Integrating Phenotypic Information of Obstructive Sleep Apnea and Deep Representation of Sleep-Event Sequences for Cardiovascular Risk Prediction

Integrating Phenotypic Information of Obstructive Sleep Apnea and Deep Representation of Sleep-Event Sequences for Cardiovascular Risk Prediction | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Integrating Phenotypic Information of Obstructive Sleep Apnea and Deep Representation of Sleep-Event Sequences for Cardiovascular Risk Prediction Yali Zheng, Zhengbi Song, Bo Cheng, Xiao Peng, Yu Huang, Min Min This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4084889/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 16 Mar, 2026 Read the published version in BMC Medical Informatics and Decision Making → Version 1 posted 9 You are reading this latest preprint version Abstract Background : Advances in mobile, wearable and machine learning (ML) technologies for gathering and analyzing long-term health data have opened up new possibilities for predicting and preventing cardiovascular diseases (CVDs). Meanwhile, the association between obstructive sleep apnea (OSA) and CV risk has been well-recognized. This study seeks to explore effective strategies of incorporating OSA phenotypic information and overnight physiological information for precise CV risk prediction in the general population. Methods : 1,874 participants without a history of CVDs from the MESA dataset were included for the 5-year CV risk prediction. Four OSA phenotypes were first identified by the K-mean clustering based on static polysomnographic (PSG) features. Then several phenotype-agnostic and phenotype-specific ML models, along with deep learning (DL) models that integrate deep representations of overnight sleep-event feature sequences, were built for CV risk prediction. Finally, feature importance analysis was conducted by calculating SHapley Additive exPlanations (SHAP) values for all features across the four phenotypes to provide model interpretability. Results : All ML models showed improved performance after incorporating the OSA phenotypic information. The DL model trained with the proposed phenotype-contrastive training strategy performed the best, achieving an area under the Receiver Operating Characteristic (ROC) curve of 0.877. Moreover, PSG and FOOD FREQUENCY features were recognized as significant CV risk factors across all phenotypes, with each phenotype emphasizing unique features. Conclusion : Models that are aware of OSA phenotypes are preferred, and lifestyle factors should be a greater focus for precise CV prevention and risk management in the general population. Cardiovascular risk prediction obstructive sleep apnea phenotyping deep representation sleep event sequences phenotype-aware models model interpretability Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. INTRODUCTION Cardiovascular diseases (CVD) are the leading cause of death globally. According to the World Health Organization, the annual deaths from CVDs exceed 17 million, accounting for approximately 32% of all mortalities [1]. Risk prediction is crucial for the prevention and treatment of CVDs. Traditionally, medical experts make qualitative predictions based on experience, supplemented by statistical models such as Framingham Risk Score [2], SCORE [3], and Cox Proportional Hazard [4]. Recently, Machine Learning (ML) and Deep Learning (DL) techniques have emerged as powerful tools for enhancing CVD risk prediction [5-11]. Despite these advances, many studies have not fully considered the complex interplay of CVD risk factors with concurrent clinical conditions. There is a recognized link between Obstructive Sleep Apnea (OSA) and increased CVD risk, with OSA prevalence in CVD patients ranging from 40% to 80% [12]. Recent research has focused on OSA phenotyping using polysomnography (PSG) data, identifying distinct CVD risk levels among different OSA phenotypes [13-17]. The pathophysiology of OSA-related cardiovascular complications has been recognized as multifaceted, encompassing four main domains: sleep architecture disturbances, autonomic dysregulation, breathing disturbances, and hypoxemia [18-21]. However, existing research have not integrated the OSA phenotypic information with traditional risk factors in predictive modeling, leaving the actual predictive effectiveness uncertain. On the other hand, the increasing ease of gathering long-term physiological data through wearable and mobile devices, combined with the advanced capabilities of deep learning (DL) in complex signal representation, has spurred studies integrating these signals into CV risk prediction. For example, Dami et al. developed a DL model to predict arterial events over a course of a few weeks/months, which outperformed traditional machine learning (ML) and other DL approaches [22]. They applied a Deep Belief Network (DBN) to select effective features from 5-minute ECG recordings, and then made predictions with a long short-term memory (LSTM) network in conjunction with features from the electronic health records (EHRs). Chen et al. employed LSTM to learn deep representations from multi-channel physiological signals. By combining these with static features, they then used a Gradient Boosting Machine (GBM) to predict six types of adverse events in the next 5 minutes of surgery with promising results [23]. Yet, existing studies [24-27] have mainly focused on the prediction of specific CV events and over relatively short period of time. There is limited research about utilizing extended periods of physiological signals for long-term CV risk assessment. Considering the advancements in wearable and mobile technologies, the integration of EHR information and daily health data for long-term CVD risk prediction holds significant potential benefits for cardiovascular healthcare. Therefore, this study aims to explore effective methods to incorporate OSA phenotypic information and leverage extended periods of physiological data for CV risk prediction over long-term periods. And, a feature importance analysis is conducted to identify the most significant risk features in the general population and across different OSA phenotypes to provide insights for precise CV risk management. The main contribution of this study is summarized as follows: 1. Proposed a novel and effective method for precise long-term CV risk prediction, by integrating deep features learned from overnight physiological sequences and employing an OSA-phenotype-contrastive training strategy; 2. Identified lifestyle-related features as significant CV risk factors among a comprehensive feature set, and revealed distinct risk features across different OSA phenotypes, offering new insights for precise management of CV risks. 2. METHODS 2.1 Datasets and Subjects This study utilized data from the Multi-Ethnic Study of Atherosclerosis (MESA), initiated and collected by the National Heart, Lung, and Blood Institute (NHLBI) [ 28 , 29 ]. MESA study recruited a total of 6,814 participants without CVDs aged between 45 to 84 years from four ethnic groups (African Americans, Chinese Americans, Hispanics, and Caucasians). Five examinations were conducted between 2000 and 2011, and clinical outcomes were assessed every 9 to 12 months during this period, including: myocardial infarction, angina, heart failure, coronary heart disease, and death. During Exam_5, 2,237 participants of the MESA population participated in an auxiliary sleep study, and 2,489 static features were recorded, including demographics, anthropometrics, medication usage, medical history, imaging risk factors, etc. Among them, 1,874 participants had provided overnight PSG data, which includes 27 continuous physiological signals recorded during sleep, including electrocardiography, electroencephalography, pulse, nasal airflow, blood oxygen, periodic limb movements (PLMS), and more. The average recording duration was 12 hours per participant. Additionally, 615 sleep-related static features were extracted from the sleep questionnaires, actigraphy-derived parameters and average counts of overnight sleep events derived from PSG recordings (so there are in total of 3,104 static features). Manual annotations were performed on the PSG data for identifying sleep stages and events (such as arousal, hypopnea, apnea, hypoxemia, periodic limb movements). This study mainly focused on the analysis of the 1,874 participants, who have complete sets of static features and PSG recordings, and 175 of them (9.3%) experienced CVD events. The whole dataset was divided into training and validation sets, comprising 1,687 and 187 subjects, respectively. The validation set proportion is 9.3%. 2.2 Study Pipeline Initially, feature selection was performed on all static features by lasso-logistic regression. Clustering was then conducted based on the overnight average values of PSG features (named PSG static features) to identify OSA phenotypes. Subsequently, several ML and DL models with different feature selection and training strategies were built to explore the most effective method of incorporating OSA phenotypic information for CVD risk prediction. Finally, the feature importance analysis was conducted for each phenotype based on the best performing model. The study pipeline is shown in Fig.1. 2.2.1 Data Preprocessing The following preprocessing steps were applied to the 3,104 static features (2,489 exam_5 static features and 615 PSG static features) of all participants. First, data outliers were removed, including blank values, duplicates, and irrelevant numerical values. Missing values were then imputed with mean interpolation. Finally, z-score normalization was applied to standardize each feature. The preprocessing process resulted in each participant having 1,600 normalized static features. Considering our goal is to predict 5-year CVD risk, while multiple high-fidelity physiological signals exhibit highly complex temporal patterns throughout the entire sleep period, directly using them as model inputs may not capture efficient and effective information related to long-term CVD risks. Therefore, this study chose to learn deep representations from the feature sequences of five sleep events known contributed to CVD risks, including arousal, hypopnea, apnea, hypoxemia and PLMS. The sequences of the five sleep events were generated according to the PSG labels through the following steps. Firstly, sleep sequences were generated with a 5-second sampling interval during the start and end sleeping time of each participant. Then, the time slots of the occurrences of each sleep event were extracted from the PSG labels to determine whether the event occurred within any of the 5-second sampling intervals. These events were marked as '1' if they occurred within an interval, and '0' otherwise. This created five overnight sleep-event sequences for further analysis. (Fig.2) illustrates the generation process of the overnight feature sequences of five sleep events. 2.2.2 Static feature selection Lasso logistic regression was used to select a subset of features from the 1,600 static features using the training set. Ten-fold cross validation was conducted to determine the regularization hyperparameter, alpha. It was adjusted across a range of values: 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.015, 0.02 and 0.025, to identify its optimal value that yields the best predictive performance on the validation set. 2.2.3 OSA clustering 29 PSG static features (the features shown in Supplementary Table 2) were employed for OSA clustering utilizing the K-means clustering. A range of 2 to 6 clusters was explored to determine the optimal cluster number by the silhouette and elbow methods. Mann-Whitney U tests were then conducted on the 29 PSG static features between different OSA phenotypes to identify the most relevant PSG features for each phenotype. A significance level below 0.05 was considered as significantly different, while a P value below 0.001 was considered highly significant. Cox proportional hazards regression analysis was also performed to estimate the hazard ratio (HR) and 95% confidence intervals (CI) of occurring CVD events within five years for different OSA phenotypes. 2.2.4 OSA phenotyping-based CVD risk prediction modeling To assess the value of incorporating OSA phenotypic information in CVD risk prediction, several classic ML models were employed in this study, including logistic regression (LR), support vector machine (SVM), decision tree (DT), random forest (RF), gradient boosting machine (GBM) and multilayer perceptron (MLP). The phenotype-agnostic method did not consider any phenotypic information. For phenotype-specific models, two different strategies to integrate the phenotypic information and three different feature sets were evaluated. The details of the phenotype-specific ML models are described as below: 1. Pheno_fuse_ML : The OSA phenotyping labels (Phenotypes 1, 2, 3, 4) were incorporated as a new feature along with selected static features, serving as input for the ML models to predict CVD risk across the entire population. 2. Pheno_specific_ML: The strategy develops CVD risk prediction models specific to each phenotypic population, using the three different sets of features: (1) Pheno_Spec_ML1 : Uses only selected static features. (2) Pheno_Spec_ML2 : Combines the 29 PSG static features with the selected static features. (3) Pheno_Spec_ML3 ：Fuses phenotype-specific static PSG features representing each phenotype with the selected static features. To further explore the value of overnight sleep-event feature sequences in CVD risk prediction, a two-layer LSTM network was employed to learn deep representations from the overnight feature sequences, which were then combined with the static features in a fully connected layer to predict CVD risk. We further proposed a phenotype-contrastive training strategy, i.e., the Contrast_pheno_DL , to enhance the model performance. To validate the effectiveness of the strategy, two other phenotype-specific DL models with different feature sets were implemented for comparison, i.e., Pheno_spec_DL1 and Pheno_spec_DL2 . The DL model architecture with different training strategies and feature sets are illustrated in (Fig.3), and the details of the three DL models are described below: 3. Pheno_spec_DL1 : The feature set of this model includes all selected static features, the 29 PSG static features, and deep representations of the five sleep-event feature sequences. 4. Pheno_spec_DL2 : The feature set comprises the selected static features, the 29 PSG static features and sleep-event feature sequences that are specifically relevant to each phenotype. The aim is to focus on features that are directly related to each phenotype. 5. Contrast_pheno_DL : This model utilizes the same features as the Pheno_spec_DL1 , while its training approach is distinct in strategically merging different phenotypic populations that have distinct risk levels. The goal is to identify and learn discriminative features among various phenotypes. 2.2.5 Performance Evaluation In this study, subjects who experienced cardiovascular events were labeled as positive samples. The evaluation metrics include: Accuracy, Precision, Recall, F1-Score, Area under Curve of the Receiver Operating Characteristic Curve (AUC-ROC), Area under Curve of the Precision-Recall Curve (AUC-PRC). 2.2.6 Feature Importance Analysis The influence of all features on the model's predictive performance was evaluated by calculating the SHapley Additive exPlanation (SHAP) values derived from the field of cooperative game theory [ 30 ], to identify the key features for different phenotypes. The SHAP value of each feature was calculated according to the following equation: 3. RESULTS 3.1. Feature selection 51 features were selected by the lasso logistic regression with the alpha parameter set to 0.015. These features encompass not only conventional CVD risk factors like smoking, high-density lipoprotein cholesterol, and total cholesterol, but also extend to features from computed tomography (CT) and magnetic resonance imaging (MRI) ones. Additionally, they include a broad spectrum of other characteristics such as sleep patterns, anthropometry, dietary habits, cognitive status, and more. A detailed list of the selected features can be found in the Supplementary Table 1. 3.2 OSA phenotypes 3 and 6 OSA phenotypes were identified by the silhouette method and 4 by the elbow method on s all subjects, respectively. For 3 phenotypes, there was insufficient differentiation among the four distinct aspects of OSA pathophysiology, i.e., sleep architecture disturbances, autonomic dysregulation, respiratory disturbances, and hypoxemia [ 18-21 ]. In contrast, the division of 6 phenotypes was excessively granular, resulting in a lack of clear distinction between phenotypes. Categorizing these 29 features into 4 phenotypes effectively reflects various aspects of OSA pathophysiology, as shown in Supplementary Table 2. In each phenotype, the most prominent PSG features are highlighted in bold. Each phenotype was named according to its dominant pathophysiological domain: Mild, Respiratory related, Sleep related, and Combined. Mild Phenotype : Characterized by the lowest positive sample proportion of 6.7%, this phenotype exhibits the fewest respiratory events, highest sleep efficiency, near-normal sleep structure, and nighttime oxygen saturation levels (SpO 2 ). It typically comprises healthy (Apnea Hypopnea Index, AHI < 5) to mild OSA (5 ≤AHI < 15) subjects. Sleep Related Phenotype: With the highest positive sample proportion of 15.1%, this phenotype presents significantly lower sleep efficiency, higher arousal and PLMS compared to the Mild type ( p <0.001). It falls under moderate OSA (15 ≤ AHI < 30). The features of breathing disturbance and hypoxemia are less prominent compared to the Respiratory and Combined phenotypes. Respiratory Related Phenotype: This phenotype is characterized by significantly more frequent respiratory events, lower levels of nighttime SpO 2 compared to the Mild type ( p <0.001) and with moderate OSA. The features of sleep architecture disturbance and autonomic dysregulation are less prominent compared to the Sleep related and Combined phenotypes. Combined Phenotype : This type exhibits a mixture of the four pathophysiological domains—sleep architecture disturbance, autonomic dysregulation, breathing disturbance, and hypoxemia. It is classified as severe OSA (AHI ≥ 30). As shown in (Table 1), the HRs for CVD events significantly differed among the four OSA phenotypes. Compared to the Mild phenotype, the Respiratory , Sleep , and Combined phenotypes show 2.708, 2.651, and 2.849 times higher risks to experience CVD events. Notably, the HRs, as determined by OSA phenotyping, does not align with the actual CVD event rates observed within each phenotype. This discrepancy suggests that additional factors are necessary to achieve more accurate predictions of CVD risk. 3.3 OSA Phenotyping based CVD Risk Prediction As illustrated in (Fig.4), the AUC-ROC values of various ML models were notably improved by incorporating OSA phenotyping for CVD risk prediction. Among all models, the MLP model achieved the best performance, with am AUC-ROC value of 0.656. Pheno_spec_ML3 presented the most significant improvement with the AUC-ROC value improved to 0.746. These findings underscore the importance of OSA phenotypic information in CVD risk prediction. A detailed comparison of model performance of various ML models with and without OSA clustering is shown in Table 3 in Supplementary material. 3.4 Integrating deep representation of feature sequences of sleep events Table 2 shows the performance of three DL modeling approaches that incorporated OSA phenotypic information and deep representations of sleep-event feature sequences. There is a notable improvement for the two phenotype-specific models ( Pheno_spec_DL1 and Pheno_spec_DL2 ) when deep features were included as compared to Pheno_spec_ML3 , with the AUC-ROC and AUC-PRC values exceeding 0.83 and 0.50, respectively. The Contrast_pheno_DL model, which involves contrastive training with a combination of the Mild phenotype and one of the other phenotypes (i.e., Mild + Sleep , Mild + Respiratory , Mild + Combined ), achieved even higher AUC-ROC and AUC-PRC values of 0.877 and 0.689, respectively. These findings highlight the substantial additional value provided by overnight sleep-event feature sequences in predicting CVD risk across different OSA phenotypes. Table 1 Cox hazard ratios for CVD events of four OSA phenotypes Parameters OSA phenotypes HR 95% CI P -value Mild Reference Respiratory related 2.708 1.913-3.888 0.003 Combined 2.849 1.566-6.424 0.008 Sleep related 2.651 1.901-4.714 0.001 Table 2 Results of different CVD risk prediction models based on the MESA dataset Related studies Model Number of subjects (Proportion of positive subjects, %) Research Objectives AUC-ROC MES-C risk score [ 31 ] 632（8.2%） Predictive value of atherogenic calcium score for 10-year risk of CVD 0.77 PCP-HFCKD risk equation [ 32 ] 2328（14.6%） Patients with chronic kidney disease_10-year Heart Failure prediction 0.79 DeepSurv [ 33 ] 6814（-） One-year Risk Prediction for Atherosclerotic Cardiovascular Disease 0.82 Our study Model Accuracy Precision Recall F1-score AUC-PRC AUC-ROC Pheno_spec_ML3 0.933 0.736 0.446 0.555 0.496 0.746 Pheno_spec_DL1 0.920 0.579 0.611 0.595 0.507 0.832 Pheno_spec_DL2 0.943 0.686 0.608 0.645 0.573 0.839 Contrast_pheno_DL 0.966 0.851 0.750 0.797 0.689 0.877 Table 3 provides a detailed illustration of the performance of the DL models trained with specific ( Pheno_spec_DL2 ) and contrastive phenotypes ( Contrast_pheno_DL ) across the four phenotypes. It reveals that combining the Mild phenotype with one of the other phenotypes generally improves model performance compared to the Mild phenotype alone across all phenotypes. Grouping the Mild phenotype with the Combined phenotype results in the most substantial improvement for the Mild Type. Table 3 The performance of the DL models across the four phenotypes trained with specific ( pheno_spec_DL2 ) and contrastive phenotypes ( contrast_pheno_DL ) AUC-ROC /AUC-PRC Pheno_spec_DL2 Contrast_pheno_DL Mild + Sleep Mild + Respiratory Mild + Combined Mild 0.750/0.534 0.785/0.590 0.833/0.619 0.833/0.689 Sleep related 0.915/0.808 0.915/0.810 - - Respiratory related 0.732/0.350 - 0.854/0.567 - Combined 0.958/0.600 - - 0.979/0.750 3.5 Feature importance analysis As shown in Fig.5, in addition to the traditional CVD risk features, PSG and FOOD FREQUENCY were recognized as very important features for all four phenotypes. Moreover, each phenotype placed emphasis on different additional features. Specifically, the Sleep Related phenotype particularly emphasized features reflecting features related to sleep status acqruied in the sleep questionnaire such as sleep duration, sleep efficiency standard deviation, etc. The Combined phenotype gave significant emphasis on CARDIAC CT features. Although all four phenotypes emphasized the PSG and FOOD FREQUENCY feature categories, they presented varying importance to specific PSG and FOOD FREQUENCY features. As shown in Fig.6 (a), the Mild phenotype gave more emphasis on breathing disturbance features such as AHI. The Sleep Related phenotype focused on features related to sleep structure disorders, such as sleep efficiency and sleep duration. The Respiratory phenotype strongly emphasized blood oxygen saturation and AHI levels. The Combined phenotype specifically highlighted autonomic disorders features such as arousal and PLMS levels. As shown in Fig.6 (b), four phenotypes gave different emphasis to distinct FOOD FREQUENCY categories. Overall, the frequency of sweet food consumption is a relatively important risk factor for all phenotypes, especially for Sleep and Combined phenotypes. These two phenotypes all gave additional significant importance to fruits. Sleep and Combined phenotypic populations should pay special attention to sugar-controlled diets, and have more fruit intake. On the other hand, the importance distribution of foods in the Mild and Respiratory Related phenotypes is relatively similar, and the two groups should reduce sweets consumption and combine more grains and fruits intake. Fig.7 further shows the top-five important foods on cardiovascular risk of the four phenotypes, which can serve as recommendations for daily dietary management. 4. DISCUSSION Recent research has examined the association between OSA and CVDs, including phenotyping OSA using PSG static features and analyzing their differences in cardiovascular risks [ 13 , 17 , 34 ]. However, it remains to be explored on how to effectively using the OSA-related information for accurate CVD risk prediction modeling. Addressing this gap, this study built several ML and DL models under various OSA phenotyping integration strategies. Additionally, the study examined the value of integrating deep representations of sleep-event feature sequences on CVD risk prediction. The findings indicate that the approach based on OSA phenotyping and integrating deep representations of overnight sleep-event feature sequences, yield the most optimal performance in CVD risk prediction. Regardless of the ML models employed, it is evident that performance improves significantly after incorporating phenotyping. Specifically, while the Pheno_fused_ML model, which integrates OSA phenotyping directly as a feature, shows some effectiveness. But, it is surpassed by the four phenotypes that were modeled separately ( Pheno_Spec_ML1, Pheno_Spec_ML2 and Pheno_Spec_ML3 ). This distinction could be attributed to the inability of Pheno_fused_ML model to adequately learn the weights of the phenotype feature amidst a multitude of other features. Furthermore, selecting phenotype-specific PSG static features to build the model further enhances the model performance ( Pheno_Spec_ML2 vs Pheno_Spec_ML3 ), a trend that was also observed in DL models ( Pheno_spec_DL2 vs Pheno_spec_DL1 ). The possible reason for this improved performance is that, features relevant to other phenotypes may introduce confounding factors in the risk prediction for a specific phenotype. By selecting PSG features tailored to specific phenotypes, models can more accurately learn and understand the relationship between these features and CVD risk within the context of that phenotype. The risk prediction accuracy was further improved by adding the deep features from sleep-event sequences. In addition, we also found that the Contrast_spec_DL model achieved optimal performance by strategically combining one of the three OSA phenotypes ( Sleep-related, Breathing-related and Combined ) with the Mild phenotype, using all PSG static features and sleep event feature sequences as inputs, compared to a training strategy (i.e., Pheno_spec_DL2 ) that incorporated phenotype-specific sequence features. This approach led to a notable increase in the AUC-ROC value to 0.877, and improved predictive accuracy for each phenotype compared to their individual performances. This finding indicates that while modeling within a single phenotype generally yields better results than non-phenotyping approaches, strategic combination of samples from different phenotypes of different risk levels can enhance risk prediction for both groups. A possible explanation for this improvement is the relatively lower and less varied risk profiles within the Mild phenotype. When these were combined with a phenotype characterized by a broader variance in risk distribution, the model's capacity to discern relevant risk features was enhanced. Consequently, this study observed a significant improvement in the predictive performance for the Mild phenotype, especially when combined with the Combined phenotype, as indicated by a HR ranging from 1.566 to 6.424. While previous studies have utilized ML or DL techniques to process short-term physiological sequences, focusing primarily on predicting specific CVD events [ 33 , 38 – 41 ], this study first leveraged the temporal information from overnight sleep-event feature sequences through deep LSTM networks. It provides a more comprehensive set of features for CVD risk prediction than only using averaged static features. Given the objective to predict CVD risk over the next five years, the study recognized the highly complex temporal relationships exhibited by multiple physiological signals throughout the entire sleep period. Directly using these signals as model inputs might fail to capture essential information pertinent to long-term CVD risk. Consequently, the study focused on extracting temporal features from five key sleep-event feature sequences (arousal, hypopnea, apnea, hypoxemia and PLMS) over the full night. Unlike previous studies that relied on manual selection of relevant risk factors for predictive modeling [ 35 – 37 ], this study employed a comprehensive feature selection process across all static features encompassing multi-dimensional aspects, including traditional risks factors, PSG-based multifaced factors, imaging markers and lifestyle factors, etc. This approach has the potential to lead to more integrated and precise predictions of CVD risk in clinical practice. Moreover, as far as we know, this is the first study to rank the importance of features among a comprehensive categories of cardiovascular risk features. One important finding is that, the four OSA phenotypes all gave particular emphasis on PSG and FOOD FREQUENCY features. This is consistent with existing understanding on the importance of PSG features for CVD risk prediction in the literature [ 42 – 44 ]. Earlier studies also highlighted the close association between dietary habits and CVD risk [ 45 – 48 ]. Through the comprehensive analysis of feature importance in this study, the importance of dietary habits in predicting CVD risk should be further emphasized in the general population. Additionally, each phenotype emphasized distinct PSG and FOOD FREQUENCY features. These findings would enable more precisive CVD risk evaluation and management for different OSA phenotypes. Despite these promising results, the study has some limitations. It utilized data from only 1,874 participants from the MESA dataset, and only 187 subjects were used in the testing set. The generalization capability of the model needs further validation on larger datasets. Additionally, the model depended on sleep-event annotation based on the PSG data, which was manually labeled, a resource-intensive and time-consuming process. However, considering ongoing research efforts in developing automated sleep staging and event detection algorithms [ 49 , 50 ], future work could leverage fully automated algorithms to provide the input features required for this model. Future studies may also investigate the minimally effective set of features that maintains the performance of CVD risk prediction with the strategies proposed in this study. 5. CONCLUSIONS This study introduces a new approach for CVD risk prediction, which integrates deep features learned from overnight sleep-event feature sequences. Additionally, the study validated the superior performance of OSA phenotype-specific models over phenotype-agnostic ones, and also introduces a training method that strategically combines the Mild phenotype with another OSA phenotypic population ( Respiratory , Sleep or Combined ) for contrastive training. The contrastive phenotype-specific model with deep features achieved an accuracy of 96.6% and an AUC-ROC value of 87.7% in predicting CVD risk over five years in general population without historical CVDs. Moreover, the development of phenotype-aware predictive models provided valuable insights into key risk features. The model placed a significant emphasis on lifestyle-related features such as the sleep factors indicated by PSG and food habits, over traditional and other risk factors for predicting long-term CVD outcomes in the general population. Furthermore, each of the four phenotypes emphasized distinct features, which may pave the way for precise risk management strategies tailored for different OSA phenotypic populations. Future research should validate these findings on additional datasets, and explore the utility of mobile and wearable devices for regularly collecting physiological and lifestyle data over extended periods, which could offer a more accurate representation of lifestyle information, potentially providing early warning of CV risks. Declarations Ethics approval and consent to participate The dataset employed in this study originates from the Multi-Ethnic Study of Atherosclerosis (MESA), with informed consent obtained from all participants in the MESA study. This research was conducted under the approval of the Shenzhen Technology University Ethics Committee (No: SZTUEA-20225011, date: 18 May 2022). Consent for publication Not applicable. Availability of data and materials The MESA datasets request should be directed to https://biolincc.nhlbi.nih.gov/login/?next=/requests/data-request/10761/view/. The MESA Sleep Ancillary study was funded by NIH-NHLBI Association of Sleep Disorders with Cardiovascular Health Across Ethnic Groups (RO1 HL098433). MESA is supported by NHLBI funded contracts HHSN268201500003I, N01-HC-95159, N01-HC-95160, N01-HC-95161, N01-HC-95162, N01-HC-95163, N01-HC-95164, N01-HC-95165, N01-HC-95166, N01-HC-95167, N01-HC-95168 and N01-HC-95169 from the National Heart, Lung, and Blood Institute, and by cooperative agreements UL1-TR-000040, UL1-TR-001079, and UL1-TR-001420 funded by NCATS. The National Sleep Research Resource was supported by the National Heart, Lung, and Blood Institute (R24 HL114473, 75N92019R002). This manuscript was not prepared in collaboration with MESA in­vestigators and does not necessarily reflect the opinions or views of MESA, or the NHLBI. Competing interests The authors declare that they have no conflict of interest. Funding This work was supported by Guangdong Basic and Applied Basic Research Foundation [2021A1515110025], Young Scientists Fund from National Natural Science Foundation of China (NSFC) [62301333], Research Foundation of Education Department of Guangdong Province [2022ZDJS115] and the Common University Innovation Team Project of Guangdong [2021KCXTD041]. Authors’ contributions Conceptualization, Yali Zheng; methodology, Zhengbi Song, Yali Zheng; validation and formal analysis, Zhengbi Song; investigation, Xiao Peng, Bo Cheng; data curation, Zhengbi Song, Yu Huang; writing—original draft preparation, Zhengbi Song; writing—review and editing, Yali Zheng, Min Min; visualization, Zhengbi Song; supervision, Yali Zheng, Min Min.; funding acquisition, Yali Zheng. All authors have read and agreed to the published version of the manuscript. Acknowledgements Not applicable. References World Health Organization[EB/OL]. https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds ). Kannel WB, Mcgee DL. Diabetes and Cardiovascular Disease: The Framingham Study[J]. JAMA. 1979;241(19):2035–8. Conroy RM, Pyörälä K, Fitzgerald AP, et al. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the SCORE project[J]. Eur Heart J. 2003;24(11):987–1003. Mehilli J, Kastrati A, Dirschinger J, et al. Sex-based analysis of outcome in patients with acute myocardial infarction treated predominantly with percutaneous coronary intervention[J]. Volume 287. JAMA-JOURNAL OF THE AMERICAN MEDICAL ASSOCIATION; 2002. pp. 210–5. 2. Mohd Faizal AS, Thevarajah TM, Khor SM, et al. A review of risk prediction models in cardiovascular disease: conventional approach vs. artificial intelligent approach[J]. Comput Methods Programs Biomed. 2021;207:106190. Than MP, Pickering JW, Sandoval Y, et al. Machine Learning to Predict the Likelihood of Acute Myocardial Infarction[J]. Circulation. 2019;140(11):899–909. Piros P, Ferenci T, Fleiner R, et al. Comparing machine learning and regression models for mortality prediction based on the Hungarian Myocardial Infarction Registry[J]. Knowl Based Syst. 2019;179:1–7. Steele AJ, Cakiroglu SA, Shah AD et al. Machine learning models in electronic health records can outperform conventional survival models for predicting patient mortality in coronary artery disease[J]. bioRxiv, 2018: 256008. Wallert J, Tomasoni M, Madison G et al. Predicting two-year survival versus non-survival after first myocardial infarction using machine learning and Swedish national register data[J]. BMC Med Inf Decis Mak, 2017, 17(1). Weng SF, Reps J, Kai J, et al. Can machine-learning improve cardiovascular risk prediction using routine clinical data?[J]. PLoS ONE. 2017;12(4):e0174944. Goff DC, Lloyd-Jones DM, Bennett G, et al. 2013 ACC/AHA Guideline on the Assessment of Cardiovascular Risk[J]. Circulation. 2014;129(25suppl2):S49–73. Javaheri S, Barbe F, Campos-Rodriguez F, et al. Sleep Apnea: Types, Mechanisms, and Clinical Cardiovascular Consequences[J]. J Am Coll Cardiol. 2017;69(7):841–58. Strassberger C, Zou D, Penzel T, et al. Beyond the AHI-pulse wave analysis during sleep for recognition of cardiovascular risk in sleep apnea patients[J]. J Sleep Res. 2021;30(6):e13364. Kim JW, Won TB, Rhee CS, et al. Polysomnographic phenotyping of obstructive sleep apnea and its implications in mortality in Korea[J]. Sci Rep. 2020;10(1):13207. Mazzotti DR, Keenan BT, Lim DC, et al. Symptom Subtypes of Obstructive Sleep Apnea Predict Incidence of Cardiovascular Outcomes[J]. Am J Respir Crit Care Med. 2019;200(4):493–506. Zinchuk AV, Jeon S, Koo BB, et al. Polysomnographic phenotypes and their cardiovascular implications in obstructive sleep apnoea[J]. Thorax. 2018;73(5):472–80. Lacedonia D, Carpagnano GE, Sabato R, et al. Characterization of obstructive sleep apnea-hypopnea syndrome (OSA) population by means of cluster analysis[J]. J Sleep Res. 2016;25(6):724–30. Shahrbabaki SS, Linz D, Hartmann S, et al. Sleep arousal burden is associated with long-term all-cause and cardiovascular mortality in 8001 community-dwelling older men and women[J]. Eur Heart J. 2021;42(21):2088–99. Azarbarzin A, Sands SA, Stone KL, et al. The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the Osteoporotic Fractures in Men Study and the Sleep Heart Health Study[J]. Eur Heart J. 2019;40(14):1149–57. Kendzerska T, Mollayeva T, Gershon AS, et al. Untreated obstructive sleep apnea and the risk for serious long-term adverse outcomes: A systematic review[J]. Sleep Med Rev. 2014;18(1):49–59. Leung RST, Comondore VR, Ryan CM, et al. Mechanisms of sleep-disordered breathing: causes and consequences[J]. Pflügers Archiv - Eur J Physiol. 2011;463(1):213–30. Dami S, Yahaghizadeh M. Predicting cardiovascular events with deep learning approach in the context of the internet of things[J]. Volume 33. Neural Computing & Applications; 2021. pp. 7979–96. 13. Chen H, Lundberg SM, Erion G, et al. Forecasting adverse surgical events using self-supervised transfer learning for physiological signals[J]. NPJ Digit Med. 2021;4(1):167. Sbrollini A, De Jongh MC, Ter Haar CC et al. Serial electrocardiography to detect newly emerging or aggravating cardiac pathology: A deep–learning approach[J]. Biomed Eng Online, 2019, 18(15). Raita Y, Goto T, Faridi MK et al. Emergency department triage prediction of clinical outcomes using machine learning models[J]. Crit Care, 2019, 23(1). Kaji DA, Zech JR, Kim JS, et al. An attention based deep learning model of clinical events in the intensive care unit[J]. PLoS ONE. 2019;14(2):e0211057. Barrett LA, Payrovnaziri SN, Bian J et al. Building computational models to predict one-year mortality in ICU patients with acute myocardial infarction and post myocardial infarction syndrome[J]. AMIA Jt Summits Transl Sci Proc, 2019, 2019: 407–416. Zhang GQ, Cui L, Mueller R, et al. The National Sleep Research Resource: towards a sleep data commons[J]. J Am Med Inf Assoc. 2018;25(10):1351–8. Chen X, Wang R, Zee P, et al. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA)[J]. Sleep. 2015;38(6):877–88. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions[J]. Adv Neural Inf Process Syst, 2017, 30. Shlomai G, Shemesh J, Segev S et al. The Multi-Ethnic Study of Atherosclerosis-Calcium Score Improves Statin Treatment Allocation in Asymptomatic Adults[J]. Front Cardiovasc Med, 2022, 9. Mehta R, Ning HY, Bansal N, et al. Ten-Year Risk-Prediction Equations for Incident Heart Failure Hospitalizations in Chronic Kidney Disease: Findings from the Chronic Renal Insufficiency Cohort Study and the Multi- Ethnic Study of Atherosclerosis[J]. J Card Fail. 2022;28(4):540–50. Hathaway QA, Yanamala N, Budoff MJ, et al. Deep neural survival networks for cardiovascular risk prediction: The Multi-Ethnic Study of Atherosclerosis (MESA)[J]. Comput Biol Med. 2021;139:104983. Yeghiazarians Y, Jneid H, Tietjens JR, et al. Obstructive Sleep Apnea and Cardiovascular Disease: A Scientific Statement From the American Heart Association[J]. Circulation. 2021;144(3):e56–67. Saki N, Babaahmadi-Rezaei H, Rahimi Z et al. Impact of modifiable risk factors on prediction of 10-year cardiovascular disease utilizing framingham risk score in Southwest Iran[J]. BMC Cardiovasc Disord, 2023, 23(1). Peng M, Hou F, Cheng Z, et al. A Cardiovascular Disease Risk Score Model Based on High Contribution Characteristics[J]. Appl Sci. 2023;13(2):893. Park S, Kim YG, Ann SH et al. Prediction of the 10-year risk of atherosclerotic cardiovascular disease in the Korean population[J]. Epidemiol Health, 2023, 45. Sharma LD, Sunkaria RK. Inferior myocardial infarction detection using stationary wavelet transform and machine learning approach[J]. SIViP. 2018;12(2):199–206. Lui HW, Chow KL. Multiclass classification of myocardial infarction with convolutional and recurrent neural networks for portable ECG devices[J]. Inf Med Unlocked. 2018;13:26–33. Lodhi AM, Qureshi AN, Sharif U et al. A novel approach using voting from ecg leads to detect myocardial infarction[C]. Adv Intell Syst Comput, 2018: 337–52. Dohare AK, Kumar V, Kumar R. Detection of myocardial infarction in 12 lead ECG using support vector machine[J]. Appl Soft Comput. 2018;64:138–47. Rundo JV, Downey R 3. Polysomnography[J] Handb Clin Neurol. 2019;160:381–92. Cuellar NG. The effects of periodic limb movements in sleep (PLMS) on cardiovascular disease[J]. Heart Lung. 2013;42(5):353–60. Sabanayagam C, Shankar A. Sleep duration and cardiovascular disease: results from the National Health Interview Survey[J]. Sleep. 2010;33(8):1037–42. Koutentakis M, Surma S, Rogula S, et al. The Effect of a Vegan Diet on the Cardiovascular System[J]. J Cardiovasc Dev Disease. 2023;10(3):94. Dyńka D, Kowalcze K, Charuta A, et al. The Ketogenic Diet and Cardiovascular Diseases[J]. Nutrients. 2023;15(15):3368. Yu E, Malik VS, Hu FB. Cardiovascular Disease Prevention by Diet Modification JACC Health Promotion Series[J]. J Am Coll Cardiol. 2018;72(8):914–26. Pan A, Lin X, Hemler E, et al. Diet and Cardiovascular Disease: Advances and Challenges in Population-Based Studies[J]. Cell Metab. 2018;27(3):489–96. Ding FH, Cotton-Clay A, Fava L, et al. Polysomnographic validation of an under-mattress monitoring device in estimating sleep architecture and obstructive sleep apnea in adults[J]. Sleep Med. 2022;96:20–7. Hassan AR, Bhuiyan MIH. An automated method for sleep staging from EEG signals using normal inverse Gaussian parameters and adaptive boosting[J]. Neurocomputing. 2017;219:76–87. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Published Journal Publication published 16 Mar, 2026 Read the published version in BMC Medical Informatics and Decision Making → Version 1 posted Editorial decision: Revision requested 06 Sep, 2024 Reviews received at journal 23 Jul, 2024 Reviews received at journal 10 Jul, 2024 Reviewers agreed at journal 07 Jul, 2024 Reviewers agreed at journal 05 Jul, 2024 Reviewers invited by journal 27 Mar, 2024 Editor assigned by journal 13 Mar, 2024 Submission checks completed at journal 13 Mar, 2024 First submitted to journal 12 Mar, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4084889","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":279478345,"identity":"cc7b1fe7-db3c-4c92-8828-ffd1559119ec","order_by":0,"name":"Yali Zheng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7klEQVRIiWNgGAWjYDACCQglw8beABU5QKQWHjaeA1DVRGthkEggUov87OZjD7+22fHwST5//PljG4Mc340Exs8FeLQwzjmWbizblszDJp1jYHCwjcFY8kYCs/QMPFqYJXLMpCXbmEFaGBKAWhI33EhgY+bBo4VNIv8bUEs9D5vk8QcHgFrqCWrhkchhk/zYdpiHTYLBsAGoJcGAkBYJiTQzaYZzx4GBnGPMcOachOHMMw+bpfFpkZ+R/EzyR1m1nHz78ccfKsps5PmOJx/8jE8LCDDzsiFsBWLGBgIagEp+/CGoZhSMglEwCkYyAAA4AUW/KaHFZwAAAABJRU5ErkJggg==","orcid":"","institution":"Shenzhen Technology University","correspondingAuthor":true,"prefix":"","firstName":"Yali","middleName":"","lastName":"Zheng","suffix":""},{"id":279478347,"identity":"767c36ee-5c02-48fe-98c6-e33ba6ed95ea","order_by":1,"name":"Zhengbi Song","email":"","orcid":"","institution":"Shenzhen Technology University","correspondingAuthor":false,"prefix":"","firstName":"Zhengbi","middleName":"","lastName":"Song","suffix":""},{"id":279478348,"identity":"4f97a87f-5fc8-4c50-82f3-6cd4c6d3d678","order_by":2,"name":"Bo Cheng","email":"","orcid":"","institution":"Shenzhen Technology University","correspondingAuthor":false,"prefix":"","firstName":"Bo","middleName":"","lastName":"Cheng","suffix":""},{"id":279478349,"identity":"945b162d-9f10-4aa0-a705-16b5b7899d2d","order_by":3,"name":"Xiao Peng","email":"","orcid":"","institution":"Shenzhen Technology University","correspondingAuthor":false,"prefix":"","firstName":"Xiao","middleName":"","lastName":"Peng","suffix":""},{"id":279478350,"identity":"ed31a7e5-b3a0-43cd-9ba4-7c27d330dd63","order_by":4,"name":"Yu Huang","email":"","orcid":"","institution":"Shenzhen Technology University","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Huang","suffix":""},{"id":279478351,"identity":"ecd7cfb0-457e-4ad3-99c4-accf4bb78c17","order_by":5,"name":"Min Min","email":"","orcid":"","institution":"Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"","lastName":"Min","suffix":""}],"badges":[],"createdAt":"2024-03-12 14:32:30","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4084889/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4084889/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12911-026-03439-8","type":"published","date":"2026-03-16T15:57:25+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":52793320,"identity":"7fae4ff7-6c0d-46cd-872e-b3831231777d","added_by":"auto","created_at":"2024-03-15 20:27:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":909943,"visible":true,"origin":"","legend":"\u003cp\u003eThe overview of the study pipeline\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/b47abfb62a06bff812798a80.png"},{"id":52793318,"identity":"ee5053f5-ef8c-43ae-b556-e8dd535c0ade","added_by":"auto","created_at":"2024-03-15 20:27:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":434897,"visible":true,"origin":"","legend":"\u003cp\u003eGeneration of the overnight feature sequences of the five sleep events, including: arousal, hypopnea, apnea, hypoxemia and PLMS. Ts: start time of sleep, Te: end time of sleep\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/4e245110e562d6dfad953b23.png"},{"id":52793737,"identity":"ebde6770-48a1-405b-a51e-4f0056043278","added_by":"auto","created_at":"2024-03-15 20:35:43","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":776575,"visible":true,"origin":"","legend":"\u003cp\u003eThe proposed OSA phenotyping based deep learning model with different training strategies and feature sets for CVD risk prediction.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/9f42d9849b4fee96302ccbf5.png"},{"id":52793322,"identity":"9abd7c89-1ad1-4a8e-8baa-8c216678f7d1","added_by":"auto","created_at":"2024-03-15 20:27:44","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":473942,"visible":true,"origin":"","legend":"\u003cp\u003eThe AUC-ROC values of various ML models with and without OSA clustering.\u003c/p\u003e","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/33a7b3d149dcbfe846fbf02a.png"},{"id":52793321,"identity":"2ab3e29f-7092-4fad-a509-9f7fab7e255e","added_by":"auto","created_at":"2024-03-15 20:27:44","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":717248,"visible":true,"origin":"","legend":"\u003cp\u003eHeatmap of feature importance of the four phenotypes\u003c/p\u003e","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/6b7791950f075e6f20d25b61.png"},{"id":52793317,"identity":"9ac01cac-df82-42a0-8f06-e9fa64eec4ab","added_by":"auto","created_at":"2024-03-15 20:27:43","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":268364,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of importances to specific PSG and FOOD FREQUENCY features across the four phenotypes\u003c/p\u003e","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/1997ffa1c2878026da3ad221.png"},{"id":52793324,"identity":"7e1d1cc2-9eb0-473d-9f85-f0879588eb55","added_by":"auto","created_at":"2024-03-15 20:27:44","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":583916,"visible":true,"origin":"","legend":"\u003cp\u003eTop-five important foods of the four phenotypes\u003c/p\u003e","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/0d005b857525bae30b71ea17.png"},{"id":105223302,"identity":"7cd96437-6c35-4d95-90be-5c071c0df6bd","added_by":"auto","created_at":"2026-03-23 16:03:10","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6098636,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/68b5af48-731e-4f60-906e-63a8d3d53ba5.pdf"},{"id":52793316,"identity":"f1acbad5-6dfa-4e58-a99d-4de422f16a7f","added_by":"auto","created_at":"2024-03-15 20:27:43","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":30228,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-4084889/v1/516981929b1988e18fb70789.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Integrating Phenotypic Information of Obstructive Sleep Apnea and Deep Representation of Sleep-Event Sequences for Cardiovascular Risk Prediction","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eCardiovascular diseases (CVD) are the leading cause of death globally. According to the World Health Organization, the annual deaths from CVDs exceed 17 million, accounting for approximately 32% of all mortalities [1]. Risk prediction is crucial for the prevention and treatment of CVDs. Traditionally, medical experts make qualitative predictions based on experience, supplemented by statistical models such as Framingham Risk Score [2], SCORE [3], and Cox Proportional Hazard [4]. Recently, Machine Learning (ML) and Deep Learning (DL) techniques have emerged as powerful tools for enhancing CVD risk prediction [5-11].\u003c/p\u003e\n\u003cp\u003eDespite these advances, many studies have not fully considered the complex interplay of CVD risk factors with concurrent clinical conditions. There is a recognized link between Obstructive Sleep Apnea (OSA) and increased CVD risk, with OSA prevalence in CVD patients ranging from 40% to 80% [12]. Recent research has focused on OSA phenotyping using polysomnography (PSG) data, identifying distinct CVD risk levels among different OSA phenotypes [13-17]. The pathophysiology of OSA-related cardiovascular complications has been recognized as multifaceted, encompassing four main domains: sleep architecture disturbances, autonomic dysregulation, breathing disturbances, and hypoxemia [18-21]. However, existing research have not integrated the OSA phenotypic information with traditional risk factors in predictive modeling, leaving the actual predictive effectiveness uncertain.\u003c/p\u003e\n\u003cp\u003eOn the other hand, the increasing ease of gathering long-term physiological data through wearable and mobile devices, combined with the advanced capabilities of deep learning (DL) in complex signal representation, has spurred studies integrating these signals into CV risk prediction. For example, Dami et al. developed a DL model to predict arterial events over a course of a few weeks/months, which outperformed traditional machine learning (ML) and other DL approaches [22]. They applied a Deep Belief Network (DBN) to select effective features from \u003cem\u003e5-minute\u003c/em\u003e ECG recordings, and then made predictions with a long short-term memory (LSTM) network in conjunction with features from the electronic health records (EHRs). Chen et al. employed LSTM to learn deep representations from multi-channel physiological signals. By combining these with static features, they then used a Gradient Boosting Machine (GBM) to predict six types of adverse events in the next 5 minutes of surgery with promising results [23]. Yet, existing studies [24-27] have mainly focused on the prediction of specific CV events and over relatively short period of time. There is limited research about utilizing extended periods of physiological signals for long-term CV risk assessment. Considering the advancements in wearable and mobile technologies, the integration of EHR information and daily health data for long-term CVD risk prediction holds significant potential benefits for cardiovascular healthcare.\u003c/p\u003e\n\u003cp\u003eTherefore, this study aims to explore effective methods to incorporate OSA phenotypic information and leverage extended periods of physiological data for CV risk prediction over long-term periods. And, a feature importance analysis is conducted to identify the most significant risk features in the general population and across different OSA phenotypes to provide insights for precise CV risk management. The main contribution of this study is summarized as follows:\u003c/p\u003e\n\u003cp\u003e1. Proposed a novel and effective method for precise long-term CV risk prediction, by integrating deep features learned from overnight physiological sequences and employing an OSA-phenotype-contrastive training strategy; \u003c/p\u003e\n\u003cp\u003e2. Identified lifestyle-related features as significant CV risk factors among a comprehensive feature set, and revealed distinct risk features across different OSA phenotypes, offering new insights for precise management of CV risks.\u003c/p\u003e"},{"header":"2. METHODS","content":"\u003ch2\u003e2.1 Datasets and Subjects\u003c/h2\u003e\n\u003cp\u003eThis study utilized data from the Multi-Ethnic Study of Atherosclerosis (MESA), initiated and collected by the National Heart, Lung, and Blood Institute (NHLBI)\u0026nbsp;[\u003ca href=\"#_ENREF_28\" title=\"Zhang, 2018 #102\"\u003e28\u003c/a\u003e,\u003ca href=\"#_ENREF_29\" title=\"Chen, 2015 #103\"\u003e29\u003c/a\u003e]. MESA study recruited a total of 6,814 participants without CVDs aged between 45 to 84 years from four ethnic groups (African Americans, Chinese Americans, Hispanics, and Caucasians). Five examinations were conducted between 2000 and 2011, and clinical outcomes were assessed every 9 to 12 months during this period, including: myocardial infarction, angina, heart failure, coronary heart disease, and death.\u003c/p\u003e\n\u003cp\u003eDuring Exam_5, 2,237 participants of the MESA population participated in an auxiliary sleep study, and 2,489 static features were recorded, including demographics, anthropometrics, medication usage, medical history, imaging risk factors, etc. Among them, 1,874 participants had provided overnight PSG data, which includes 27 continuous physiological signals recorded during sleep, including electrocardiography, electroencephalography, pulse, nasal airflow, blood oxygen, periodic limb movements (PLMS), and more. The average recording duration was 12 hours per participant. Additionally, 615 sleep-related static features were extracted from the sleep questionnaires, actigraphy-derived parameters and average counts of overnight sleep events derived from PSG recordings (so there are in total of 3,104 static features). \u0026nbsp;Manual annotations were performed on the PSG data for identifying sleep stages and events (such as arousal, hypopnea, apnea, hypoxemia, periodic limb movements). This study mainly focused on the analysis of the 1,874 participants, who have complete sets of static features and PSG recordings, and 175 of them (9.3%) experienced CVD events. The whole dataset was divided into training and validation sets, comprising 1,687 and 187 subjects, respectively. The validation set proportion is 9.3%.\u003c/p\u003e\n\u003ch2\u003e2.2 Study Pipeline\u003c/h2\u003e\n\u003cp\u003eInitially, feature selection was performed on all static features by lasso-logistic regression. Clustering was then conducted based on the overnight average values of PSG features (named PSG static features) to identify OSA phenotypes. Subsequently, several ML and DL models with different feature selection and training strategies were built to explore the most effective method of incorporating OSA phenotypic information for CVD risk prediction. Finally, the feature importance analysis was conducted for each phenotype based on the best performing model. The study pipeline is shown in\u0026nbsp;Fig.1.\u003c/p\u003e\n\u003ch3\u003e2.2.1 Data Preprocessing\u003c/h3\u003e\n\u003cp\u003eThe following preprocessing steps were applied to the 3,104 static features (2,489 exam_5 static features and 615 PSG static features) of all participants. First, data outliers were removed, including blank values, duplicates, and irrelevant numerical values. Missing values were then imputed with mean interpolation. Finally, z-score normalization was applied to standardize each feature. The preprocessing process resulted in each participant having 1,600 normalized static features.\u003c/p\u003e\n\u003cp\u003eConsidering our goal is to predict 5-year CVD risk, while multiple high-fidelity physiological signals exhibit highly complex temporal patterns throughout the entire sleep period, directly using them\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eas model inputs may not capture efficient and effective information related to long-term CVD risks. Therefore, this study chose to learn deep representations from the feature sequences of five sleep events known contributed to CVD risks, including arousal, hypopnea, apnea, hypoxemia and PLMS. The sequences of the five sleep events were generated according to the PSG labels through the following steps. Firstly, sleep sequences were generated with a 5-second sampling interval during the start and end sleeping time of each participant. Then, the time slots of the occurrences of each sleep event were extracted from the PSG labels to determine whether the event occurred within any of the 5-second sampling intervals. These events were marked as \u0026apos;1\u0026apos; if they occurred within an interval, and \u0026apos;0\u0026apos; otherwise. This created five overnight sleep-event sequences for further analysis. (Fig.2) illustrates the generation process of the overnight feature sequences of five sleep events.\u003c/p\u003e\n\u003ch3\u003e2.2.2 Static feature selection\u003c/h3\u003e\n\u003cp\u003eLasso logistic regression was used to select a subset of features from the 1,600 static features using the training set. Ten-fold cross validation was conducted to determine the regularization hyperparameter, alpha. It was adjusted across a range of values: 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.015, 0.02 and 0.025, to identify its optimal value that yields the best predictive performance on the validation set.\u0026nbsp;\u003c/p\u003e\n\u003ch3\u003e2.2.3 OSA clustering\u003c/h3\u003e\n\u003cp\u003e29 PSG static features (the features shown in Supplementary Table 2) were employed for OSA clustering utilizing the K-means clustering. A range of 2 to 6 clusters was explored to determine the optimal cluster number by the silhouette and elbow methods.\u003c/p\u003e\n\u003cp\u003eMann-Whitney U tests were then conducted on the 29 PSG static features between different OSA phenotypes to identify the most relevant PSG features for each phenotype. A significance level below 0.05 was considered as significantly different, while a \u003cem\u003eP\u003c/em\u003e value below 0.001 was considered highly significant.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCox proportional hazards regression analysis was also performed to estimate the hazard ratio (HR) and 95% confidence intervals (CI) of occurring CVD events within five years for different OSA phenotypes.\u003c/p\u003e\n\u003ch3\u003e2.2.4 OSA phenotyping-based CVD risk prediction modeling\u003c/h3\u003e\n\u003cp\u003eTo assess the value of incorporating OSA phenotypic information in CVD risk prediction, several classic ML models were employed in this study, including logistic regression (LR), support vector machine (SVM), decision tree (DT), random forest (RF), gradient boosting machine (GBM) and multilayer perceptron (MLP). The phenotype-agnostic method did not consider any phenotypic information. For phenotype-specific models, two different strategies to integrate the phenotypic information and three different feature sets were evaluated. The details of the phenotype-specific ML models are described as below:\u003c/p\u003e\n\u003cp\u003e1.\u003cem\u003ePheno_fuse_ML\u003c/em\u003e: The OSA phenotyping labels (Phenotypes 1, 2, 3, 4) were incorporated as a new feature along with selected static features, serving as input for the ML models to predict CVD risk across the entire population.\u003c/p\u003e\n\u003cp\u003e2.\u003cem\u003ePheno_specific_ML:\u0026nbsp;\u003c/em\u003eThe strategy develops CVD risk prediction models specific to each phenotypic population, using the three different sets of features:\u003c/p\u003e\n\u003cp\u003e(1) \u003cem\u003ePheno_Spec_ML1\u003c/em\u003e: Uses only selected static features.\u003c/p\u003e\n\u003cp\u003e(2) \u003cem\u003ePheno_Spec_ML2\u003c/em\u003e: Combines the 29 PSG static features with the selected static features.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003e(3) \u003cem\u003ePheno_Spec_ML3\u003c/em\u003e：Fuses phenotype-specific static PSG features representing each phenotype with the selected static features.\u003c/p\u003e\n\u003cp\u003eTo further explore the value of overnight sleep-event feature sequences in CVD risk prediction, a two-layer LSTM network was employed to learn deep representations from the overnight feature sequences, which were then combined with the static features in a fully connected layer to predict CVD risk. We further proposed a phenotype-contrastive training strategy, i.e., the \u003cem\u003eContrast_pheno_DL\u003c/em\u003e, to enhance the model performance. To validate the effectiveness of the strategy, two other phenotype-specific DL models with different feature sets were implemented for comparison, i.e., \u003cem\u003ePheno_spec_DL1\u003c/em\u003e and \u003cem\u003ePheno_spec_DL2\u003c/em\u003e. The DL model architecture with different training strategies and feature sets are illustrated in (Fig.3), and the details of the three DL models are described below:\u003c/p\u003e\n\u003cp\u003e3. \u003cem\u003ePheno_spec_DL1\u003c/em\u003e: The feature set of this model includes all selected static features, the 29 PSG static features, and deep representations of the five sleep-event feature sequences.\u003c/p\u003e\n\u003cp\u003e4. \u003cem\u003ePheno_spec_DL2\u003c/em\u003e: The feature set comprises the selected static features, the 29 PSG static features and sleep-event feature sequences that are specifically relevant to each phenotype. The aim is to focus on features that are directly related to each phenotype.\u003c/p\u003e\n\u003cp\u003e5. \u003cem\u003eContrast_pheno_DL\u003c/em\u003e: This model utilizes the same features as the \u003cem\u003ePheno_spec_DL1\u003c/em\u003e, while its training approach is distinct in strategically merging different phenotypic populations that have distinct risk levels. The goal is to identify and learn discriminative features among various phenotypes.\u003c/p\u003e\n\u003ch3\u003e2.2.5 Performance Evaluation\u003c/h3\u003e\n\u003cp\u003eIn this study, subjects who experienced cardiovascular events were labeled as positive samples. The evaluation metrics include: Accuracy, Precision, Recall, F1-Score, Area under Curve of the Receiver Operating Characteristic Curve (AUC-ROC), Area under Curve of the Precision-Recall Curve (AUC-PRC).\u003c/p\u003e\n\u003ch3\u003e2.2.6 Feature Importance Analysis\u003c/h3\u003e\n\u003cp\u003eThe influence of all features on the model\u0026apos;s predictive performance was evaluated by calculating the SHapley Additive exPlanation (SHAP) values derived from the field of cooperative game theory\u0026nbsp;[\u003ca href=\"#_ENREF_30\" title=\"Lundberg, 2017 #130\"\u003e30\u003c/a\u003e], to identify the key features for different phenotypes. The SHAP value of each feature was calculated according to the following equation:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"1085\" height=\"253\"\u003e\u003cbr\u003e\u003c/p\u003e"},{"header":"3. RESULTS","content":"\u003ch2\u003e3.1. Feature selection\u003c/h2\u003e\n\u003cp\u003e51 features were selected by the lasso logistic regression with the alpha parameter set to 0.015. These features encompass not only conventional CVD risk factors like smoking, high-density lipoprotein cholesterol, and total cholesterol, but also extend to features from computed tomography (CT) and magnetic resonance imaging (MRI) ones. Additionally, they include a broad spectrum of other characteristics such as sleep patterns, anthropometry, dietary habits, cognitive status, and more. A detailed list of the selected features can be found in the Supplementary Table 1.\u003c/p\u003e\n\u003ch2\u003e3.2 OSA phenotypes\u003c/h2\u003e\n\u003cp\u003e3 and 6 OSA phenotypes were identified by the silhouette method and 4 by the elbow method on s all subjects, respectively. For 3 phenotypes, there was insufficient differentiation among the four distinct aspects of OSA pathophysiology, i.e., sleep architecture disturbances, autonomic dysregulation, respiratory disturbances, and hypoxemia\u0026nbsp;[\u003ca href=\"#_ENREF_18\" title=\"Shahrbabaki, 2021 #15\"\u003e18-21\u003c/a\u003e]. In contrast, the division of 6 phenotypes was excessively granular, resulting in a lack of clear distinction between phenotypes. Categorizing these 29 features into 4 phenotypes effectively reflects various aspects of OSA pathophysiology, as shown in Supplementary Table 2. In each phenotype, the most prominent PSG features are highlighted in bold. Each phenotype was named according to its dominant pathophysiological domain: \u003cem\u003eMild, Respiratory related, Sleep related, and Combined.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMild Phenotype\u003c/em\u003e: Characterized by the lowest positive sample proportion of 6.7%, this phenotype exhibits the fewest respiratory events, highest sleep efficiency, near-normal sleep structure, and nighttime oxygen saturation levels (SpO\u003csub\u003e2\u003c/sub\u003e). It typically comprises healthy (Apnea Hypopnea Index, AHI \u0026lt; 5) to mild OSA (5\u0026nbsp;\u0026le;AHI \u0026lt; 15) subjects.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSleep Related Phenotype:\u0026nbsp;\u003c/em\u003eWith the highest positive sample proportion of 15.1%, this phenotype presents significantly lower sleep efficiency, higher arousal and PLMS compared to the \u003cem\u003eMild\u003c/em\u003e type (\u003cem\u003ep\u003c/em\u003e\u0026lt;0.001). It falls under moderate OSA (15 \u0026le; AHI \u0026lt; 30). The features of breathing disturbance and hypoxemia are less prominent compared to the \u003cem\u003eRespiratory\u0026nbsp;\u003c/em\u003eand \u003cem\u003eCombined\u0026nbsp;\u003c/em\u003ephenotypes.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eRespiratory Related Phenotype:\u0026nbsp;\u003c/em\u003eThis phenotype is characterized by significantly more frequent respiratory events, lower levels of nighttime SpO\u003csub\u003e2\u003c/sub\u003e compared to the \u003cem\u003eMild\u003c/em\u003e type (\u003cem\u003ep\u003c/em\u003e\u0026lt;0.001) and with moderate OSA. The features of sleep architecture disturbance and autonomic dysregulation are less prominent compared to the \u003cem\u003eSleep related\u003c/em\u003e and \u003cem\u003eCombined\u003c/em\u003e phenotypes.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eCombined Phenotype\u003c/em\u003e: This type exhibits a mixture of the four pathophysiological domains\u0026mdash;sleep architecture disturbance, autonomic dysregulation, breathing disturbance, and hypoxemia. It is classified as severe OSA (AHI\u0026nbsp;\u0026ge;\u0026nbsp;30).\u003c/p\u003e\n\u003cp\u003eAs shown in (Table 1), the HRs for CVD events significantly differed among the four OSA phenotypes. Compared to the \u003cem\u003eMild\u003c/em\u003e phenotype, the \u003cem\u003eRespiratory\u003c/em\u003e, \u003cem\u003eSleep\u003c/em\u003e, and\u003cem\u003e\u0026nbsp;Combined\u003c/em\u003e phenotypes show 2.708, 2.651, and 2.849 times higher risks to experience CVD events. Notably, the HRs, as determined by OSA phenotyping, does not align with the actual CVD event rates observed within each phenotype. This discrepancy suggests that additional factors are necessary to achieve more accurate predictions of CVD risk.\u003c/p\u003e\n\u003ch2\u003e3.3 OSA Phenotyping based\u0026nbsp;CVD Risk Prediction\u003c/h2\u003e\n\u003cp\u003eAs illustrated in (Fig.4), the AUC-ROC values of various ML models were notably improved by incorporating OSA phenotyping for CVD risk prediction. Among all models, the MLP model achieved the best performance, with am AUC-ROC value of 0.656. \u003cem\u003ePheno_spec_ML3\u003c/em\u003e presented the most significant improvement with the AUC-ROC value improved to 0.746. These findings underscore the importance of OSA phenotypic information in CVD risk prediction. A detailed comparison of model performance of various ML models with and without OSA clustering is shown in Table 3 in Supplementary material.\u003c/p\u003e\n\u003ch2\u003e3.4 Integrating deep representation of feature sequences of sleep events\u003c/h2\u003e\n\u003cp\u003eTable 2\u0026nbsp;shows the performance of three DL modeling approaches that incorporated OSA phenotypic information and deep representations of sleep-event feature sequences. There is a notable improvement for the two phenotype-specific models (\u003cem\u003ePheno_spec_DL1\u003c/em\u003e and \u003cem\u003ePheno_spec_DL2\u003c/em\u003e) when deep features were included as compared to \u003cem\u003ePheno_spec_ML3\u003c/em\u003e, with the AUC-ROC and AUC-PRC values exceeding 0.83 and 0.50, respectively. The\u003cem\u003e\u0026nbsp;Contrast_pheno_DL\u003c/em\u003e model, which involves contrastive training with a combination of the\u003cem\u003e\u0026nbsp;Mild\u003c/em\u003e phenotype and one of the other phenotypes (i.e., \u003cem\u003eMild + Sleep\u003c/em\u003e, \u003cem\u003eMild + Respiratory\u003c/em\u003e, \u003cem\u003eMild + Combined\u003c/em\u003e), achieved even higher AUC-ROC and AUC-PRC values of 0.877 and 0.689, respectively. These findings highlight the substantial additional value provided by overnight sleep-event feature sequences in predicting CVD risk across different OSA phenotypes.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e Cox hazard ratios for CVD events of four OSA phenotypes\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"368\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"40.76086956521739%\"\u003e\n \u003cp\u003eParameters\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eOSA phenotypes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.652173913043477%\"\u003e\n \u003cp\u003eHR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.097826086956523%\"\u003e\n \u003cp\u003e95% CI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.48913043478261%\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"40.76086956521739%\" valign=\"top\"\u003e\n \u003cp\u003eMild\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.652173913043477%\"\u003e\n \u003cp\u003eReference\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.097826086956523%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.48913043478261%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"40.76086956521739%\" valign=\"top\"\u003e\n \u003cp\u003eRespiratory related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.652173913043477%\"\u003e\n \u003cp\u003e2.708\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.097826086956523%\"\u003e\n \u003cp\u003e1.913-3.888\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.48913043478261%\"\u003e\n \u003cp\u003e0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"40.76086956521739%\" valign=\"top\"\u003e\n \u003cp\u003eCombined\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.652173913043477%\"\u003e\n \u003cp\u003e2.849\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.097826086956523%\"\u003e\n \u003cp\u003e1.566-6.424\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.48913043478261%\"\u003e\n \u003cp\u003e0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"40.76086956521739%\" valign=\"top\"\u003e\n \u003cp\u003eSleep related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.652173913043477%\"\u003e\n \u003cp\u003e2.651\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.097826086956523%\"\u003e\n \u003cp\u003e1.901-4.714\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.48913043478261%\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2\u003c/strong\u003e Results of different CVD risk prediction models based on the MESA dataset\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"570\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"99.47368421052632%\" colspan=\"9\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelated studies\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.157894736842106%\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.157894736842106%\" colspan=\"2\"\u003e\n \u003cp\u003eNumber of subjects (Proportion of positive subjects, %)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"39.824561403508774%\" colspan=\"5\"\u003e\n \u003cp\u003eResearch Objectives\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.333333333333334%\"\u003e\n \u003cp\u003eAUC-ROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.157894736842106%\"\u003e\n \u003cp\u003eMES-C risk score\u0026nbsp;[\u003ca href=\"#_ENREF_31\" title=\"Shlomai, 2022 #116\"\u003e31\u003c/a\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.157894736842106%\" colspan=\"2\"\u003e\n \u003cp\u003e632（8.2%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"39.824561403508774%\" colspan=\"5\"\u003e\n \u003cp\u003ePredictive value of atherogenic calcium score for 10-year risk of CVD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.333333333333334%\"\u003e\n \u003cp\u003e0.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.157894736842106%\"\u003e\n \u003cp\u003ePCP-HFCKD risk equation\u0026nbsp;[\u003ca href=\"#_ENREF_32\" title=\"Mehta, 2022 #112\"\u003e32\u003c/a\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.157894736842106%\" colspan=\"2\"\u003e\n \u003cp\u003e2328（14.6%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"39.824561403508774%\" colspan=\"5\"\u003e\n \u003cp\u003ePatients with chronic kidney disease_10-year Heart Failure prediction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.333333333333334%\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.157894736842106%\"\u003e\n \u003cp\u003eDeepSurv\u0026nbsp;[\u003ca href=\"#_ENREF_33\" title=\"Hathaway, 2021 #36\"\u003e33\u003c/a\u003e]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.157894736842106%\" colspan=\"2\"\u003e\n \u003cp\u003e6814（-）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"39.824561403508774%\" colspan=\"5\"\u003e\n \u003cp\u003eOne-year Risk Prediction for Atherosclerotic Cardiovascular Disease\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.333333333333334%\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"99.47368421052632%\" colspan=\"9\"\u003e\n \u003cp\u003e\u003cstrong\u003eOur study\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5263157894736842%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.117338003502628%\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.309982486865149%\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.908931698774081%\" colspan=\"2\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.982486865148863%\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.558669001751314%\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.8861646234676%\"\u003e\n \u003cp\u003eAUC-PRC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.711033274956216%\" colspan=\"2\"\u003e\n \u003cp\u003eAUC-ROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5253940455341506%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.117338003502628%\"\u003e\n \u003cp\u003e\u003cem\u003ePheno_spec_ML3\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.309982486865149%\"\u003e\n \u003cp\u003e0.933\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.908931698774081%\" colspan=\"2\"\u003e\n \u003cp\u003e0.736\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.982486865148863%\"\u003e\n \u003cp\u003e0.446\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.558669001751314%\"\u003e\n \u003cp\u003e0.555\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.8861646234676%\"\u003e\n \u003cp\u003e0.496\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.711033274956216%\" colspan=\"2\"\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5253940455341506%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.117338003502628%\"\u003e\n \u003cp\u003e\u003cem\u003ePheno_spec_DL1\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.309982486865149%\"\u003e\n \u003cp\u003e0.920\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.908931698774081%\" colspan=\"2\"\u003e\n \u003cp\u003e0.579\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.982486865148863%\"\u003e\n \u003cp\u003e0.611\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.558669001751314%\"\u003e\n \u003cp\u003e0.595\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.8861646234676%\"\u003e\n \u003cp\u003e0.507\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.711033274956216%\" colspan=\"2\"\u003e\n \u003cp\u003e0.832\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5253940455341506%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.117338003502628%\"\u003e\n \u003cp\u003e\u003cem\u003ePheno_spec_DL2\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.309982486865149%\"\u003e\n \u003cp\u003e0.943\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.908931698774081%\" colspan=\"2\"\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.982486865148863%\"\u003e\n \u003cp\u003e0.608\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.558669001751314%\"\u003e\n \u003cp\u003e0.645\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.8861646234676%\"\u003e\n \u003cp\u003e0.573\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.236427320490368%\" colspan=\"3\"\u003e\n \u003cp\u003e0.839\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"23.117338003502628%\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003eContrast_pheno_DL\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.309982486865149%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.966\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.908931698774081%\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.851\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"9.982486865148863%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.750\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"11.558669001751314%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.797\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.8861646234676%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.689\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"14.711033274956216%\" colspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.877\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"0.5253940455341506%\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable 3\u0026nbsp;provides a detailed illustration of the performance of the DL models trained with specific (\u003cem\u003ePheno_spec_DL2\u003c/em\u003e) and contrastive phenotypes (\u003cem\u003eContrast_pheno_DL\u003c/em\u003e) across the four phenotypes. It reveals that combining the \u003cem\u003eMild\u003c/em\u003e phenotype with one of the other phenotypes generally improves model performance compared to the \u003cem\u003eMild\u003c/em\u003e phenotype alone across all phenotypes. Grouping the \u003cem\u003eMild\u003c/em\u003e phenotype with the \u003cem\u003eCombined\u003c/em\u003e phenotype results in the most substantial improvement for the \u003cem\u003eMild\u003c/em\u003e Type.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e3\u003c/strong\u003e The performance of the DL models across the four phenotypes trained with specific (\u003cem\u003epheno_spec_DL2\u003c/em\u003e) and contrastive phenotypes (\u003cem\u003econtrast_pheno_DL\u003c/em\u003e)\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"567\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.631393298059965%\" rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC-ROC /AUC-PRC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.34215167548501%\" rowspan=\"2\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003ePheno_spec_DL2\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"55.026455026455025%\" colspan=\"3\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u003cem\u003eContrast_pheno_DL\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"32.79742765273312%\"\u003e\n \u003cp\u003eMild + Sleep\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"39.87138263665595%\"\u003e\n \u003cp\u003eMild + Respiratory\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"27.331189710610932%\"\u003e\n \u003cp\u003eMild + Combined\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.67844522968198%\"\u003e\n \u003cp\u003eMild\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.374558303886925%\"\u003e\n \u003cp\u003e0.750/0.534\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.021201413427562%\"\u003e\n \u003cp\u003e0.785/0.590\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.908127208480565%\"\u003e\n \u003cp\u003e0.833/0.619\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.833/0.689\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.67844522968198%\"\u003e\n \u003cp\u003eSleep related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.374558303886925%\"\u003e\n \u003cp\u003e0.915/0.808\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.021201413427562%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.915/0.810\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.908127208480565%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.67844522968198%\"\u003e\n \u003cp\u003eRespiratory related\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.374558303886925%\"\u003e\n \u003cp\u003e0.732/0.350\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.021201413427562%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.908127208480565%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.854/0.567\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"26.67844522968198%\"\u003e\n \u003cp\u003eCombined\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.374558303886925%\"\u003e\n \u003cp\u003e0.958/0.600\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.021201413427562%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.908127208480565%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.979/0.750\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003ch2\u003e3.5 Feature importance analysis\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eAs shown in\u0026nbsp;Fig.5, in addition to the traditional CVD risk features, PSG and FOOD FREQUENCY were recognized as very important features for all four phenotypes. Moreover, each phenotype placed emphasis on different additional features. Specifically, the\u003cem\u003e\u0026nbsp;Sleep Related\u003c/em\u003e phenotype particularly emphasized features reflecting features related to sleep status acqruied in the sleep questionnaire such as sleep duration, sleep efficiency standard deviation, etc. The \u003cem\u003eCombined\u0026nbsp;\u003c/em\u003ephenotype gave significant emphasis on CARDIAC CT features.\u003c/p\u003e\n\u003cp\u003eAlthough all four phenotypes emphasized the PSG and FOOD FREQUENCY feature categories, they presented varying importance to specific PSG and FOOD FREQUENCY features. As shown in Fig.6 (a), the \u003cem\u003eMild\u0026nbsp;\u003c/em\u003ephenotype gave more emphasis on breathing disturbance features such as AHI. The \u003cem\u003eSleep Related\u0026nbsp;\u003c/em\u003ephenotype focused on features related to sleep structure disorders, such as sleep efficiency and sleep duration. The \u003cem\u003eRespiratory\u0026nbsp;\u003c/em\u003ephenotype strongly emphasized blood oxygen saturation and AHI levels. The \u003cem\u003eCombined\u003c/em\u003e phenotype specifically highlighted autonomic disorders features such as arousal and PLMS levels.\u003c/p\u003e\n\u003cp\u003eAs shown in Fig.6 (b), four phenotypes gave different emphasis to distinct FOOD FREQUENCY categories. Overall, the frequency of sweet food consumption is a relatively important risk factor for all phenotypes, especially for \u003cem\u003eSleep\u0026nbsp;\u003c/em\u003eand\u003cem\u003e\u0026nbsp;Combined\u0026nbsp;\u003c/em\u003ephenotypes. These two phenotypes all gave additional significant importance to fruits. \u003cem\u003eSleep\u0026nbsp;\u003c/em\u003eand\u003cem\u003e\u0026nbsp;Combined\u0026nbsp;\u003c/em\u003ephenotypic populations should pay special attention to sugar-controlled diets, and have more fruit intake. On the other hand, the importance distribution of foods in the \u003cem\u003eMild\u003c/em\u003e and \u003cem\u003eRespiratory Related\u003c/em\u003e phenotypes is relatively similar, and the two groups should reduce sweets consumption and combine more grains and fruits intake. Fig.7 further shows the top-five important foods on cardiovascular risk of the four phenotypes, which can serve as recommendations for daily dietary management.\u003c/p\u003e"},{"header":"4. DISCUSSION","content":"\u003cp\u003eRecent research has examined the association between OSA and CVDs, including phenotyping OSA using PSG static features and analyzing their differences in cardiovascular risks [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. However, it remains to be explored on how to effectively using the OSA-related information for accurate CVD risk prediction modeling. Addressing this gap, this study built several ML and DL models under various OSA phenotyping integration strategies. Additionally, the study examined the value of integrating deep representations of sleep-event feature sequences on CVD risk prediction. The findings indicate that the approach based on OSA phenotyping and integrating deep representations of overnight sleep-event feature sequences, yield the most optimal performance in CVD risk prediction.\u003c/p\u003e \u003cp\u003eRegardless of the ML models employed, it is evident that performance improves significantly after incorporating phenotyping. Specifically, while the \u003cem\u003ePheno_fused_ML\u003c/em\u003e model, which integrates OSA phenotyping directly as a feature, shows some effectiveness. But, it is surpassed by the four phenotypes that were modeled separately (\u003cem\u003ePheno_Spec_ML1, Pheno_Spec_ML2\u003c/em\u003e and \u003cem\u003ePheno_Spec_ML3\u003c/em\u003e). This distinction could be attributed to the inability of \u003cem\u003ePheno_fused_ML\u003c/em\u003e model to adequately learn the weights of the phenotype feature amidst a multitude of other features. Furthermore, selecting phenotype-specific PSG static features to build the model further enhances the model performance (\u003cem\u003ePheno_Spec_ML2\u003c/em\u003e vs \u003cem\u003ePheno_Spec_ML3\u003c/em\u003e), a trend that was also observed in DL models (\u003cem\u003ePheno_spec_DL2\u003c/em\u003e vs \u003cem\u003ePheno_spec_DL1\u003c/em\u003e). The possible reason for this improved performance is that, features relevant to other phenotypes may introduce confounding factors in the risk prediction for a specific phenotype. By selecting PSG features tailored to specific phenotypes, models can more accurately learn and understand the relationship between these features and CVD risk within the context of that phenotype.\u003c/p\u003e \u003cp\u003eThe risk prediction accuracy was further improved by adding the deep features from sleep-event sequences. In addition, we also found that the \u003cem\u003eContrast_spec_DL\u003c/em\u003e model achieved optimal performance by strategically combining one of the three OSA phenotypes (\u003cem\u003eSleep-related, Breathing-related and Combined\u003c/em\u003e) with the \u003cem\u003eMild\u003c/em\u003e phenotype, using all PSG static features and sleep event feature sequences as inputs, compared to a training strategy (i.e., \u003cem\u003ePheno_spec_DL2\u003c/em\u003e) that incorporated phenotype-specific sequence features. This approach led to a notable increase in the AUC-ROC value to 0.877, and improved predictive accuracy for each phenotype compared to their individual performances. This finding indicates that while modeling within a single phenotype generally yields better results than non-phenotyping approaches, strategic combination of samples from different phenotypes of different risk levels can enhance risk prediction for both groups. A possible explanation for this improvement is the relatively lower and less varied risk profiles within the \u003cem\u003eMild\u003c/em\u003e phenotype. When these were combined with a phenotype characterized by a broader variance in risk distribution, the model's capacity to discern relevant risk features was enhanced. Consequently, this study observed a significant improvement in the predictive performance for the \u003cem\u003eMild\u003c/em\u003e phenotype, especially when combined with the \u003cem\u003eCombined\u003c/em\u003e phenotype, as indicated by a HR ranging from 1.566 to 6.424.\u003c/p\u003e \u003cp\u003eWhile previous studies have utilized ML or DL techniques to process short-term physiological sequences, focusing primarily on predicting specific CVD events [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan additionalcitationids=\"CR39 CR40\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], this study first leveraged the temporal information from overnight sleep-event feature sequences through deep LSTM networks. It provides a more comprehensive set of features for CVD risk prediction than only using averaged static features. Given the objective to predict CVD risk over the next five years, the study recognized the highly complex temporal relationships exhibited by multiple physiological signals throughout the entire sleep period. Directly using these signals as model inputs might fail to capture essential information pertinent to long-term CVD risk. Consequently, the study focused on extracting temporal features from five key sleep-event feature sequences (arousal, hypopnea, apnea, hypoxemia and PLMS) over the full night.\u003c/p\u003e \u003cp\u003eUnlike previous studies that relied on manual selection of relevant risk factors for predictive modeling [\u003cspan additionalcitationids=\"CR36\" citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], this study employed a comprehensive feature selection process across all static features encompassing multi-dimensional aspects, including traditional risks factors, PSG-based multifaced factors, imaging markers and lifestyle factors, etc. This approach has the potential to lead to more integrated and precise predictions of CVD risk in clinical practice. Moreover, as far as we know, this is the first study to rank the importance of features among a comprehensive categories of cardiovascular risk features. One important finding is that, the four OSA phenotypes all gave particular emphasis on PSG and FOOD FREQUENCY features. This is consistent with existing understanding on the importance of PSG features for CVD risk prediction in the literature [\u003cspan additionalcitationids=\"CR43\" citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Earlier studies also highlighted the close association between dietary habits and CVD risk [\u003cspan additionalcitationids=\"CR46 CR47\" citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. Through the comprehensive analysis of feature importance in this study, the importance of dietary habits in predicting CVD risk should be further emphasized in the general population. Additionally, each phenotype emphasized distinct PSG and FOOD FREQUENCY features. These findings would enable more precisive CVD risk evaluation and management for different OSA phenotypes.\u003c/p\u003e \u003cp\u003eDespite these promising results, the study has some limitations. It utilized data from only 1,874 participants from the MESA dataset, and only 187 subjects were used in the testing set. The generalization capability of the model needs further validation on larger datasets. Additionally, the model depended on sleep-event annotation based on the PSG data, which was manually labeled, a resource-intensive and time-consuming process. However, considering ongoing research efforts in developing automated sleep staging and event detection algorithms [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e], future work could leverage fully automated algorithms to provide the input features required for this model. Future studies may also investigate the minimally effective set of features that maintains the performance of CVD risk prediction with the strategies proposed in this study.\u003c/p\u003e"},{"header":"5. CONCLUSIONS","content":"\u003cp\u003eThis study introduces a new approach for CVD risk prediction, which integrates deep features learned from overnight sleep-event feature sequences. Additionally, the study validated the superior performance of OSA phenotype-specific models over phenotype-agnostic ones, and also introduces a training method that strategically combines the \u003cem\u003eMild\u003c/em\u003e phenotype with another OSA phenotypic population (\u003cem\u003eRespiratory\u003c/em\u003e, \u003cem\u003eSleep\u003c/em\u003e or \u003cem\u003eCombined\u003c/em\u003e) for contrastive training. The contrastive phenotype-specific model with deep features achieved an accuracy of 96.6% and an AUC-ROC value of 87.7% in predicting CVD risk over five years in general population without historical CVDs. Moreover, the development of phenotype-aware predictive models provided valuable insights into key risk features. The model placed a significant emphasis on lifestyle-related features such as the sleep factors indicated by PSG and food habits, over traditional and other risk factors for predicting long-term CVD outcomes in the general population. Furthermore, each of the four phenotypes emphasized distinct features, which may pave the way for precise risk management strategies tailored for different OSA phenotypic populations. Future research should validate these findings on additional datasets, and explore the utility of mobile and wearable devices for regularly collecting physiological and lifestyle data over extended periods, which could offer a more accurate representation of lifestyle information, potentially providing early warning of CV risks.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthics approval and consent to participate\u003c/h2\u003e\n\u003cp\u003eThe dataset employed in this study originates from the Multi-Ethnic Study of Atherosclerosis (MESA), with informed consent obtained from all participants in the MESA study. This research was conducted under the approval of the Shenzhen Technology University Ethics Committee (No: SZTUEA-20225011, date: 18 May 2022).\u003c/p\u003e\n\u003ch2\u003eConsent for publication\u003c/h2\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003ch2\u003eAvailability of data and materials\u003c/h2\u003e\n\u003cp\u003eThe MESA datasets request should be directed to https://biolincc.nhlbi.nih.gov/login/?next=/requests/data-request/10761/view/. The MESA Sleep Ancillary study was funded by NIH-NHLBI Association of Sleep Disorders with Cardiovascular Health Across Ethnic Groups (RO1 HL098433). MESA is supported by NHLBI funded contracts HHSN268201500003I, N01-HC-95159, N01-HC-95160, N01-HC-95161, N01-HC-95162, N01-HC-95163, N01-HC-95164, N01-HC-95165, N01-HC-95166, N01-HC-95167, N01-HC-95168 and N01-HC-95169 from the National Heart, Lung, and Blood Institute, and by cooperative agreements UL1-TR-000040, UL1-TR-001079, and UL1-TR-001420 funded by NCATS. The National Sleep Research Resource was supported by the National Heart, Lung, and Blood Institute (R24 HL114473, 75N92019R002). This manuscript was not prepared in collaboration with MESA in\u0026shy;vestigators and does not necessarily reflect the opinions or views of MESA, or the NHLBI.\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis work was supported by Guangdong Basic and Applied Basic Research Foundation [2021A1515110025], Young Scientists Fund from National Natural Science Foundation of China (NSFC) [62301333], Research Foundation of Education Department of Guangdong Province [2022ZDJS115] and the Common University Innovation Team Project of Guangdong [2021KCXTD041].\u003c/p\u003e\n\u003ch2\u003eAuthors\u0026rsquo; contributions\u003c/h2\u003e\n\u003cp\u003eConceptualization, Yali Zheng; methodology, Zhengbi Song, Yali Zheng; validation and formal analysis, Zhengbi Song; investigation, Xiao Peng, Bo Cheng; data curation, Zhengbi Song, Yu Huang; writing\u0026mdash;original draft preparation, Zhengbi Song; writing\u0026mdash;review and editing, Yali Zheng, Min Min; visualization, Zhengbi Song; supervision, Yali Zheng, Min Min.; funding acquisition, Yali Zheng. All authors have read and agreed to the published version of the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eAcknowledgements\u003c/h2\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eWorld Health Organization[EB/OL]. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds\u003c/span\u003e\u003cspan address=\"https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKannel WB, Mcgee DL. Diabetes and Cardiovascular Disease: The Framingham Study[J]. JAMA. 1979;241(19):2035\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eConroy RM, Py\u0026ouml;r\u0026auml;l\u0026auml; K, Fitzgerald AP, et al. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the SCORE project[J]. Eur Heart J. 2003;24(11):987\u0026ndash;1003.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehilli J, Kastrati A, Dirschinger J, et al. Sex-based analysis of outcome in patients with acute myocardial infarction treated predominantly with percutaneous coronary intervention[J]. Volume 287. JAMA-JOURNAL OF THE AMERICAN MEDICAL ASSOCIATION; 2002. pp. 210\u0026ndash;5. 2.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohd Faizal AS, Thevarajah TM, Khor SM, et al. A review of risk prediction models in cardiovascular disease: conventional approach vs. artificial intelligent approach[J]. Comput Methods Programs Biomed. 2021;207:106190.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThan MP, Pickering JW, Sandoval Y, et al. Machine Learning to Predict the Likelihood of Acute Myocardial Infarction[J]. Circulation. 2019;140(11):899\u0026ndash;909.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePiros P, Ferenci T, Fleiner R, et al. Comparing machine learning and regression models for mortality prediction based on the Hungarian Myocardial Infarction Registry[J]. Knowl Based Syst. 2019;179:1\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSteele AJ, Cakiroglu SA, Shah AD et al. Machine learning models in electronic health records can outperform conventional survival models for predicting patient mortality in coronary artery disease[J]. bioRxiv, 2018: 256008.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWallert J, Tomasoni M, Madison G et al. Predicting two-year survival versus non-survival after first myocardial infarction using machine learning and Swedish national register data[J]. BMC Med Inf Decis Mak, 2017, 17(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWeng SF, Reps J, Kai J, et al. Can machine-learning improve cardiovascular risk prediction using routine clinical data?[J]. PLoS ONE. 2017;12(4):e0174944.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoff DC, Lloyd-Jones DM, Bennett G, et al. 2013 ACC/AHA Guideline on the Assessment of Cardiovascular Risk[J]. Circulation. 2014;129(25suppl2):S49\u0026ndash;73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJavaheri S, Barbe F, Campos-Rodriguez F, et al. Sleep Apnea: Types, Mechanisms, and Clinical Cardiovascular Consequences[J]. J Am Coll Cardiol. 2017;69(7):841\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStrassberger C, Zou D, Penzel T, et al. Beyond the AHI-pulse wave analysis during sleep for recognition of cardiovascular risk in sleep apnea patients[J]. J Sleep Res. 2021;30(6):e13364.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim JW, Won TB, Rhee CS, et al. Polysomnographic phenotyping of obstructive sleep apnea and its implications in mortality in Korea[J]. Sci Rep. 2020;10(1):13207.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMazzotti DR, Keenan BT, Lim DC, et al. Symptom Subtypes of Obstructive Sleep Apnea Predict Incidence of Cardiovascular Outcomes[J]. Am J Respir Crit Care Med. 2019;200(4):493\u0026ndash;506.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZinchuk AV, Jeon S, Koo BB, et al. Polysomnographic phenotypes and their cardiovascular implications in obstructive sleep apnoea[J]. Thorax. 2018;73(5):472\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLacedonia D, Carpagnano GE, Sabato R, et al. Characterization of obstructive sleep apnea-hypopnea syndrome (OSA) population by means of cluster analysis[J]. J Sleep Res. 2016;25(6):724\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahrbabaki SS, Linz D, Hartmann S, et al. Sleep arousal burden is associated with long-term all-cause and cardiovascular mortality in 8001 community-dwelling older men and women[J]. Eur Heart J. 2021;42(21):2088\u0026ndash;99.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAzarbarzin A, Sands SA, Stone KL, et al. The hypoxic burden of sleep apnoea predicts cardiovascular disease-related mortality: the Osteoporotic Fractures in Men Study and the Sleep Heart Health Study[J]. Eur Heart J. 2019;40(14):1149\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKendzerska T, Mollayeva T, Gershon AS, et al. Untreated obstructive sleep apnea and the risk for serious long-term adverse outcomes: A systematic review[J]. Sleep Med Rev. 2014;18(1):49\u0026ndash;59.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeung RST, Comondore VR, Ryan CM, et al. Mechanisms of sleep-disordered breathing: causes and consequences[J]. Pfl\u0026uuml;gers Archiv - Eur J Physiol. 2011;463(1):213\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDami S, Yahaghizadeh M. Predicting cardiovascular events with deep learning approach in the context of the internet of things[J]. Volume 33. Neural Computing \u0026amp; Applications; 2021. pp. 7979\u0026ndash;96. 13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen H, Lundberg SM, Erion G, et al. Forecasting adverse surgical events using self-supervised transfer learning for physiological signals[J]. NPJ Digit Med. 2021;4(1):167.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSbrollini A, De Jongh MC, Ter Haar CC et al. Serial electrocardiography to detect newly emerging or aggravating cardiac pathology: A deep\u0026ndash;learning approach[J]. Biomed Eng Online, 2019, 18(15).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRaita Y, Goto T, Faridi MK et al. Emergency department triage prediction of clinical outcomes using machine learning models[J]. Crit Care, 2019, 23(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKaji DA, Zech JR, Kim JS, et al. An attention based deep learning model of clinical events in the intensive care unit[J]. PLoS ONE. 2019;14(2):e0211057.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarrett LA, Payrovnaziri SN, Bian J et al. Building computational models to predict one-year mortality in ICU patients with acute myocardial infarction and post myocardial infarction syndrome[J]. AMIA Jt Summits Transl Sci Proc, 2019, 2019: 407\u0026ndash;416.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang GQ, Cui L, Mueller R, et al. The National Sleep Research Resource: towards a sleep data commons[J]. J Am Med Inf Assoc. 2018;25(10):1351\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X, Wang R, Zee P, et al. Racial/Ethnic Differences in Sleep Disturbances: The Multi-Ethnic Study of Atherosclerosis (MESA)[J]. Sleep. 2015;38(6):877\u0026ndash;88.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLundberg SM, Lee S-I. A unified approach to interpreting model predictions[J]. Adv Neural Inf Process Syst, 2017, 30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShlomai G, Shemesh J, Segev S et al. The Multi-Ethnic Study of Atherosclerosis-Calcium Score Improves Statin Treatment Allocation in Asymptomatic Adults[J]. Front Cardiovasc Med, 2022, 9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehta R, Ning HY, Bansal N, et al. Ten-Year Risk-Prediction Equations for Incident Heart Failure Hospitalizations in Chronic Kidney Disease: Findings from the Chronic Renal Insufficiency Cohort Study and the Multi- Ethnic Study of Atherosclerosis[J]. J Card Fail. 2022;28(4):540\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHathaway QA, Yanamala N, Budoff MJ, et al. Deep neural survival networks for cardiovascular risk prediction: The Multi-Ethnic Study of Atherosclerosis (MESA)[J]. Comput Biol Med. 2021;139:104983.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYeghiazarians Y, Jneid H, Tietjens JR, et al. Obstructive Sleep Apnea and Cardiovascular Disease: A Scientific Statement From the American Heart Association[J]. Circulation. 2021;144(3):e56\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaki N, Babaahmadi-Rezaei H, Rahimi Z et al. Impact of modifiable risk factors on prediction of 10-year cardiovascular disease utilizing framingham risk score in Southwest Iran[J]. BMC Cardiovasc Disord, 2023, 23(1).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng M, Hou F, Cheng Z, et al. A Cardiovascular Disease Risk Score Model Based on High Contribution Characteristics[J]. Appl Sci. 2023;13(2):893.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark S, Kim YG, Ann SH et al. Prediction of the 10-year risk of atherosclerotic cardiovascular disease in the Korean population[J]. Epidemiol Health, 2023, 45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharma LD, Sunkaria RK. Inferior myocardial infarction detection using stationary wavelet transform and machine learning approach[J]. SIViP. 2018;12(2):199\u0026ndash;206.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLui HW, Chow KL. Multiclass classification of myocardial infarction with convolutional and recurrent neural networks for portable ECG devices[J]. Inf Med Unlocked. 2018;13:26\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLodhi AM, Qureshi AN, Sharif U et al. A novel approach using voting from ecg leads to detect myocardial infarction[C]. Adv Intell Syst Comput, 2018: 337\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDohare AK, Kumar V, Kumar R. Detection of myocardial infarction in 12 lead ECG using support vector machine[J]. Appl Soft Comput. 2018;64:138\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRundo JV, Downey R 3. Polysomnography[J] Handb Clin Neurol. 2019;160:381\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCuellar NG. The effects of periodic limb movements in sleep (PLMS) on cardiovascular disease[J]. Heart Lung. 2013;42(5):353\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSabanayagam C, Shankar A. Sleep duration and cardiovascular disease: results from the National Health Interview Survey[J]. Sleep. 2010;33(8):1037\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoutentakis M, Surma S, Rogula S, et al. The Effect of a Vegan Diet on the Cardiovascular System[J]. J Cardiovasc Dev Disease. 2023;10(3):94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDyńka D, Kowalcze K, Charuta A, et al. The Ketogenic Diet and Cardiovascular Diseases[J]. Nutrients. 2023;15(15):3368.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu E, Malik VS, Hu FB. Cardiovascular Disease Prevention by Diet Modification\u0026thinsp;\u0026lt;\u0026thinsp;i\u0026thinsp;\u0026gt;\u0026thinsp;JACC\u0026thinsp;Health Promotion Series[J]. J Am Coll Cardiol. 2018;72(8):914\u0026ndash;26.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePan A, Lin X, Hemler E, et al. Diet and Cardiovascular Disease: Advances and Challenges in Population-Based Studies[J]. Cell Metab. 2018;27(3):489\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing FH, Cotton-Clay A, Fava L, et al. Polysomnographic validation of an under-mattress monitoring device in estimating sleep architecture and obstructive sleep apnea in adults[J]. Sleep Med. 2022;96:20\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassan AR, Bhuiyan MIH. An automated method for sleep staging from EEG signals using normal inverse Gaussian parameters and adaptive boosting[J]. Neurocomputing. 2017;219:76\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Cardiovascular risk prediction, obstructive sleep apnea phenotyping, deep representation, sleep event sequences, phenotype-aware models, model interpretability","lastPublishedDoi":"10.21203/rs.3.rs-4084889/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4084889/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e: Advances in mobile, wearable and machine learning (ML) technologies for gathering and analyzing long-term health data have opened up new possibilities for predicting and preventing cardiovascular diseases (CVDs). Meanwhile, the association between obstructive sleep apnea (OSA) and CV risk has been well-recognized. This study seeks to explore effective strategies of incorporating OSA phenotypic information and overnight physiological information for precise CV risk prediction in the general population.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e: 1,874 participants without a history of CVDs from the MESA dataset were included for the 5-year CV risk prediction. Four OSA phenotypes were first identified by the K-mean clustering based on static polysomnographic (PSG) features. Then several phenotype-agnostic and phenotype-specific ML models, along with deep learning (DL) models that integrate deep representations of overnight sleep-event feature sequences, were built for CV risk prediction. Finally, feature importance analysis was conducted by calculating SHapley Additive exPlanations (SHAP) values for all features across the four phenotypes to provide model interpretability.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e: All ML models showed improved performance after incorporating the OSA phenotypic information. The DL model trained with the proposed phenotype-contrastive training strategy performed the best, achieving an area under the Receiver Operating Characteristic (ROC) curve of 0.877. Moreover, PSG and FOOD FREQUENCY features were recognized as significant CV risk factors across all phenotypes, with each phenotype emphasizing unique features.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e: Models that are aware of OSA phenotypes are preferred, and lifestyle factors should be a greater focus for precise CV prevention and risk management in the general population.\u003c/p\u003e","manuscriptTitle":"Integrating Phenotypic Information of Obstructive Sleep Apnea and Deep Representation of Sleep-Event Sequences for Cardiovascular Risk Prediction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-03-15 20:27:39","doi":"10.21203/rs.3.rs-4084889/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-09-06T11:32:57+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-23T12:41:21+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-11T03:29:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"212673512338194710926023105094126776899","date":"2024-07-07T18:35:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"192630262918648723239706843700845390084","date":"2024-07-05T06:33:57+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-03-27T07:28:22+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-03-13T11:45:06+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-03-13T11:41:20+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2024-03-12T12:49:11+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"29226da0-4106-497d-bf26-92b4647d858f","owner":[],"postedDate":"March 15th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-23T16:00:43+00:00","versionOfRecord":{"articleIdentity":"rs-4084889","link":"https://doi.org/10.1186/s12911-026-03439-8","journal":{"identity":"bmc-medical-informatics-and-decision-making","isVorOnly":false,"title":"BMC Medical Informatics and Decision Making"},"publishedOn":"2026-03-16 15:57:25","publishedOnDateReadable":"March 16th, 2026"},"versionCreatedAt":"2024-03-15 20:27:39","video":"","vorDoi":"10.1186/s12911-026-03439-8","vorDoiUrl":"https://doi.org/10.1186/s12911-026-03439-8","workflowStages":[]},"version":"v1","identity":"rs-4084889","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4084889","identity":"rs-4084889","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}