(PDF) Transformer based deep learning approach for obstructive sleep apnea detection using single lead ECG

TYPE Original Research PUBLISHED 11 February 2026 DOI 10.3389/frai.2026.1727091 REVIEWED BY Transformer-based deep learning approach for obstructive sleep apnea detection using single-lead ECG Shuaiwei, SongTongji University, China Yiming Qin, Tsinghua University, China Malak Abdullah Almarshad   1*, Saad Al-Ahmadi   2, Saiful Islam   3, Adel Soudani   2 and Ahmed S. BaHammam   4,5 OPEN ACCESS EDITED BY Lin-Ching Chang, The Catholic University of America, United States *CORRESPONDENCE Malak Abdullah Almarshad

RECEIVED 17 October 2025 REVISED 20 January 2026 ACCEPTED 23 January 2026 PUBLISHED 11 February 2026 Computer Science Department, College of Computer and Information Sciences, Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, Saudi Arabia, 2Computer Science Department, College of Computer and Information Sciences, King Saud University, Riyadh, Saudi Arabia, 3Department of Computer Engineering, TED University, Ankara, Türkiye, 4The University Sleep Disorders Center, Department of Medicine, College of Medicine, King Saud University, Riyadh, Saudi Arabia, 5Strategic Technologies Program of the National Plan for Sciences and Technology and Innovation in the Kingdom of Saudi Arabia, Riyadh, Saudi Arabia 1 CITATION Almarshad MA, ​Al-Ahmadi S, Islam S, Soudani A and BaHammam AS (2026) Transformer-based deep learning approach for obstructive sleep apnea detection using single-lead ECG. Front. Artif. Intell. 9:1727091. doi: 10.3389/frai.2026.1727091 COPYRIGHT © 2026 Almarshad, Al-Ahmadi, Islam, Soudani and BaHammam. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms. Obstructive sleep apnea (OSA) results from repeated collapses of the upper airway during sleep, which can lead to serious health complications. Although polysomnography (PSG) is the diagnostic gold standard, it is costly, labor-intensive, and associated with long waiting times. With the rapid evolution of automated scoring solutions and the emergence of machine learning (ML) and deep learning (DL) in many disciplines, there is a need for tools that use fewer signals and can provide accurate diagnoses. DL models can an process large amounts of data and often generalize effectively to new instances. This makes them a suitable choice for classifying continuous time series data. This study introduces a transformerbased deep learning approach using a single-lead electrocardiogram (ECG) for OSA detection. The proposed architecture, designed to handle raw signals with high sampling rates, preserves temporal continuity over unlimited durations. Without any preprocessing, the model tolerates high-noise raw data. The model is tested with different positional embedding techniques. Additionally, a novel positional encoding technique using an autoencoder is introduced. The proposed approach achieves a high F1 score, outperforming other published work by an average margin of more than 13%. In addition, the model classifies apnea episodes at one-second intervals, providing clinicians with nuanced insights. KEYWORDS artificial intelligence (AI), autoscoring, deep learning (DL), electrocardiogram (ECG), healthcare, obstructive sleep apnea (OSA), polysomnography (PSG), time-series classification (TSC) 1 Introduction Obstructive sleep apnea (OSA) is a disorder characterized by repeated partial or complete blockage of the upper airway during sleep (Halani, n.d.). A study in 2019 reported a prevalence of OSA with almost one billion people affected and around 50% prevalence in middle-aged and older adults in some countries (Benjafield et al., 2019). Undiagnosed and untreated OSA can have profound implications, including cardiovascular diseases (CVDs) (Zhao et al., 2024; Cai et al., 2023), stroke, metabolic disease, lower quality of life, and decreased productivity. Given the high prevalence and serious consequences of OSA, greater efforts are required to achieve accurate and early diagnosis (Benjafield et al., 2019). Frontiers in Artificial Intelligence 01 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 Polysomnography (PSG) is considered the gold-standard diagnostic test for OSA. PSG is a full or split night study that is done in the sleep unit at the hospital, to monitor patients’ sleep architecture and several respiratory parameters by collecting a bundle of signals. The American Academy of Sleep Medicine (AASM) guidelines for PSG include electroencephalogram (EEG), and chin electromyogram (EMG) for sleep staging, electrocardiogram (ECG) for heart rate (HR) and arrhythmias, thermal sensors and a nasal pressure transducer to monitor respiratory flow, and photoplethysmography (PPG) for oxygen saturation (Almarshad et al., 2022). Sleep studies (PSG) are expressed as thirty-second epochs of raw data, recorded for 8 h, which are interpreted into around 900 pages (Gupta et al., 2018). Recorded PSG needs a trained technician to score it, which is tedious and time-consuming. To date, the standard PSG scoring process is performed manually (Figure 1). The entire process is complex and costly, potentially causing delays in the diagnosis and treatment of patients with OSA. Over the last 10 years, various automated scoring solutions have evolved, including statistical analysis, signal processing, machine learning (ML), and deep learning (DL) methods. Several DL network architectures are used for apnea detection, including multilayer perceptron (MLP), convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory (LSTM) (Jahrami et al., 2025). So far, LSTM has achieved the best results (Faust et al., 2021). Although currently available auto-scoring algorithms for sleep apnea have shown promise, there is still a need for further development and validation of these methods that utilize fewer signals and provide more accurate diagnosis. Among the various signals, ECG signal has a significant role in the detection of OSA as presented in Table 1. Cyclic variations in RR intervals of ECG signals have been reported to correlate with OSA events, resulting in a pattern of bradycardia and tachycardia (Veasey and Rosen, 2019; Mangrum and Dimarco, 2014). Heart rate variability, which can be accurately extracted from the ECG signal, is a key biomarker for the detection of sleep apnea (Verma et al., 2020). This encourages the development of DL models that detect OSA utilizing single-lead ECG exclusively (Table 1). This pattern can be promising in detecting patients with clinical sleep apnea symptoms. However, further research is necessary to validate these findings and assess the reliability of classifying OSA from ECG signals using deep learning. Generally, four main architectures of deep networks are used, deep vanilla neural network (DNN) (Li et al., 2018; De Falco et al., 2019; Tagluk and Sezgin, 2011), convolution neural network (CNN) (Dey and Chaudhuri, 2018; Choi et al., 2018; Taghizadegan et al., 2021; Mashrur et al., 2021; Urtnasan et al., 2018), recurrent neural network (RNN) (Signals, 2020; Urtnasan and Lee, 2020) and long short-term memory (LSTM) (Faust et al., 2021; Drzazga and Cyganek, 2021). Some researchers also developed hybrid models (Li et al., 2018; Zhang et al., 2021; Almutairi et al., 2021; Chang et al., 2020; Guijarroberdiñas et al., 2012; Hu et al., 2022). Shuaicong, et al. developed a hybrid attention model (Hu et al., 2022), utilizing the sinusoidal positional encoding. Recently, Biswas and Abu Yousuf, (2025) achieved state-of-the-art (SOTA) on the apnea-ECG database, using a Transformers based model with a 1D CNNs. Most of these research with coarse-grained apnea labeling (Li et al., 2018; Mostafa et al., 2019; Ismail Fawaz et al., 2019). Additionally, advanced filtering techniques are employed to minimize noise (Bernardini et al., 2022; Bernardini et al., 2021; Sheta et al., 2021b). Furthermore, some approaches depend heavily on extensive data preprocessing and feature extraction (Faust et al., 2021; Sheta et al., 2021a; Wang et al., 2019). Transformers and their variants, such as BERT, GPT, and ChatGPT, have been proved efficient for multiple natural language processing (NLP) tasks. In this paper, we propose a Transformerbased deep learning framework for classifying PSG as a time series data. An autoencoder with a convolutional transpose (convt) layer is proposed to focus on learning the best representation of the positions FIGURE 1 Polysomnographic recording, scored manually by a professional technician at the University Sleep Disorders Center, at King Saud University Medical City (KSUMC), showing some different hypopneas events, within 2 min time window. Frontiers in Artificial Intelligence 02 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 TABLE 1 Different works for apnea detection from PSG using deep learning, arranged by year of publication. Paper Tagluk and Sezgin (2011) Guijarro-berdiñas et al. (2012) Year Dataset No. of recordings 2011 Proprietary 20 2012 Proprietary 6 Li et al. (2018) 2018 Choi et al. (2018) 2018 Urtnasan et al. (2018) Dey and Chaudhuri (2018) De Falco et al. (2019) 2018 2018 2019 Wang et al. (2019) 2019 Arslan et al. (2019) 2019 Erdenebayar et al. (2019) Singh and Majumder (2019) Apnea-ECG Database (Penzel et al., 2000) Proprietary and MESA (Bild et al., 2002) Proprietary Apnea-ECG Database (Penzel et al., 2000) Sleep Heart Health Study Apnea-ECG Database (Penzel et al., 2000) Proprietary Signal type Classifier EEG ANN 96.15% ANN 90.27% ECG DNN 85% Nasal pressure CNN 96.6% – ECG CNN 90–93% 70 ECG CNN 98.91% 17 ECG DNN 72.95% 70 RR Residual network 94.4% – PTT AlexNet and VGG-16 92.78% Nasal airflow and thoracic 70 179 + 50 Accuracy (DNN, 1D CNN, 2D 2019 Proprietary 86 ECG CNN, RNN, LSTM, 99.0% and GRU) 2019 Apnea-ECG Database (Penzel et al., 2000) 70 ECG AlexNet – Respiratory signals Autoencoder 86.22% Apnea-ECG Database Abreu et al. (2020) 2020 (Penzel et al., 2000) And Proprietary 95 ± 3.5% and 87 ± 6.6% (BrainAnswer RGBT) Deep RNN model Urtnasan and Lee (2020) utilizes long short2020 Proprietary – ECG term memory 98.5–99.0% (LSTM) and a gatedrecurrent unit (GRU) MIT-BIH Polysomnographic Chang et al. (2020) 2020 Database (Ichimaru and Moody, 1999) One-dimensional – ECG Apnea-ECG Database (1D) deep convolutional neural 97.1% network (CNN) (Penzel et al., 2000) Jarchi et al. (2020) 2020 Proprietary – ECG and EMG DNN 72% Deep BiLSTM-based 85.6 LSTMRNN 89.3% Oronasal thermal airflow (FlowTh), nasal pressure (NPRE), and Signals (2020) 2020 Proprietary – abdominal respiratory inductance plethysmography (ABD). Chang et al. (2020) 2020 Proprietary – SpO2 and ECG (Continued) Frontiers in Artificial Intelligence 03 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 TABLE 1 (Continued) Paper Year Dataset No. of recordings Signal type Classifier Accuracy 70 + 30 ECG and SpO2 CNN-LSTM 98.7% – RR derived from ECG LSTM 99.80% 70 ECG and SpO2 Apnea-ECG Database Bernardini et al. (2021) 2021 (Penzel et al., 2000) and OSASUD (Bernardini et al., 2022) MIT-BIH Faust et al. (2021) 2021 Polysomnographic Database (Ichimaru and Moody, 1999) Apnea-ECG Database Mashrur et al. (2021) 2021 (Penzel et al., 2000) , UCDDB datasets scalogram-based convolutional neural 95.71 network (SCNN) (Heneghan, 2011) Multiscale dilation attention 1-D convolutional neural Sheta et al. (2021a) 2021 Apnea-ECG Database (Penzel et al., 2000) 70 RR from ECG network (MSDA1DCNN) and 89.4% a weighted-loss timedependent (WLTD) classification Yue et al. (2021) Zhang et al. (2021) Almutairi et al. (2021) 2021 2021 2021 2 datasets, their own dataset Apnea-ECG Database (Penzel et al., 2000) Apnea-ECG Database (Penzel et al., 2000) Nasal pressure airflow – signals Cyganek (2021) Polysomnographic CNN and LSTM 96.1% 70 ECG CNN-LSTM 94.27% Oronasal airflow, the thoracic and abdominal – respiratory Database (Ichimaru and Sheta et al. (2021b) Mukherjee et al. (2021) 2021 2021 2021 Proprietary Apnea-ECG Database (Penzel et al., 2000) Apnea-ECG Database (Penzel et al., 2000) LSTM 80.66%/82.04% effort signals. Moody, 1999) Leino et al. (2021) 91.2%, (Mr-ResNet) ECG 1997) and MIT-BIH 2021 residual network 70 SHHS-1 (Quan et al., Drzazga and multi-resolution – SpO2 CNN 88.3% 70 ECG CNNLSTM 86.25% 70 ECG 2 CNNs, and CNN + LSTM, MLP 85.58% based ensemble. an ensemble of Taghizadegan et al. (2021) 2021 UCDDB datasets (Heneghan, 2011) recurrence plots – EEG and ECG (RPs) and pre-trained 89.45% convolutional neural networks (RPCNNs) Zarei et al. (2022) 2022 UCDDB datasets (Heneghan, 2011) – ECG CNN and LSTM 97.21% a hybrid model that Hu et al. (2022) 2022 Apnea-ECG Database (Penzel et al., 2000) contains an altered 70 ECG self-attention 90.5% mechanism from Transformers (Continued) Frontiers in Artificial Intelligence 04 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 TABLE 1 (Continued) Paper Almarshad et al. (2023) Biswas and Abu Yousuf, (2025) Year 2023 2024 Dataset OSASUD (Bernardini et al., 2022) Apnea-ECG Database (Penzel et al., 2000) No. of recordings Signal type Classifier Accuracy 30 SpO2 Transformers 80.0% 70 ECG of the samples. These samples, along with their learned positional embeddings, were input into Transformer encoder blocks to capture the most relevant information using the self-attention mechanism. Such a model can serve as a component within a larger system to analyze raw signals prior to further processing. Unlike traditional methods that rely on handcrafted features, our approach leverages the Transformer’s self-attention mechanism to automatically capture relevant patterns. In addition, it does not require any preprocessing step, having the ability to deal with raw data of high noise level. Apnea related abnormalities are detected at a one-second granularity, enabling fine-grained temporal localization of clinically marked apnea events. Which provide physicians with detailed insights into the patient’s condition and facilitate the interpretation and validation of the model’s results. The key contributions of this study can be summarized as follows: Transformers 91.97% processing patients’ videos while sleeping, detecting apnea from wrist actigraphy, smartwatches, or detecting apnea from snoring sounds, are outside the intended context. Table 1 presents a concise yet comprehensive chronological summary of studies utilizing PSG recordings. For each entry, it provides details on the type of analyzed signal (ECG, SpO₂, or both), the dataset type and name (public or proprietary), population size, signal characteristics, the employed deep learning (DL) model, and its reported accuracy. Several studies have utilized only one physiological signal to detect apnea events. The vast majority of them used DL on ECG only (Li et al., 2018; De Falco et al., 2019; Dey and Chaudhuri, 2018; Urtnasan et al., 2018; Urtnasan and Lee, 2020; Zhang et al., 2021; Almutairi et al., 2021; Sheta et al., 2021a,b; Erdenebayar et al., 2019; Singh and Majumder, 2019; Zarei et al., 2022; Mukherjee et al., 2021; Chang et al., 2020; Sun et al., 2022); in comparison, some considered SpO2 (Almarshad et al., 2023; Denker, 2022; Leino et al., 2021). Fewer studies took advantage of more than one signal, such as: ECG and SpO2 together (Mashrur et al., 2021; Chang et al., 2020; Bernardini et al., 2021). However, a couple of studies rely on the RR interval derived from ECG (Faust et al., 2021; Sheta et al., 2021a; Wang et al., 2019). While a few researchers choose to build their own datasets from scratch, most rely on publicly available benchmark datasets. Three datasets are widely used in the literature, namely St. Vincent’s University Hospital/University College Dublin Sleep Apnea Database (Heneghan, 2011), MIT-BIH polysomnographic database (Ichimaru and Moody, 1999), and APNEA-ECG database (Penzel et al., 2000). Evidence-based medical research depends critically on the availability of raw data of sufficient quantity and quality, not to mention that there are multiple concerns about patients’ privacy, organizational structures, and legal challenges (BaHammam and Chee, 2022). All these contributed to the fact that the most used apnea dataset is two decades old (Chang et al., 2020). Since then, several sleep study standards practices have changed (Padovano et al., 2022). However, multiple papers gained much attention, using their own datasets. Bernardini et al. (2022) published an interesting, comprehensive dataset for 30 admitted patients with precise apnea syndrome severity annotation, and it has been used in their previous work (Bernardini et al., 2021). However, unlike previous datasets, its primary focus is on apnea, and the exclusion criteria were minimal. Consequently, the data are highly susceptible to noise. • Developing a Transformer-based model for OSA detection to support clinical decision-making, achieving optimal performance through a novel learnable positional encoding implemented via a simple convolutional autoencoder with a single transposed convolution layer; • Investigated the impact of various positional embedding strategies on model performance, using static and learnable embedding, and how the proposed learnable embedding using an autoencoder improves the overall model performance; • Multiple encoding models were proposed for OSA detection; our scheme is the first one that utilizes learnable positioning encoding via an autoencoder and it outperforms all previous models on OSASUD. The rest of this paper is organized as follows. Section 2 reviews related work and presents different DL approaches to classify apnea. In section 3, we discuss the proposed model and the used dataset. Section 4 explained the experimental setting. Section 5 reports the results in comparison with other deep learning approaches evaluated on the same dataset (Bernardini et al., 2021). Finally, Section 6 concludes the paper and discusses potential directions for future work. 2 Related work Different DL approaches have been proposed in the literature to identify sleep apnea and hypopnea events (Jahrami et al., 2025; Veasey and Rosen, 2019; Mangrum and Dimarco, 2014). We investigated the literature trying to cover published articles in the last two decades. Articles that classify apnea using other techniques, such as statistical methods, signal processing, classical machine learning like support vector machines (SVM), and decision trees (DT), were excluded. Also, papers that deal with other types of apnea detection methods, like Frontiers in Artificial Intelligence 1D CNNs + 3 Materials and methods This section presents the dataset, preprocessing steps, and the proposed Transformer-based architecture in a unified and reproducible manner. Figure 2 illustrates the complete processing pipeline, starting from raw ECG input to apnea event classification. The model consists of three main components: (i) input normalization 05 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 and segmentation, (ii) positional encoding (static or learnable), and (iii) a Transformer encoder followed by a classification head. Our work differs from existing studies in four key ways: (i) it employs a real-world dataset that is both noisy and imbalanced; (ii) apnea events are detected at one-second granularity; (iii) the proposed architecture effectively processes high-frequency raw signals while preserving temporal and spatial dependencies over long durations; and (iv) the model is validated on entirely unseen raw data. Figure 2 provides a general overview of the proposed architecture. The encoder component is shown in Figures 3, 4 describes the autoencoder positional embedding; The model processes the ECG sampling window as a whole, without shuffling. All experiments were conducted on a local machine equipped with an AMD Ryzen™ 95,900X CPU, an NVIDIA GeForce RTX 3080 GPU, and 32 GB of RAM. The models were developed using TensorFlow 2.10. For reproducibility, the source code of the proposed models are available at: https://github.com/ malakalmarshad/TOSA. time, high-pass and low-pass filters were set to ECG at 0.3 and 70 Hz, respectively. Each recording has a duration of approximately 7 to 12 h. To train our model, we used ECG lead II (signal_ecg_ii) as on (Bernardini et al., 2021) to compare the obtained result with previous DL models. 3.2 Data preprocessing Trying to comply with the fact that DL models tend to extract features by themselves and require a minimum amount of preprocessing, we preferred not to apply any filters and to assume that our model is capable of dealing with high levels of noise present in the OSUSA dataset. All input arrays were independently rescaled to the range [0, 1] using min-max normalization (Goodfellow et al., 2016). Moreover, all NaN values are replaced by zero. Ground truth anomaly data were segmented into non-overlapping 30-s windows, with each window represented by a list of 30 binary labels indicating the presence (true) or absence (false) of an anomaly at one-second intervals. In the OSASUD dataset, apnea and hypopnea events are originally annotated by clinicians as continuous temporal intervals, following the standard clinical definition of respiratory events lasting at least 10 s. To obtain one-second level labels, we project each event annotation onto a one-second time grid: all seconds whose timestamps fall within the temporal boundaries of a clinically annotated apnea or hypopnea event are labeled as positive, forming contiguous segments of positive labels that preserve the true event duration. No isolated positive seconds are introduced outside annotated events. When an apnea event spans across two adjacent 30-s windows, the corresponding seconds in both windows are labeled as positive, thus preserving temporal continuity across window boundaries. The learning objective of the model is second-level classification, aimed at fine-grained temporal detection. However, event-level performance is obtained by aggregating contiguous positive seconds into detected events and comparing them with clinically annotated events, ensuring consistency with the clinical definition of apnea. 3.1 Dataset Most previous studies have been evaluated on idealized datasets, which limits their applicability in real-world scenarios. In contrast, our work utilizes imbalanced and challenging datasets. The dataset we considered is the OSASUD (Bernardini et al., 2022) for detecting OSA syndrome. It consists of 30 patients’ overnight PSG recordings. Patients are affected by different OSAS severity, including 7 subjects without OSA (AHI < 5). Three reasons make OSASUD dataset challenging (Leino et al., 2021); first, PSG (ECG and PPG) recordings are continuous and not segmented into discrete time windows. Second, the dataset is highly contaminated by noise and contains missing and null samples, as is typical in real-world monitoring scenarios. Third, given that the ratio between apnea and normal samples is around 0.17, it is a highly skewed dataset. A trained sleep technician annotated the collected PSG data for apnea and hypopnea events with one-second temporal resolution. The sampling frequency is 80 Hz for both ECG and PPG. At the same FIGURE 2 General workflow of the proposed model. Frontiers in Artificial Intelligence 06 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 To preserve temporal order information, positional embeddings are added to the input sequence prior to the encoder stack. Three positional encoding strategies are evaluated: (i) naive positional encoding, (ii) fixed sinusoidal encoding, and (iii) the proposed learnable encoding based on a convolutional autoencoder. The output of the encoder stack is aggregated using a GlobalAveragePooling1D layer and passed to a multilayer perceptron (MLP) for binary classification. 3.3.1 Encoder stack The encoder comprises 6 identical layers, each with two sub-layers. The first sub-layer is a multi-head self-attention mechanism with 4 heads of size 256. The second sub-layer uses two convolutions with a ReLU activation in between. Following the original Transformer design, each sub-layer is enclosed by a residual connection and a subsequent normalization layer. The core part of our model is now completed. Multiple of those Transformer encoder blocks can be piled, but the best results we achieved empirically is by stacking 6 of them. A random search was obtained to find optimized combinations of hyperparameters to train the model. These hyperparameters include the number of heads and the number of encoder blocks (Table 2). To compress the output tensor of the encoder, a GlobalAveragePooling1D layer is added before the final MLP classification head. Alongside multi-head attention, positional embeddings allow Transformers to outperform earlier architectures. Without recurrence or convolution, positional encodings added to input embeddings at the base of the encoder and decoder stacks preserve positional information throughout all Transformer blocks (Vaswani et al., 2017). Sequence order is critical in time series data. Unlike CNNs, RNNs, and LSTMs, which inherently capture order, Transformers replace recurrence with multi-head self-attention for faster, parallelized training. We evaluated three positional embedding strategies: naive, fixed sinusoidal, and learned embeddings (Wang and Chen, 2020; Wang et al., 2022). Without positional encoding, samples are treated as a bag of words. Positional embeddings are directly added to the sequence representation as (Equation 1): FIGURE 3 Proposed transformers-base model general architecture. 3.3 Transformers_based model At the core of our approach for apnea event classification is a Transformer encoder, following the design proposed by Vaswani et al. (2017); We employ only the Transformer encoder, as our aim is to detect apnea events in the ECG signal rather than reconstruct it, a task typically performed by the decoder. Figure 3 illustrates the generic component of our model, used across all experiments. We refer to Vaswani et al. (2017) for a detailed description of the Transformer, and here highlight our modifications that adapt it for continuous univariate time series classification instead of generating sequences of discrete tokens. The proposed model takes as input a tensor of shape (batch size, sequence length, 1), where the sequence corresponds to a 30-s ECG window sampled at 80 Hz. Each window is associated with 30 binary labels, representing apnea presence at one-second intervals. The core of the architecture is a stack of six Transformer encoder blocks, each consisting of a multi-head self-attention layer (4 heads, head size 256) followed by a convolution-based feed-forward sublayer. Residual connections and layer normalization are applied after each sublayer, following the original Transformer design (Vaswani et al., 2017). Frontiers in Artificial Intelligence Zi =inputE( xi ) + PE(i ) (1) Here, xi is the sequence at the i-th position, inputE the input embedding, and PE the positional encoding, which may be learnable or predefined. 3.3.2 Naive positional encoding It iss a finite-dimensional representation of each sample’s index in a sequence. For a sequence, X = [ x0 , …, x k , xn−1 , xn ], the encoding tensor informs the model of the position of each element xi is in the sequence X. A fixed positional encoding can be calculated from the normalized sequence index as follows (Equation 2): PE ( i ) = 07 ( pos ( xi ) − min pos ( x ) ( ) ( ) max pos ( x ) − min pos ( x ) ) (2) frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 FIGURE 4 General proposed autoencoder architecture. Autoencoder learns a better representation of each input epoch, then feeds it to the transformer component (Figure 3). configuration for the dataset. Figure 5, depicts the PE vectors (Vaswani et al., 2017). TABLE 2 Hyperparameter tuning using random search. Name Tested Best 4, 8, 16, 32, 64, more than 64 Batch size 32 generates OOM error RMSprop, Adam, Optimizer AdamW Adam(amsgrad), AdamW Initial learning rate Learning rate scheduler 3.4 Learned positional embedding 1e-2, 1e-3, 1e-4, 1e-5, 1e-6 Positional encoding is a technique that incorporates information about the position of each token in the sequence to the input embeddings (Vaswani et al., 2017). This allows Transformers to understand the relative or absolute position of tokens which is important for differentiating between events in different positions and capturing the structure of a segment. Standard positional encoding techniques in Transformers can be broadly categorized into absolute encodings and relative encodings that model pairwise position differences. These methods assign positional information solely based on the index of a token in the sequence, independently of the input signal itself. In contrast, the proposed autoencoder-based positional representation is a data-driven, learnable positional encoding. Instead of encoding position as a function of the time index alone, the autoencoder learns a position-dependent representation directly from the raw ECG waveform. Therefore, the proposed method provides a content-aware positional embedding. The autoencoder consists of two one-dimensional convolutional layers with ReLU activation (filters: 132 and 64), followed by a dropout layer (rate = 0.1), and a transposed convolution layer used for upsampling and reconstruction. The latent representation learned by the encoder serves as a positional embedding that is added to the input signal before entering the Transformer encoder blocks (Figure 4). This design enables the model to capture both local temporal structure and global positional context, improving robustness to noise and signal variability commonly observed in real-world PSG recordings. Autoencoders come in handy for data denoising, reducing dimensionality, or even learning a better representation of the samples’ distribution. The critical task was to tweak an autoencoder that fits the 1e-3 Constant, Learning rate 1e-4 decay head_size 128, 256 256 num_heads 4, 8 4 num_transformer_blocks 4, 6, 8, 12 6 mlp_units 258, 128 128 dropout = 0.1 None, 0.1, 0.5 0.1 Where pos is the position and i is the dimension. 3.3.3 Sinusoidal positional encoding Sinusoidal positional encoding employs sine and cosine functions of different frequencies to represent each sequence position as a vector of size dmodel (Equation 3): PE( pos ,2i ) = sin pos / 10000 2i / d model ( ) ( ) PE( pos ,2i +1) = cos pos / 10000 2i / d model (3) Where pos is the position and i is the dimension. 2i and 2i + 1 are used to alternate between even and odd sequences. We experimented with different lengths and depths of the sinusoidal embedding. A length of 64 and a depth (dmodel) of 32 provided the most suitable Frontiers in Artificial Intelligence 08 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 TABLE 3 Confusion matrix. PR/GT Predicted (PR) True positive (TP) False positive (FP) False negative (FN) True negative (TN) used for training and data from 7 patients were used for validation, with no patient appearing in both sets. This protocol prevents data leakage across folds and ensures that all reported results are generalized to unseen subjects. We conducted two sets of experiments: first, evaluating the impact of three different positional encoding strategies, and second, investigating training acceleration through weight decay (Loshchilov and Hutter, 2019). The first set of experiments looks at the effect of different positional embeddings with the base Transformers model on the apnea classification task. To overcome the skewed distribution of the classes in this dataset, different weights were assigned to both the majority and minority classes; this influenced the training to be fair for both classes. Let TP denote samples correctly identified as apnea, and TN denote samples correctly identified as normal. The evaluation metrics computed from the confusion matrix (Table 3) include accuracy, recall (sensitivity), specificity, and F1-score. Accuracy: measures the proportion of correct predictions made by the model. Since the used dataset is imbalanced, which is common in problems of similar nature, where there are fewer anomalies than normal events, accuracy alone is not sufficient to correctly evaluate the model’s performance. To overcome its limitations, we also took into account recall (Sensitivity), specificity and. F1-Score. We report accuracy, sensitivity (recall), specificity, precision, and F1-score using their standard definitions commonly adopted in the literature. In addition, we considered the area under the receiver operating characteristic (ROC) curve as a complementary evaluation metric, which shows TP rates against FP rate, illustrating the model’s ability to distinguish between the two classes, whereas a random classifier would not exceed an AUC of 0.5. It is worth noting that we were able to speed up the model convergence 0.3x faster, while achieving comparable performance, using AdamW with weight decay equal to 0.0001 (Loshchilov and Hutter, 2019). FIGURE 5 The 64-dimensional positional encoding for a sequence with a maximum depth (dmodel) of 32. Each row represents the embedding vector. job by determining how many layers, different filters are in those layers, and what is the size of the kernel, where kernel size defines the size of the sliding window. In this work, we employ two 1D convolution kernels with ReLU activation and a filter size of 132 and 64, respectively, and a dropout of 0.1 between them. Then a transposed convolution (convt) as the final layer of the autoencoder. Transposed convolution (convt) reverses the standard convolution by dimensions and is usually carried out for upsampling. Convt can be defined to increase the spatial resolution of feature maps in AE (Dumoulin and Visin, 2016). We choose convolutional autoencoders over feedforward autoencoders, as the convolution layer is better for capturing spatial information (Zhang, 2015). 4 Experimental setting Guided by Occam’s Razor, we started with a simple design and gradually increased complexity, conducting experiments that produced a model performing effectively on a real-world dataset of approximately 1 million samples. More specifically, we started the hyperparameter tuning through a random search through K-fold cross-validation on the training samples. Cross-validation (CV) is a technique for evaluating machine learning models. In k-fold CV, the dataset is split into k roughly equal folds, with one fold used for validation and the remaining k-1 folds for training. This process is repeated k times, reducing variance by utilizing the entire dataset for both training and validation. Nevertheless, it requires a higher computational cost and takes more time because the model needs to be trained K times at the validation step and an additional one at the test phase. In the experiment, we did k = 5 CV. The dataset was first split at the patient level into two disjoint subsets: 80% of the patients were used for model development and 20% were held out as an independent test set for final evaluation. Within the training subset, we performed 5-fold cross-validation at the patient level, such that in each fold, data from 23 patients were Frontiers in Artificial Intelligence Ground truth (GT) 5 Results and discussion We conducted five experiments using a slightly modified version of the model each time. All experiments were conducted on the OSASUD dataset (Bernardini et al., 2022), with results summarized in Table 4. Initially, we used only the encoder component of the Transformer without any positional embedding (model_1) and employed the AMSGrad variant of the Adam optimizer. In the second and third experiments (model_2 and model_3), the sample order for each batch was incorporated using the naive and sinusoidal positional encoding strategies, respectively. These two fixed positional encodings did not noticeably improve the model’s overall 09 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 TABLE 4 Performance on the OSASUD dataset. No. 1 2 3 4 5 6 7 8 9 Model Accuracy Sensitivity Specificity F1 AUC 0.716 0.168 0.824 0.162 0.523 0.769 0.628 0.769 0.471 0.750 0.737 0.107 0.865 0.737 0.525 0.752 0.643 0.752 0.468 0.760 0.7754 0.8903 0.8750 0.8638 0.7928 0.7278 0.8909 0.8590 0.8286 0.7644 0.7428 0.8884 0.8670 0.8404 0.7801 0.7736 0.8848 0.9075 0.8635 0.8276 0.7745 0.9058 0.9277 0.8606 0.8520 ResNet (all patients) (Bernardini et al., 2021) LSTM + CNN (all patients) (Bernardini et al., 2021) ResNet (without validation patients) (Bernardini et al., 2021) LSTM + CNN (without validation patients) (Bernardini et al., 2021) Transformers (encoder + Adam) (amsgrad = True) (model_1) Naïve positional embedding encoder (model_2) Sinusoidal positional embeddings Transformer (model_3) Transformers (encoder + weight decay) (model_4) Transformers (with autoencoder) (model_5) The bold values represent the maximum value for each metric. performance. Nevertheless, we further explored different parameter values for the sinusoidal positional encoding. After that, in the fourth experiment (model_4), we tried to speed up the training process through the use of weight decay, and it produced similar results and converged about 0.3 times faster with slightly better specificity and AUC. Autoencoders have attracted so much research attention; they have long been thought to be a potential avenue for cracking unsupervised learning problems, i.e., learning useful representations without labels. To be more precise, autoencoders are a self-supervised technique, where the targets are generated from the input. To get self-supervised models to learn features, you have to come up with the right combination of autoencoder layers. To achieve that, we started with the simplest autoencoder based on fully-connected layers (Goodfellow et al., 2016), which predicted approximately similar to a random classifier (AUC = 0.501), and classifies all apnea events as normal events (the majority class). After that, we attempted a convolution autoencoder, that contained only 1-D convolution layers. At this phase, the model shows better separability reaching AUC equal to 0.814. Finally, after adding a transposed convolution layer (model_5) the model achieved its best performance, reaching AUC = 0.852 (Figure 6). The superiority of convolutional autoencoder over dense autoencoder was expected; Convolutional shows exceptional performance before on time series encoding, i.e., Rocket (Dempster et al., 2020). In the final experiment (model_5), the static positional encoding was replaced with a learnable positional encoding implemented through a convolutional autoencoder containing a transposed convolution (ConvT) layer, added prior to the encoder components. As shown in Table 4, this configuration achieved the highest performance across all evaluation metrics. It is important to consider the implications of emphasizing different metrics. In medical applications for disease prediction, sensitivity is particularly critical, as it is generally preferable to classify a healthy individual as diseased Frontiers in Artificial Intelligence rather than to misclassify a diseased individual as healthy. Notably, model_5 achieved the highest sensitivity among all evaluated models. Transformer-based approaches have recently been applied to obstructive sleep apnea detection from ECG signals. For instance, Biswas and Abu Yousuf, (2025) proposed a CNN–Transformer model using static positional encoding and segmented inputs. While effective, such designs rely on predefined positional representations and may be less robust when applied to continuous, noisy clinical recordings. In contrast, the proposed framework employs a learnable positional encoding based on a convolutional autoencoder, enabling temporal position information to be inferred directly from raw ECG signals. By avoiding handcrafted feature extraction and operating at a one-second temporal resolution, the model better adapts to real-world data variability and demonstrates improved generalization on the imbalanced OSASUD dataset (Bernardini et al., 2022). When it comes to selecting a dataset that is better suited for Transformer-based models, the Apnea-ECG dataset is useful for benchmarking and consists of continuous single-lead ECG overnight recordings lasting approximately 7–10 h, with apnea annotations at one-minute resolution. OSASUD is a more challenging, realistic, and clinically relevant dataset, with apnea annotations at one-minute resolution. Moreover, OSASUD provides a higher sampling rate, which better exploits the strengths of Transformer architectures in modeling fine-grained temporal dependencies and long-range relationships in dense timeseries data. From the beginning, the proposed Transformer-based models showed performance that was comparable to, and sometimes better than, existing deep learning architectures for OSA classification. Adding the autoencoder helped the model capture temporal dependencies and represent the input data more effectively, improving its understanding of each sample’s position and context 10 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 FIGURE 6 Performance comparison of different AEs as a positional encoding component, AE with a convt layer achieved the best results. FIGURE 7 Transformer-based architecture after adding the autoencoder component (model_5): (a) Training loss, (b) learning rate decreased exponentially after the 10th epoch, (c) ROC curve. Frontiers in Artificial Intelligence 11 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 FIGURE 8 Predicted (PR) and ground truth (GT) events for 4 different samples from the testset, at 1-s intervals. over time. However, In comparison with Bernardini et al. (2021), our model was tested on an independent test set. The AIOSA (Bernardini et al., 2021) results were obtained by dividing the dataset into separate training and validation sets, a practice that may introduce data leakage and consequently cast doubt on the generalizability of the reported outcomes (Goodfellow et al., 2016). Figure 7 shows the variable learning rate and loss for the first 18 epochs. Figure 8, showcases random samples from the test set for TP, TN, FP, and FP. the performance of learnable positioned encoding was better than static positional encoding. In addition, we tried to speed up the training process using AdamW and weight decay. Our proposed scheme, based on the Transformer encoder with a convolutional autoencoder, as a positioning encoding, that contains a convt layer, detects OSA events better than ResNet, LSTM, and CNN encoding (Bernardini et al., 2021); with F1 score equivalent to 0.863 and AUC-ROC equivalent to 0.852. For future work, we plan to test the model on diverse datasets and incorporate additional PSG signals, such as thoracic effort (THO), abdominal effort (ABD), and EEG. Most current deep learning approaches detect apnea from a single lead, usually ECG, which offers limited clinical accuracy; as an expert physician remarked, “It would not be very accurate to depend solely on one lead. In practice, we look to different signals at the same time.” In addition, future work will include the collection of larger, multi-center datasets, external validation on independent cohorts, and comprehensive robustness analyses to strengthen model reliability and generalizability. We also plan to integrate Explainable AI techniques to provide clinicians with interpretable insights into model predictions. However, this is an expected direction 6 Conclusions and future work The goal of this work is to develop an efficient tool to support clinical decision-making for OSA. To this end, we propose a Transformer-based deep learning framework for OSA detection using only ECG, capable of handling noisy waveforms without extensive preprocessing. We first applied the framework to OSASUD (Bernardini et al., 2022) dataset and proved that it outperforms other solutions. Then, we focused on trying different positional encoding, Frontiers in Artificial Intelligence 12 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 Conflict of interest for a new multi-disciplinary area, not to mention the expensive computational equipment that must be acquired to develop a decent DL model for multivariate time series data with a high sampling rate. The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Data availability statement Generative AI statement The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author. The author(s) declared that Generative AI was not used in the creation of this manuscript. Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us. Author contributions MA: Writing – original draft, Writing – review & editing. SA-A: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – review & editing. SI: Supervision, Writing – review & editing. AS: Supervision, Writing – review & editing. AB: Supervision, Writing – review & editing. Publisher’s note All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. Funding The author(s) declared that financial support was not received for this work and/or its publication. References Abreu, M., Fred, A., Valente, J., Wang, C., and Plácido, H. (2020). Morphological autoencoders for apnea detection in respiratory gating radiotherapy. Comput. Methods Prog. Biomed. 195:105675. doi: 10.1016/j.cmpb.2020.105675 Cai, X., Song, S., Hu, J., Zhu, Q., Yang, W., and Hong, J. (2023). Body roundness index improves the predictive value of cardiovascular disease risk in hypertensive patients with obstructive sleep apnea: a cohort study. Clin. Exp. Hypertens. 45. doi: 10.1080/10641963.2023.2259132 Almarshad, M. A., Al-Ahmadi, S., Islam, M. S., BaHammam, A. S., and Soudani, A. (2023). Adoption of transformer neural network to improve the diagnostic performance of oximetry for obstructive sleep apnea. Sensors 23:7924. doi: 10.3390/ s23187924 Chang, H.-C., Wu, H.-T., Huang, P.-C., Ma, H.-P., Lo, Y.-L., and Huang, Y.-H. (2020). Portable sleep apnea syndrome screening and event detection using long short-term memory recurrent neural network. Sensors 20. doi: 10.3390/s20216067 Almarshad, M. A., Islam, M. S., Al-Ahmadi, S., and Bahammam, A. S. (2022). Diagnostic features and potential applications of PPG signal in healthcare: a systematic review. Healthc. 10, 1–28. doi: 10.3390/healthcare10030547 Chang, H. Y., Yeh, C. Y., Te Lee, C., and Lin, C. C. (2020). A sleep apnea detection system based on a one-dimensional deep convolution neural network model using single-lead electrocardiogram. Sensors (Switzerland) 20, 1–15. doi: 10.3390/s20154157 Almutairi, H., Hassan, G. M., and Datta, A. (2021). Classification of obstructive sleep apnoea from single-lead ECG signals using convolutional neural and long short term memory networks. Biomed. Signal Process. Control 69. doi: 10.1016/j. bspc.2021.102906 Choi, S. H., Yoon, H., Kim, H. S., Kim, H. B., Kwon, H. B., Oh, S. M., et al. (2018). Real-time apnea-hypopnea event detection during sleep by convolutional neural networks. Comput. Biol. Med. 100, 123–131. doi: 10.1016/j.compbiomed.2018.06.028 Arslan, S., Ak, B., and Toraman, S. (2019). A deep learning-based decision support system for diagnosis of OSAS using PTT signals. Med. Hypotheses 127, 15–22. doi: 10.1016/j.mehy.2019.03.026 De Falco, I., De Pietro, G., Della, A., Sannino, G., Scafuri, U., and Tarantino, E. (2019). Evolution-based configuration optimization of a deep neural network for the classification of obstructive sleep apnea episodes. Futur. Gener. Comput. Syst. 98, 377–391. doi: 10.1016/j.future.2019.01.049 BaHammam, A. S., and Chee, M. W. (2022). Publicly available Health Research datasets: opportunities and responsibilities. Nat. Sci. Sleep 14, 1709–1712. doi: 10.2147/ nss.s390292 Dempster, A., Petitjean, F., and Webb, G. I. (2020). Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Min. Knowl. Discov. 34, 1454–1495. doi: 10.1007/s10618-020-00701-z Benjafield, A. V., Ayas, N. T., Eastwood, P. R., Heinzer, R., Ip, M. S. M., Morrell, M. J., et al. (2019). Estimation of the global prevalence and burden of obstructive sleep apnoea: a literature-based analysis. Lancet Respir. Med. 7, 687–698. doi: 10.1016/ S2213-2600(19)30198-5 Denker, T. D. (2022). Study of a transformer-inspired data-driven diagnostic algorithm for automatic detection of cardiac arrhythmia in 12-Lead. The University of Manchester (United Kingdom) ProQuest Dissertations & Theses, 30159531. Dey, D., and Chaudhuri, S. (2018). Obstructive sleep apnoea detection using convolutional neural network based deep learning framework. Biomed. Eng. Lett. 8, 95–100. doi: 10.1007/s13534-017-0055-y Bernardini, A., Brunello, A., Gigli, G. L., Montanari, A., and Saccomanno, N. (2022). OSASUD: a dataset of stroke unit recordings for the detection of obstructive sleep apnea syndrome. Sci Data 9, 1–10. doi: 10.1038/s41597-022-01272-y Drzazga, J., and Cyganek, B. (2021). An lstm network for apnea and hypopnea episodes detection in respiratory signals. Sensors 21. doi: 10.3390/s21175858 Bernardini, A., Brunello, A., Luigi, G., Montanari, A., and Saccomanno, N. (2021). AIOSA: an approach to the automatic identification of obstructive sleep apnea events based on deep learning. Artif. Intell. Med. 118:102133. doi: 10.1016/j.artmed.2021.102133 Dumoulin, V., and Visin, F. (2016) “A guide to convolution arithmetic for deep learning,” pp. 1–31. Available online at: http://arxiv.org/abs/1603.07285 Bild, D. E., et al. (2002). Multi-ethnic study of atherosclerosis: objectives and design. Am. J. Epidemiol. 156, 871–878. Erdenebayar, U., Kim, Y. J., Park, J. U., Joo, E. Y., and Lee, K. J. (2019). Deep learning approaches for automatic detection of sleep apnea events from an electrocardiogram. Comput. Methods Prog. Biomed. 180. doi: 10.1016/j.cmpb.2019.105001 Biswas, P., and Abu Yousuf, M. (2025). Leveraging Transformer Models for Accurate Detection of Obstructive Sleep Apnea from Single-Lead ECG Signals. In Proceedings of the 3rd International Conference on Computing Advancements (ICCA ‘24). New York, NY, USA: Association for Computing Machinery, 556–563. doi: 10.1145/3723178.3723252 Frontiers in Artificial Intelligence Faust, O., Barika, R., Shenfield, A., Ciaccio, E. J., and Acharya, U. R. (2021). Accurate detection of sleep apnea with long short-term memory network based on RR interval signals. Knowl.-Based Syst. 212:106591. doi: 10.1016/j.knosys.2020.106591 13 frontiersin.org Almarshad et al. 10.3389/frai.2026.1727091 Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep learning: MIT Press. Sheta, A., Turabieh, H., and Thaher, T. (2021b). “Applied sciences diagnosis of obstructive sleep apnea from ECG signals using machine learning and deep learning classifiers.” Guijarro-berdiñas, B., Hernández-pereira, E., and Peteiro-barral, D. (2012). A mixture of experts for classifying sleep apneas. Expert Syst. Appl. 39, 7084–7092. doi: 10.1016/j. eswa.2012.01.037 Signals, R. (2020). Deep recurrent neural networks for automatic detection of sleep apnea from Single Channel. Sensors 20, 1–20. doi: 10.3390/s20185037 Gupta, R., Pandi-Perumal, and BaHammam, A. S. (2018). Clinical atlas of polysomnography. 1st Editio Edn: Apple Academic Press. Singh, S. A., and Majumder, S. (2019). A novel approach Osa detection using singleLead Ecg Scalogram based on deep neural network. J. Mech. Med. Biol. 19. doi: 10.1142/ S021951941950026X Halani, Varun, “Obstructive Sleep Apnea (OSA).” Available online at: https:// emedicine.medscape.com/article/295807-overview (accessed Feb. 23, 2023). Sun, C., Hong, S., Wang, J., Dong, X., Han, F., and Li, H. (2022). A systematic review of deep learning methods for modeling electrocardiograms during sleep. Physiol. Meas. 43. doi: 10.1088/1361-6579/ac826e Heneghan, C. (2011). St. Vincent’s university hospital / University College Dublin sleep apnea database. Hu, S., Cai, W., Gao, T., and Wang, M. (2022). A hybrid transformer model for obstructive sleep apnea detection based on self-attention mechanism using single-lead ECG. IEEE Trans. Instrum. Meas. 71. doi: 10.1109/TIM.2022.3193169 Taghizadegan, Y., Dabanloo, N. J., Maghooli, K., and Sheikhani, A. (2021). Prediction of obstructive sleep apnea using ensemble of recurrence plot convolutional neural networks (RPCNNs) from polysomnography signals. Med. Hypotheses 154:110659. doi: 10.1016/j.mehy.2021.110659 Ichimaru, Y., and Moody, D. G. B. (1999). Development of the polysomnographic database on CD-ROM. Psychiatry Clin. Neurosci. 53, 175–177. doi: 10.1046/j.1440-1819.1999.00527.x Ismail Fawaz, H., et al. (2019). Classification of ford motor data. Data Min. Knowl. Discov. 33, 917–963. doi: 10.1007/s10618-019-00619-1 Tagluk, M. E., and Sezgin, N. (2011). A new approach for estimation of obstructive sleep apnea syndrome. Expert Syst. Appl. 38, 5346–5351. doi: 10.1016/j. eswa.2010.10.022 Jahrami, H., Husain, W., Trabelsi, K., Penzel, T., Hirshkowitz, M., Razjouyan, J., et al. (2025). Artificial intelligence and sleep medicine II: a scoping review of applications, advancements, and future directions ☆. Sleep Med. Rev. 85:102212. doi: 10.1016/j. smrv.2025.102212 Urtnasan, E., and Lee, J. P. K. (2020). Automatic detection of sleep-disordered breathing events using recurrent neural networks from an electrocardiogram signal. Neural Comput. & Applic. 32, 4733–4742. doi: 10.1007/s00521-018-3833-2 Urtnasan, E., Park, J. U., and Lee, K. J. (2018). Multiclass classification of obstructive sleep apnea/hypopnea based on a convolutional neural network from a single-lead electrocardiogram. Physiol. Meas. 39:065003. doi: 10.1088/1361-6579/aac7b7 Jarchi, D., Andreu-Perez, J., and Kiani, M. (2020). Recognition of patient groups with sleep related disorders using bio-signal processing and deep learning. Sensors 20:2594. doi: 10.3390/s20092594 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., and Gomez, A. N. I. (2017). Attention is all you need. Adv. Neural Inf. Process. Syst, 30. Leino, A., Nikkonen, S., Kainulainen, S., Korkalainen, H., Töyräs, J., Myllymaa, S., et al. (2021). Neural network analysis of nocturnal SpO2 signal enables easy screening of sleep apnea in patients with acute cerebrovascular disease. Sleep Med. 79, 71–78. doi: 10.1016/j.sleep.2020.12.032 Veasey, S. C., and Rosen, I. M. (2019). Obstructive sleep apnea in adults. N. Engl. J. Med. 380, 1442–1449. doi: 10.1056/nejmcp1816152 Verma, O. P., Roy, S., Pandey, S. C., and Mittal, M. (2020) Advancement of machine intelligence in interactive medical image analysis Li, K., Pan, W., Li, Y., Jiang, Q., and Liu, G. (2018). A method to detect sleep apnea based on deep neural network and hidden Markov model using single-lead ECG signal. Neurocomputing 294, 94–101. doi: 10.1016/j.neucom.2018.03.011 Wang, Y. A., and Chen, Y. N. (2020). “What do position embeddings learn? An empirical study of pre-trained language model positional encoding,” EMNLP 2020–2020 Conf. Empir. Methods Nat. Lang. Process. Proc. Conf, vol. 2, pp. 6840–6849 Loshchilov, I., and Hutter, F. (2019). “Decoupled weight decay regularization,” 7th Int. Conf. Learn. Represent. ICLR. Mangrum, J. M., and Dimarco, J. (2014). The evaluation and management of bradycardia. Prim. CARE Rev. 342, 1–19. Wang, L., Lin, Y., and Wang, J. (2019). A RR interval based automated apnea detection approach using residual network. Comput. Methods Prog. Biomed. 176, 93–104. doi: 10.1016/j.cmpb.2019.05.002 Mashrur, F. R., Islam, M. S., Saha, D. K., Islam, S. M. R., and Moni, M. A. (2021). SCNN: Scalogram-based convolutional neural network to detect obstructive sleep apnea using single-lead electrocardiogram signals. Comput. Biol. Med. 134:104532. doi: 10.1016/j.compbiomed.2021.104532 Wang, G., Lu, Y., Cui, L., Lv, T., Florencio, D., and Zhang, C. (2022) “A simple yet effective learnable positional encoding method for improving document transformer model,” Find. Assoc. Comput. Linguist. AACL-IJCNLP, pp. 453–463. Available at: https://aclanthology.org/2022.findings-aacl.42 Mostafa, S. S., Mendonça, F., Ravelo-García, A. G., and Morgado-Dias, F. (2019). A systematic review of detecting sleep apnea using deep learning. Sensors 19. doi: 10.3390/ s19224934 Yue, H., Lin, Y., Wu, Y., Wang, Y., Li, Y., Guo, X., et al (2021). Deep learning for diagnosis and classification of obstructive sleep apnea: A nasal airflow-based multiresolution residual network. Nature and Science of Sleep, 361–373. Mukherjee, D., Dhar, K., Schwenker, F., and Sarkar, R. (2021). Ensemble of Deep Learning Models for sleep apnea detection: an experimental study. Sensors, 1–17. Zarei, A., Beheshti, H., and Asl, B. M. (2022). Detection of sleep apnea using deep neural networks and single-lead ECG signals. Biomed. Signal Process. Control 71. doi: 10.1016/j.bspc.2021.103125 Padovano, D., Martinez-Rodrigo, A., Pastor, J. M., Rieta, J. J., and Alcaraz, R. (2022). On the generalization of sleep apnea detection methods based on heart rate variability and machine learning. IEEE Access 10, 1–1. doi: 10.1109/access.2022.3201911 Zhang, Y. (2015). A better autoencoder for image: Convolutional autoencoder, 1–7. Penzel, T., Moody, G. B., Mark, R. G., Goldberger, A. L., and Peter, J. H. (2000). The apnea-ECG database. Comput. Cardiol., 255–258. Zhang, J., Tang, Z., Gao, J., Lin, L., Liu, Z., Wu, H., et al. (2021). Automatic detection of obstructive sleep apnea events using a deep CNN-LSTM model. Comput. Intell. Neurosci. 2021:5594733. doi: 10.1155/2021/5594733 Quan, S. F., Howard, B. V., Iber, C., Kiley, J. P., Nieto, F. J., O’Connor, G. T., et al. (1997). The sleep heart health study: design, rationale, and methods. Sleep 20, 1077–1085. Zhao, J., Cai, X., Hu, J., Song, S., Zhu, Q., and Shen, D. (2024). J-shaped relationship between weight-adjusted- waist index and cardiovascular disease risk in hypertensive patients with obstructive sleep apnea: a cohort study. Diabetes Metab. Syndr. Obes. 17, 2671–2681. doi: 10.2147/dmso.s469376 Sheta, A., Turabieh, H., Thaher, T., Too, J., Mafarja, M., Hossain, M. S. (2021a). Diagnosis of Obstructive Sleep Apnea from ECG Signals Using Machine Learning and Deep Learning Classifiers. Applied Sciences, 11:6622. doi: 10.3390/app11146622 Frontiers in Artificial Intelligence 14 frontiersin.org