Health Recommender System for Sleep Apnea Using Computational Intelligence

We have explored databases such as PubMed and IEEE-Xplore to find the interaction between the health and biological sciences literature with the computer science and engineering literature. Google Scholar and NCBI were also used to manually screen papers from various journals and conferences based on title and abstract; see Table 1. In this review article, we have included/excluded articles based on some criteria; those are shown in Table 2.

The available literature suggests that the occurrence of SA and factors that contribute to obstructive sleep apnea (OSA) in India are increasing urbanization, a stressful life, and lifestyle changes. India is expected to face an obesity epidemic, which is highly associated with OSA [1]. OSA is the most common breathing disorder worldwide. While obesity remains a major cause of OSA in Asians, factors like age, body mass index (BMI), smoking, and alcohol consumption may contribute more to OSA development; see Figure 1 for types, causes, and consequences of SA. Research highlights the need for public awareness, early diagnosis, and effective treatment to tackle SA [2].

The following research articles show that SA is highly prevalent in India and in other parts of the world. The data set “Sleep-Cohort-Study by Wisconsin” discusses the occurrence of sleep-disordered breathing (SDB) for two distinct time frames in the US: 1988–1994 and 2007–2010. The findings reveal a significant increase in the prevalence of moderateto-severe SA over the past two decades. The relative increases in various subgroups (by age) range from 14 percent to 55 percent [3]. A decade ago, Suri pioneered work on sleep medicine in India, which led to the establishment of multiple sleep centers across the country. The article underscores the need for educational initiatives, local production of affordable sleep analysis tools, and integrating sleep medicine into medical education. The article advocates collaborations with premier engineering institutes like IITMadras for developing cost-effective equipment’s for SA diagnosis [4].

Types of SA: The existing literature suggests three categories of SA: OSA, central sleep apnea (CSA), and mixed/complex sleep apnea; see Figure 2 for more details. Of these, OSA) is most prevalent. Individuals with OSA commonly experience repeated instances of pharyngeal airway narrowing or collapse during sleep, as highlighted by Campana et al. [5].

Figure 2.

SA symptoms, causes, and consequences

OSA is characterized by repeated collapse of the upper airway during sleep, leading to disrupted sleep, hypoxemia, and increased risk of conditions like hypertension and cardiovascular disease (CVD). Figure 3 shows a comparison between various breathing patterns associated with breathing disorders. With a global prevalence of 2 percent to 10 percent, risk factors include age, gender, obesity, genetics, craniofacial anomalies, smoking, and alcohol consumption. Identifiable symptoms include loud snoring, observed apneas, and daytime sleepiness [6].

Figure 3.

Breathing patterns

Sleep-Disordered-Breathing (SDB) Recognized as a significant health issue in young children, SDB has estimated prevalence rates of 1 percent to 4 percent, with contributing factors being a narrow airway and reduced neuromuscular tone. Children with conditions like Down syndrome are at higher risk. SDB symptoms include snoring, frequent arousal, enuresis, and hyperactivity, and if untreated, it can lead to learning challenges, stunted growth, and increased risks of hypertension and cardiovascular problems [7].

Since SA is sleep-related disorder, knowledge of sleep stages is important because it has been observed through the literature that SA occurs when a patient goes into deep sleep. There are two main types of sleep stages: the first one is nonrapid eye movement (NREM) sleep and the second one is rapid eye movement (REM) sleep.

NREM is subdivided into stages N1, N2, and N3. The characteristics of REM sleep are rapid eye movements, muscle atonia, and desynchronized brain activity [8].

Risk Factors of OSA: Obesity: The literature analyzed here suggests that patients with high BMI value is very likely to develop SA. Kandala et al. [9] investigated OSA in 30 participants with snoring history and a high Epworth sleepiness scale scores. Based on BMI, 23 patients were diagnosed with OSA (AHI > 5); out of these, 13 were obese and 10 were nonobese. Obese individuals exhibited lower mean oxygen saturation levels (SpO₂) and experienced reduced total sleep time, sleep efficiency, N3 stage, and REM stage compared to nonobese patients. The same thing is highlighted by Reddy et al. [10]; they have assessed OSA occurrence and contributing factors in an urban Indian population aged 30 to 65, estimating a 9.3 percent population prevalence for OSA and 2.8 percent for OSA syndrome. Male gender, high BMI, and abdominal obesity were associated with OSA, with obesity found to be a major risk factor; they noted associations with hypertension, while the least significant risk factors were smoking and drinking. Their study also highlighted the fact that OSA is a common condition that causes interrupted breathing during sleep, affects various age groups (prevalence: 2 percent to 14 percent), and leads to fragmented sleep and daytime sleepiness. Linked health issues include hypertension and heart disease, with predictive features like snoring and obesity. Diagnosis involves polysomnography (PSG), and the primary treatment is the continuous positive airway pressure (CPAP) machine, with bariatric surgery considered for obese patients.

If SA is left untreated, its consequences can be fatal. Some of the following articles discuss links between SA and other diseases. According to Gurubhagavatula et al. [11], OSA can cause daytime sleepiness and cognitive impairments, affecting vigilance, memory, concentration, and executive function, thus increasing the risk of accidents and diminishing overall quality of life. Treatment of OSA has demonstrated improvements in these areas. Researchers like Martin et al. [12] highlight the impact of OSA during pregnancy, linking it to adverse outcomes such as preeclampsia and gestational diabetes. Fatima et al. [13] emphasize that OSA is prevalent, affecting about 20 percent of US adults, with a higher incidence in obese individuals and males. Home studies reveal that 7.5 percent of middle-aged Indian men have OSA. The connection between OSA and CVD is of interest due to their potential mechanisms, but the impact of OSA treatment on cardiovascular risk remains uncertain.

Many researchers discuss SA and kidney diseases. Anker et al. [14] highlight the common occurrence of SDB in CVD and its impact on outcomes. It discusses the two main SDB types, diagnostic considerations using PSG or portable devices, and effective treatments like CPAP for OSA. The optimal treatment for CSA in heart failure is uncertain, with emerging therapies like phrenic nerve stimulation under exploration. Targa et al. [15], in their study, explored the link between OSA events and their impact on sleep patterns, Alzheimer markers, and cognitive decline in 116 patients (median age: 76, AHI: 25.9). Obstructive apneas were related to sleep disruptions, while hypopneas were linked to increased arousal, and both mixed and central apneas affected sleep structure. At the 12-month follow-up, hypopneas were the most significant predictor of greater cognitive deterioration, and OSA was connected with raised neurofilament light levels. Mavanur et al. [16] showed that SDB is common in advanced chronic kidney disease (CKD), affecting over 50 percent. It leads to upper airway blockage during sleep, causing physiological reactions and increased cardiovascular risks. Treatments like renal transplantation and specialized dialysis methods show promise in reducing SDB severity in CKD patients.

For diagnosis of SA, sleep experts usually monitor ECG [17–21], EEG [22], SpO₂ [23], snoring sound, and various other body parameters of the patient. Sharma et al. [24] emphasized that due to lack of awareness about OSA in India, the Ministry of Health and Family Welfare established INdian initiative on Obstructive sleep apnoea (INOSA) guidelines. These guidelines recommend a sleep assessment, particularly for individuals showing symptoms like snoring and daytime sleepiness or those who were considered high-risk cases. PSG is the standard diagnostic method, and positive airway pressure (PAP) therapy is the primary treatment, with oral appliances and bariatric surgery considered for specific cases.

Table 3 shows variety of data used by researchers across the world to address research questions concerning SA classification. The data components encompass subject demographics, data sets of ECG signal data, EEG signals, and SpO₂ related to SA. Data sets were also obtained through various kinds of wearable sensors like smart watches, pulse oximeters, SA rings, smart body sensors, and nasal cannulas.

A large volume of literature is available that shows how the CPAP machine is effective in SA treatment. CPAP is the first choice of medical experts. Some of the following literature highlighted the use of CPAP. Jané [30] highlighted that standard treatment involves CPAP therapy; their research focuses on adaptive pressure algorithms and emerging trends aimed at enhancing patient engagement through mobile apps and web platforms such m-Health and Tele-Health. Similarly, authors like Senavongse et al. [31] discuss the affordability issue of CPAP machines for treating OSA and present a study designing a functional prototype. The prototype, demonstrating accurate measurements and potential improvements, shows promise for clinical trials in addressing OSA and snoring. Amrulloh et al. [32] presented an on-demand CPAP (OCPAP) controller as an alternative SA treatment, adapting air pressure based on respiratory needs for enhanced comfort and upper airway muscle training. Developed with LabView, the model demonstrated promising results in performance tests, offering potential for a low-cost treatment option, particularly in Indonesia, to reduce dependence on traditional CPAP systems.

Some authors have discussed advances in the CPAP machine and their use in SA treatment. Boisteanu et al. [33] compared automatic CPAP (APAP) to a fixedpressure CPAP in moderate to severe OSA patients; the study suggests that APAP is as effective, offering a slightly lower effective pressure. Intelligent CPAPs with remote-monitoring capabilities can reduce costs and doctor visits, making them a potential choice for long-term treatment. Penzel et al. [34] investigated how OSA affects cardiovascular and respiratory regulation during different sleep stages, focusing on patients with OSA and normal blood pressure, OSA and hypertension, and normal controls.

Researchers have discussed other popular treatment options as well, such as the mandibular advancement device (MAD). Daga et al. [35] assessed sleep quality before and after using an MAD and participating in yoga. While the MAD group showed immediate improvements with compliance issues, the yoga group demonstrated sustained benefits over the long term, indicating the effectiveness of yoga and pranayama practices in long-term management of OSA.

The literature available on SA suggests the use of advanced technological options such as adaptive servo-ventilation (ASV) for CSA treatment. Aurora et al. [36], according to a recent systematic analysis, showed that ASV enhanced left ventricular ejection fraction (LVEF) and normalized the apnea-hypopnea index (AHI) in patients with CSA associated with congestive heart failure. Webster et al. [37] proposed a novel SA treatment device comprising a mask, hose, and CO₂ chamber, that automatically adjusts rebreathed air to reduce apneas without PAP, offering a potentially more effective and comfortable treatment for the over 25 million Americans affected by SA.

In conclusion, the available literature suggests that there are continuous improvements going on to make CPAP, MAD, ASV, and SA monitoring devices better and more patient-friendly. There is future research scope in this field to incorporate artificial intelligence (AI) and advanced ML and DL algorithms to make these devices intelligent, with automated pressure settings.

An ample amount of literature is available on the application of ML in SA diagnosis; see Table 6 for a comparative study. Han et al. [27], in their study, evaluated ML techniques for OSAS severity assessment using demographic and questionnaire data from 313 patients. For classification, random forest and support vector machine (SVM) models performed best, yielding the highest accuracy of 44.7 percent, with misclassification observed in only 5.7 percent of cases. Linear regression and the SVM model performed well in predicting the AHI, with regression models achieving a minimum RMSE = 22.17. Similarly, Alvarez et al. [44] used regression and SVM techniques, discovering that in terms of forecasting the AHI, the dual-channel technique performed better than individual oximetry and airflow, showcasing high complementary value, and significantly improving accuracy for efficient at-home screening of OSA.

Researchers have found that ML algorithms, along with body parameters like SpO₂, can be used for SA diagnosis. Shi et al. [28], in their study involving 1493 OSA patients and 27 variables, including hyper tension, learned that ML algorithms, and particularly, the gradient-boosting machine (GBM), to be the most reliable in predicting hypertension associated with OSA (AUC = 0.873, accuracy = 0.885, sensitivity = 0.713). The identified key variables, including age, minimum arterial oxygen saturation, BMI, and percentage of time with SaO₂ < 90 percent, led to the development of an online tool for clinicians [28]. Researchers like Liu et al. [45] emphasize the importance of detecting non-apnea-related arousals during sleep for assessing sleep quality. The proposed algorithm trained and tested on PSG data utilizing convolutional neural networks (CNNs) and a random-forest module, achieved an Area Under the Precision-Recall Curve (AUPRC) of 0.552.While effective, the method may have limitations for certain patients.

Authors have discussed the use of ML for judging/monitoring treatment provided to SA patients. Mitri et al. [46] discussed the use of CPAP, applying air pressure for conditions like SA and preterm infants, with a focus on anomaly detection aided by ML using the Numenta Anomaly Benchmark (NAB) and Hierarchical Temporal Memory (HTM). An experiment using infant breathing patterns demonstrated effective anomaly prediction, emphasizing the potential of HTM in anomaly detection, though its immaturity limits progress, suggesting the need for future work to solidify its competitiveness in ML research. Fallamnn et al. [47] comprehensively reviewed technological advancements in sleep monitoring, addressing sleep behavior characterization, assessment methodologies, monitoring techniques, and analysis methods within personalized smart health care, emphasizing the potential for data-driven techniques to bridge the gap between clinic-based and home-based sleep assessments. Rao [48] explored wearable sensors for respiratory and pulse monitoring, emphasizing technologies like heart rate monitoring, GPS, GSM, IoT, and infrared-based breath sensors, aiming to enable selfmonitoring of health parameters and improve health care technology through ML techniques and an innovative system leveraging IoT and GSM platforms.

The accuracy of ML methods applied for SA diagnosis can be improved by considering combination of signals instead of individual signals; ECG-SpO₂ [54], ECG-nasal pressure, SpO₂-EEG [44], ECG-EEG, ECGchest-abdomen signal combinations can be tried. Limitations with standalone ML algorithms include the lack of feature extraction/optimal feature selection ability; hence, hybrid combinations ML models with optimization algorithms should be tested [e.g., particle swarm optimization (PSO), genetic algorithms (GAs), blue whale (BLO), and gray wolf (GWO)]. In order to enhance performance measures of ML approaches, researchers have tried hybrid approaches like CNNLSTM, PSO-SVM, and HRV-LDA [52].

DL has shown very promising results in diagnosis and processing of signals received from SA-suspected patients. Feature extraction and optimal feature selection were the key points of DL algorithms like CNNs and recurrent neural networks (RNNs). We have presented a comparative study of various DL methods applied in SA prediction; see Table 7. Sun et al. [55] presented a classification of sleep stages, a two-stage neural network strategy that uses an RNN for temporal input processing and feature learning, along with a pretraining procedure to address sample imbalance. Tests on sleep databases demonstrate superior performance compared to advanced methods, achieving significant F1 scores and Kappa coefficients.

Researchers like Kristiansen et al. [51] have investigated the use of data mining methods, such as decision trees, random forests, SVMs, K-nearest neighbor (KNN), and artificial neural networks to examine physiological signals for the purpose of detecting OSA, utilizing data sets from the MIT-BIH and Apnea-ECG databases [51]. Other authors like Yang et al. [58] utilized a 1D-CNN model on onechannel EEG data with sleep-stage annotations, demonstrating higher accuracy for intrapatient insomnia identification, particularly leveraging REM and SWS epochs, compared to baseline methods, while finding no significant differences in interpatient identification.

Many researchers stressed the use of ensembled ML/DL algorithms as Kwon et al. [59] have; they presented a novel method using IR-UWB radar and DL algorithms for real-time apnea-hypopnea identification in SA and hypopnea syndrome, achieving high performance with a strong association between estimated and reference AHIs. Zarei et al. [29] introduced an automated approach utilizing ECG signals and a combination of CNNs with LSTM networks for SA detection, achieving impressive results with higher sensitivity (94.41 percent), specificity (98.94 percent), and accuracy (97.21 percent) on Apneic and UCDDB data sets. The LSTM-CNN model outperforms traditional methods, providing accurate per-segment and per-recording classifications, thereby enhancing SA diagnosis for physicians.

Researchers have also presented the use of pretrained neural networks in SA diagnosis. Hemrajani et al. [57] presented MobileNet-V1, LSTM, and GRU networks for identifying single-lead ECG signals in undiagnosed OSA cases, achieving accuracy rates of 89.5 percent, 90 percent, and 90.29 percent on authentic cases. Sharma et al. [49] work presented an automated technique that uses pulse-oximeterrecorded SpO₂ and PR data to identify episodes of SA. For epoch-based apnea detection, the DL model obtained a test performance of 90.4 percent area under the ROC curve and a 58.9 percent area under the precision-recall curve after being trained on a heterogeneous cohort of patients.

The IoT, along with health-monitoring devices, has a very wide scope in the field of health care monitoring. Kwon et al. [60] introduced a portable athome solution that uses wearable electronics with embedded ML and wireless sleep sensors to solve the problem of undetected sleep disorders. Clinical testing demonstrates a comparable performance to PSG in capturing brain, eye, and muscle signals. The wearable system accurately identifies OSA with 88.5 percent precision. Steblin et al. [61] introduced an IoT-based solution to improve the treatment of OSA. To assist patients with OSA and provide feedback to lung specialists, the suggested technology transmits patient data to the cloud for analysis.

Many researchers have tried to emphasize the important role of the IoT and ML/DL combinations in SA diagnosis. Abdel-Basit et al. [62] explored the challenges in detecting and treating OSA and highlighted the potential of AI-driven IoT technologies for remote patient monitoring, providing an overview of developments from 2016 to 2019 in big data, cloud computing, ML, smart devices, and fog computing.

Recommender systems (RSs) give recommendations based on user profile. Here, in the case of SA treatment recommendations, authors have proposed various approaches of RSs in collaboration with ML/DL algorithms; see Table 8. Nanehkaran et al. [63] introduced a medical recommendation system utilizing IoT devices, employing KNN classification for disease identification and collaborative filtering for treatment recommendation. The approach shows high precision in diagnosing chronic diseases and recommending treatments, surpassing previous methods. Other authors, like Casal-Guisande et al. [64], presented an intelligent system for diagnosing OSA. The system combines patient health data and symptom information to generate risk indicators for OSA. The early testing of the system showed promise, but further clinical validation and improvements are needed before it can be widely used in hospitals.

We have analyzed research articles on the application of RS in the health care domain utilizing electronic health records (EHRs). Raza et al. [65] introduced a two-stage recommender system for clinical decisionmaking with the help of EHRs. The first stage retrieves candidate items based on patient records using a deep neural network and a language model. The second stage ranks and recommends relevant items considering patient history and context. Pinion et al. [66] proposed a federated learning architecture for health recommender systems (HRSs) in precision medicine, addressing privacy concerns and enabling a real federated HRS without compromising confidentiality. In this research, they evaluated an HRS developed for the TeNDER-project, which provides personalized recommendations based on monitoring device data. The notifications covered various aspects of daily life.

However, other authors, like del Rio et al. [67], have proposed a recommendation system that suggests a healthy lifestyle schedule to mitigate SA severity. A probabilistic Markov model (PMM) guides activities based on patient time allocation, aiming to reduce apnea cycles and improve sleep patterns. The system uses a hidden Markov model for condition-directed recommendations, focusing on flexibility and user preferences.

Figure 4 shows a proposed HRS. It involves processes like feature engineering and selection, initial stages of HRSs we need to identify, and relevant features (e.g., SpO₂, blood pressure, sleep habits, patient demographics such as age, gender, BMI, and neck circumference. Data should be preprocessed (i.e., cleaned and normalized) before feeding into an ML/DL algorithm. Before algorithm selection, we need to investigate various ML algorithms. HRSs may require DL techniques (CNNs and RNNs) for better prediction results. There is a need to optimize hyperparameters (learning rate, regularization). Ensemble methods can be used by combining predictions from multiple models. Performance evaluation metrics such as accuracy and recall can be considered, and cross-validation methods like k-fold cross-validation can be used to improve HRS performance. Domainspecific knowledge, such as collaboration with sleep medicine specialists, is essential. Understanding clinical implications is also very important.

Figure 4.

Workflow for an HRS for SA

To demonstrate its efficiency and effectiveness in the health care industry, the study [71] presents a Healthcare Monitoring System (HMS) that combines IoT and ML technologies. It does this by using wearable sensors for real-time monitoring and a medical decision support system for the detection and analysis of health issues.

Researchers across the world have explored the field of AI for SA diagnosis and the role of AI in assisting in SA treatment. In line with this, Kaneria et al. [68] developed an automated DL method called DOSED to detect sleep-breathing events in PSG recordings, which are used to diagnose OSA. The performance of the method was compared to the precision of human sleep experts in diagnosing the severity of OSA and detecting individual breathing events. Furthermore, Thorey et al. [72] reviewed publications from 1999 to 2022 to explore AI’s role in improving OSA treatment. AI can predict treatment outcomes, evaluate current treatment effectiveness, and enhance understanding of OSA mechanisms. Strumpf et al. [56] have designed Belun-Ring and tested its performance, involving 84 participants and comparing the Belun-Ring results with in-lab PSG. They found that the Belun Ring with BSP2 algorithms accurately detected OSA, classified its severity, and classified sleep stages.

SA detection using ML and DL approaches is strongly reliant on high-quality data. Researchers face numerous issues linked to data availability, privacy, and secrecy. Here are the main issues:

• 1)

  Obtaining labeled SA data is challenging owing to the specialized nature of the study and the necessity for expert comments.

• 2)

  Researchers frequently use publicly available data sets to construct and test their models. These databases are critical to developing SA research.

• 3)

  Data privacy and confidentiality. Sleep-related information, including physiological signals (electroencephalography, electrocardiography, and breathing patterns), is sensitive and personal. Maintaining patient privacy and adhering to ethical rules are critical. Researchers must keep data anonymous and secure.

• 4)

  National Sleep Research Repository (NSRR). The NSRR is an excellent resource for SA researchers. It hosts a variety of sleep problem data sets, including SA. Researchers can examine many sorts of data, including PSG recordings, actigraphy data, and clinical information.

Table 9 contains information about specific data sets important for SA research.