identified research gaps in AI-powered sleep apnea screening using SpO2 measurements, future research should focus on enhancing dataset quality and standardization. Developing and sharing large-scale, open-access datasets that include diverse populations is crucial for training robust and generalizable AI models. Collaborative data initiatives involving academia, industry, and healthcare organizations can facilitate the creation of comprehensive datasets. Additionally, establishing standardized protocols for data collection and signal preprocessing will improve reproducibility and comparability across studies. Formulating and disseminating guidelines for data collection and preprocessing, along with creating benchmarking frameworks, can help ensure consistency and facilitate meaningful cross-study comparisons.

Moreover, advancing model evaluation and generalizability is essential for improving AI performance in sleep apnea screening. Future research should emphasize evaluating models on diverse datasets and adopting robust evaluation metrics that account for data imbalance and biases. Encouraging transparency by sharing methodologies, preprocessing scripts, and source code can enhance reproducibility and foster a more reliable research environment. Exploring multi-modal data approaches and innovative sensor technologies could also improve the accuracy and robustness of models. By addressing these areas, the field can make significant strides toward developing more effective and generalizable AI solutions for sleep apnea detection.

Finally, research on AI-based sleep apnea research has been lacking a user-centered perspective, and this needs to be addressed in future studies. In addition to improving model performance, future research must also prioritize clinical applicability and user adoption (59). Keeping experts involved by collecting their feedback to continuously retrain the models can help better align the inner workings of the models with expert decision-making processes. Engaging with both patients and clinicians through pilot studies, usability testing, and observational trials can provide valuable insights into model interpretability, user trust, and practical deployment. Last but not the least, designing user-friendly interfaces that visualize model predictions in an easy-to-understand manner is important for fostering user acceptance and real-world applicability.