Machine learning for scalable obstructive sleep apnea risk screening using digital phenotyping from wearable devices and clinical scales

Abstract

Obstructive sleep apnea (OSA) is common yet frequently underdiagnosed, partly because overnight polysomnography (PSG) is logistically burdensome and access to specialized testing is limited. We aimed to develop machine-learning models for OSA risk screening using multimodal digital phenotyping from consumer-grade wearable devices, smartphone-based assessments, and clinical scales. We enrolled 338 participants and collected data over four weeks. After preprocessing, 107 features were derived from wearable-derived physiological and activity measures, smartphone-based records, and questionnaire-based clinical risk profiles, and used to classify high- versus low-risk OSA groups defined by the Berlin Questionnaire. Across multiple model configurations, predictive performance was high, with the best-performing model achieving an AUC of up to 0.94 and an F1 score of 0.80 in the internal validation set. Consistently influential predictors included body mass index, Insomnia Severity Index score, Smartphone Overuse Screening Questionnaire score, resting heart rate, and heart rate recovery. These findings suggest that multimodal digital phenotyping from accessible consumer technologies may support scalable pre-screening for OSA risk in real-world settings. Further validation against PSG-confirmed OSA outcomes is needed.Trial Registration: Clinical Research Information Service (CRIS) KCT0009175 (Registration data: Feb-15-2024) (https://cris.nih.go.kr/cris/search/detailSearch.do?search_lang=E&focus=reset_12&search_page=M&pageSize=10&page=undefined&seq=26133&status=5&seq_group=26133).

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-58772-9.

Introduction

Obstructive sleep apnea (OSA) is a highly prevalent and often underdiagnosed chronic sleep disorder, characterized by recurrent episodes of upper airway obstruction during sleep, leading to intermittent hypoxia, sleep fragmentation, and significant long-term health consequences, including increased cardiovascular morbidity and mortality¹. Despite its widespread impact, a substantial proportion of patients with OSA remain undiagnosed, primarily because of the logistical and practical limitations of the current gold-standard diagnostic tool, overnight polysomnography (PSG)². PSG is resource-intensive and requires specialized sleep laboratories, trained personnel, and considerable time and costs, making it inaccessible for population-level screening and routine clinical practice. Epidemiological analyses estimate that 80–90% of adults with clinically significant OSA have never been formally diagnosed³. Furthermore, recent global assessments suggest that nearly one billion individuals worldwide may have some degree of OSA, underscoring the urgent need for more scalable screening approaches⁴.

Consequently, the research focus has shifted to identifying alternative, scalable, cost-effective, and noninvasive approaches for OSA risk detection and screening. In this regard, digital phenotyping derived from wearable devices and smartphone applications offers continuous monitoring of behavioral and physiological parameters for real-world health assessment^5–7. Recent advancements in wearable technology and smartphone applications have enabled the collection of rich, longitudinal data, providing unprecedented opportunities to derive digital biomarkers indicative of various health conditions^8,9. Previous studies have also demonstrated the utility of wearable-derived data, such as accelerometry, peripheral oxygen saturation (SpO2) sensors, and photoplethysmography (PPG) signals, for inferring sleep architecture and detecting sleep-disordered breathing using machine learning techniques¹⁰. These approaches have shown promising results in OSA screening using smartbands¹¹, optical sensors, and multimodal signals from smartphones and smartwatches¹².

However, many existing studies relied on specialized medical devices, focused on single data modalities, or utilized complex sensor arrays that may not be readily available on consumer-grade devices. Therefore, this study aimed to evaluate whether integrated digital phenotyping using widely accessible consumer-grade wearable devices, smartphone-based self-reports, and clinical scales can be used to predict OSA risk. Importantly, the aim of this study was not to establish a definitive diagnosis of OSA without PSG, but to investigate whether multimodal digital phenotypes collected longitudinally in daily life could approximate a clinically meaningful OSA risk state captured by an established screening instrument. In this context, digital phenotyping may serve as a scalable pre-screening or triage-support approach rather than a substitute for formal diagnostic testing.

Specifically, we sought to (1) identify behavioral and physiological features derived from wearable devices, smartphone applications, and self-reported clinical scales that are significantly associated with high OSA risk, as classified by the Berlin Questionnaire (BQ); (2) assess the predictive utility of these features using advanced machine learning models (XGBoost, LightGBM, Random Forest, and XGBRF); and (3) interpret the contributions of individual variables to model predictions using explainable AI techniques (SHapley Additive exPlanations, SHAP), thereby providing insights into the underlying factors driving OSA risk. In contrast to prior studies, we incorporated a wide range of passive and active monitoring signals collected continuously over a four-week period. This study provides a comprehensive and ecologically valid design for exploring novel digital biomarkers of OSA risk in real-world settings, potentially complementing conventional diagnostic workflows and bridging the diagnostic gap in sleep healthcare.

Methods

Data collection

Participants aged 17–70 are initially recruited from the Korea University Anam Hospital (Seoul, South Korea) and Datamaker (Daejeon, South Korea) via online platforms and university bulletin boards between January 2023 and July 2024, ultimately collecting valid data from 338 individuals aged 18–65 after filtering out data recording issues. Although the broader research objective involved comprehensive sleep-related phenotyping, the current study specifically focused on OSA risk. Based on the Insomnia Severity Index (ISI)¹³, participants experiencing subjective insomnia symptoms at least three times per week in the past 3 months were grouped into the moderate-to-severe insomnia group (ISI score 15), mild insomnia group (ISI score 8 and < 15), or normal sleep group (ISI score < 8)¹⁴. Exclusion criteria included a diagnosis of major psychiatric or neurological disorders, current treatment for sleep-related conditions, and inability or unwillingness to use a smartphone for data collection.

Data on baseline demographics, psychiatric and medical histories, and current medications were collected. Core data collection spanned a four-week period and integrated three primary sources: a consumer-grade wearable device, smartphone-based ecological momentary assessment (EMA), and self-reported clinical questionnaires. Each participant was provided with a Fitbit Inspire 3 device (Google, USA), which continuously recorded the heart rate, cumulative step count, moving distance, and exercise time every 5 min. Each participant wore a Fitbit Inspire 3 (Google, USA), which continuously recorded heart rate, cumulative step count, distance, and exercise time at 5-minute intervals, and also provided daily sleep metrics including total sleep time, sleep stage durations, awake time, and number of awakenings. Concurrently, participants used a mobile EMA application (SOMDAY, Lumanlab Inc., South Korea) on their smartphones for daily self-reported behavioral logs, such as caffeine, alcohol, and food intakes, stress levels, and nap time. Additionally, a comprehensive set of self-reported clinical questionnaires were administered to assess various sleep-related and psychological factors, including the ISI¹³, Dysfunctional Beliefs and Attitudes about Sleep (DBAS-16)¹⁵, Alcohol Use Disorder Identification Test Korean version (AUDIT-K)¹⁶, State-Trait Anxiety Inventory State-anxiety (STAI-S)¹⁷, International Restless Legs Syndrome Study Group Rating Scale (IRLS)¹⁸, Korean version of Biological Rhythms Interview of Assessment in Neuropsychiatry (K-BRIAN)¹⁹, Patient Health Questionnaire-9 (PHQ-9)²⁰, Generalized Anxiety Disorder-7 (GAD-7)²¹, and Smartphone Overuse Screening Questionnaire (SOS-Q)²². Finally, OSA risk was assessed using the Berlin Questionnaire (BQ)²³, which classified participants into high-risk (n = 85) and low-risk (n = 253) groups based on the established BQ criteria. The integration of these diverse data sources allowed for a holistic digital phenotyping approach to capture the physiological, behavioral, and psychological aspects relevant to OSA risk.

Ethical considerations

Ethical approval was obtained from the Institutional Review Board of Korea University Anam Hospital (IRB No. 2022AN0587). The study adhered to the principles of the Declaration of Helsinki. All participants provided written informed consent after receiving a full explanation of the study procedures. To safeguard privacy, personally identifiable information was anonymized. Participants received approximately $80 as compensation for their time.

Machine learning methodology and experimental design

Digital phenotyping and feature extraction

Digital phenotyping data collected from wearable devices over a four-week period were subjected to rigorous preprocessing. Heart rate, steps, distance, and exercise time were recorded every 5 min, with exercise time, distance, and step count stored as cumulative values. Meanwhile, sleep parameters were recorded daily. Missing values for the heart rate data were imputed using the moving average method. Missing values for cumulative variables, such as step count, distance traveled, and exercise duration, were interpolated by distributing the difference between adjacent cumulative records across the number of missing intervals, with small random noise added to avoid artificial uniformity. The periods during which the wearable device was completely removed were excluded from the analysis. The data were grouped into 1-h intervals and averaged for the entire recording period according to day conditions (daytime/nighttime, weekdays/holidays) to smooth out outliers during feature extraction.

Feature extraction from wearable device data involved daily statistics, cosinor analysis, and activity parameters. Cosinor analysis was applied to heart rate and step data using a 2-day time window to estimate circadian rhythm parameters, including midline estimating statistics of rhythm (MESOR), amplitude, acrophase, and goodness of fit²⁴. Step data were further used to compute activity parameters such as the least active 5-h period (L5), the most active 10-h period (M10), relative amplitude (RA), intradaily stability (IS), and intradaily variability (IV) to quantify circadian activity patterns. Daily statistics (maximum, minimum, and average values) for heart rate, steps, distance, and exercise time were calculated during the daytime (8:00–18:00) and nighttime (18:00–08:00) as well as separately for weekdays and holidays (including Korean public holidays and weekends). Sleep quality was calculated from sleep parameters as the ratio of sleep time to total time spent in bed (excluding awake time). Additionally, the first heart rate recovery (difference between peak heart rate and heart rate 5 min later), the second heart rate recovery (difference between heart rate 5 min after peak and heart rate 10 min later), and moderate intensity steps (steps when heart rate was ≥ 65% of maximum of heart rate, reflecting a wide range of age groups)²⁵ were calculated using heart rate and steps data. All features were processed daily and averaged over the entire recording period for each participant. Sleep parameters, daily statistics, and activity parameters were extracted separately for weekdays and weekends. Preprocessing details such as cosinor analysis and daily statistics are provided in the Supplementary Information.

Self-reported data from smartphone applications, including stress levels, caffeine and alcohol intake, and naptime, consisting of quantitative (numerical) and time-of-day (categorical) data. Naptime, a categorical variable, was encoded as follows: based on duration, 0.5 for naps < 1 h, 1 for 1–2 h, 2 for 2–3 h, 3 for > 3 h; weighted by time: 1 for morning, 1.5 for afternoon, 2 for before bedtime). Caffeine intake was weighted as 1 for morning, 1.5 for afternoon, and 2 before bedtime, and stress levels were categorized by intensity (mild = 1, moderate = 2, severe = 3) and adjusted by time of day using the same weighting scheme as caffeine. Alcohol consumption was weighted as follows: 1.5 for morning, 1 for afternoon, 2 and bedtime). Each questionnaire of the self-reported clinical scales yielded scores for the questions and categorical responses to the total scores, which were subsequently encoded.

A total of 107 features were extracted from 338 participants (Fig. 1). Missing values were imputed using the average of each group’s features. For the detailed process of preprocessing and feature extraction, see the Supplementary Information.

Fig. 1.

Feature extraction process of multimodal digital phenotypes. HR; Heart rate.

Machine learning models and experimental setup

Four tree-based machine learning models were used to predict OSA risk groups: eXtreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), Random Forest (RF), and eXtreme Gradient Boosting Random Forest (XGBRF)^26–29. A robust training and evaluation pipeline was implemented to optimize model performance.

The full dataset was divided into a training set and an internal validation set. The internal validation set consisted of 50 participants, including 25 from the low-risk OSA group and 25 from the high-risk OSA group, to provide a balanced hold-out sample for performance evaluation. During the model development, the Synthetic Minority Oversampling Technique (SMOTE)³⁰ was applied only to the training set to address the inherent class imbalance in training dataset (253 low-risk vs. 85 high-risk participants). Hyperparameter tuning was performed using Bayesian optimization³¹ in conjunction with stratified k-fold cross-validation³² to optimize the model parameters and prevent overfitting (Fig. 2).

Fig. 2.

Training and evaluation of OSA risk groups. OSA; Obstructive sleep apnea.

Final model performance was evaluated in the internal validation set using accuracy, precision, recall, F1 score, area under the receiver operating characteristic curve (AUC), and area under the precision-recall curve (PR-AUC). SHapley Additive extensions (SHAP)³³ were used to calculate the feature importance and identify the key factors influencing the trained model’s predictions to provide interpretability for the model’s decisions. The analysis was structured across three distinct data configurations to comprehensively compare the influence of different data types on model performance: (1) All data, encompassing features derived from wearable devices, smartphone applications, and self-reported questionnaires; (2) Least data, representing a minimal set of traditionally verifiable information (BMI, Sex, and Age), and (3) Selected data, which comprised features chosen from the ‘All data’ set using Mutual Information (MI)³⁴ as the feature selection method. All data processing and analyses were performed using Python version 3.12.7 using the scikit-learn 1.5.2, bayes_opt 2.0.3, shap 0.46.0, xgboost 2.1.3, lightgbm 4.5.0, scipy 1.15.1, and imblearn 0.13.0 packages.

Statistical analysis

Welch’s t-test was used to analyze the baseline characteristics of the participants³⁵, comparing the high- and low-risk OSA groups. Statistical analysis was performed to identify significant differences in demographic and feature variables between the two OSA risk groups. After Welch’s t-test, the machine-learning model was trained using the data and its performance was evaluated. All missing values were imputed using groupwise means.

Result

Baseline characteristics of participants

Table 1 presents the baseline characteristics of the participants, highlighting significant differences between the low-risk (n = 253) and high-risk (n = 85) OSA groups. Patients in the high-risk group were significantly older (t = −3.164, p < 0.05) and exhibited a higher Body Mass Index (BMI) (t = −10.712, p < 0.05) than those in the low-risk group.

Table 1.

Significantly different baseline characteristics of participants.

Variables (mean, SD)	OSA low-risk (n = 253)	OSA high-risk (n = 85)	t-value	df	p-value
Demographic information
BMI	21.26 (2.71)	25.57 (3.36)	−10.71	122.72	< 0.01
Age	28.82 (9.26)	32.72 (8.63)	−3.16	135.61	< 0.01
Sex (male, %)	70 (27.67)	49 (57.65)
Heartrate from wearable device
HR_CR_mesorc	74.25 (7.12)	76.38 (6.77)	−2.46	150.41	0.02
mean_HR_week_daya, b	79.9 (8.09)	81.98 (7.69)	−2.1	148.03	0.04
mean_HR_week_nighta, b	71.14 (7.46)	73.13 (7.21)	−2.16	148.36	0.03
mean_HR_holiday_nighta, b	71.376 (8.27)	73.47 (8.4)	−1.99	142.26	0.049
HR_second_recovery_phase	9.15 (3.78)	7.87 (3.52)	2.83	153.56	0.005
Steps from wearable device
RA_weekb	0.92 (0.1)	0.94 (0.06)	−2.13	249.16	0.03
Sleep parameters from wearable device
REM_holidayb	225.48 (50.12)	212.83 (49.92)	2.0	145.62	0.048
Self-reported questionnaire
ISI
ISI_score	11.18 (6.51)	13.99 (5.23)	−4.01	178.08	< 0.01
SOS-Q
SOS-Q_score	56.9 (19.46)	50.68 (19.16)	2.58	146.48	0.011

^aDay means daytime (8:00–18:00), and night means nighttime (18:00 ~ 8:00).

^bWeek means weekdays other than holidays, holiday means holidays including weekends and Koreanpublic holidays.

^cThese parameters were generated using cosinor analysis to reflect the circadian rhythm. Detailed descriptions are provided in Supplementary Information.

OSA; Obstructive Sleep Apnea, BMI; Body Mass Index, REM; Rapid Eye Movement, CR; Circadian Rhythm, L5; the least active 5-hour period, M10; the most active 10-h period, RA; Relative Amplitude, IS; Intradaily Stability, IV; Intradaily Variability, HR; Heart Rate, MESOR; Midline Estimating Statistic Of Rhythm. IRLS, International Restless Legs Syndrome Study Group Rating Scale; K-BRIAN, Korean version of the Biological Rhythms Interview of Assessment in Neuropsychiatry; PHQ, Patient Health Questionnaire; GAD, Generalized Anxiety Disorder; SOS-Q, Smartphone Overuse Screening Questionnaire.

Regarding heart rate data derived from wearable device, the mesor (HR_CR_mesor) of the high-risk group was significantly higher than that of the low-risk group (t = −2.46, p = 0.02). Daily heart rate statistics also showed significant differences; the mean heart rate during the daytime on weekdays (mean_HR_week_day) (t = −2.10, p = 0.04) and the mean heart rate during nighttime on weekdays (mean_HR_week_night) (t = −2.16, p = 0.03) were higher in the high-risk group than those in the low-risk group. Similarly, the mean heart rate during nighttime on holidays (mean_HR_holiday_night) was higher in the high-risk group than that in the low-risk group (t = −1.99, p = 0.049). However, the second heart rate recovery (HR_ second _recovery) was significantly greater in the low-risk group than that in the high-risk group (t = 2.83, p = 0.048).

In terms of steps data from wearable devices, the relative amplitude on weekdays (RA_week) was found to be higher in the high-risk group than that in the low-risk group (t = −2.13, p = 0.03). Among the sleep parameters, the REM sleep duration during holidays (REM holidays) was significantly shorter in the high-risk group than that in the low-risk group (t = 2.0, p = 0.048).

Self-reported clinical scales from the questionnaires showed significant differences between the SOS-Q and ISI. The high-risk group had higher ISI scores than the low-risk group (ISI_score, t = −4.01, p < 0.01), indicating greater insomnia severity. However, the high-risk group had significantly lower SOS-Q scores than the low-risk group (t = 2.58, p = 0.011). No significant differences were observed for the other analyzed variables (see Supplementary Information for detailed baseline characteristics).

Performance and feature importance of OSA prediction model

Table 2 presents the predictive performance of the four tree-based machine learning models (XGB, LGBM, RF, and XGBRF) across three distinct input datasets: (1) all data, (2) least data, and (3) selected data.

Table 2.

Performance comparison of prediction models across different datasets.

Dataset	Models	Accuracy	Precision	Recall	F1	AUC	PR-AUC
All data	XGB	0.82	0.90	0.72	0.80	0.94	0.94
	LGBM	0.80	0.86	0.72	0.78	0.90	0.91
	RF	0.80	0.86	0.72	0.78	0.91	0.90
	XGBRF	0.76	0.88	0.60	0.71	0.89	0.88
Least data	XGB	0.78	0.85	0.68	0.76	0.87	0.89
	LGBM	0.78	0.85	0.68	0.76	0.88	0.89
	RF	0.82	0.90	0.72	0.80	0.86	0.79
	XGBRF	0.78	0.82	0.72	0.77	0.87	0.88
Selected data	XGB	0.82	0.94	0.68	0.79	0.91	0.92
	LGBM	0.80	0.86	0.72	0.78	0.90	0.89
	RF	0.84	0.90	0.76	0.83	0.92	0.92
	XGBRF	0.72	0.87	0.52	0.65	0.87	0.88

All data included wearable devices, smartphone apps, and self-reported questionnaires.

The least amount of data is the minimum amount of data that can be traditionally verified.

Data were selected from all data using the feature-selection method.

XGB; eXtreme Gradient boosting, LGBM; Light gradient Boosting Machine, RF; Random Forest, XGBRF; eXtreme Gradient boosting Random Forest.

When utilizing all data (wearable devices, smartphone apps, and self-reported questionnaire data), XGBachieved the highest AUC of 0.94 and PR-AUC of 0.94, indicating strong discriminative power.It also demonstrated excellent performance across all metrics of other models. In contrast, LGBM showed the highest AUC of 0.88, along with a stable accuracy (0.78) and F1 score (0.76) in the Least Data configuration (BMI, Sex, and Age) Meanwhiile RF showed the best overall balance with an accuracy (0.82), F1 score (0.80), and lower PR-AUC (0.79). Since a low PR-AUC indicates weak confidence, LGBM can perform efficiently with small feautures data, while RF maintains decent performance but may see a drop in performance for ambiguous data. RF achieved the highest AUC of 0.92 for the Selected data (features chosen via Mutual Information), whereas XGB attained the highest precision (0.94) and lower AUC (0.91) than RF, indicating efficient performance with a reduced feature set. The Receiver Operating Characteristic (ROC) curves for all models across the three datasets are shown in Fig. 3A–C.

Fig. 3.

Comparison of OSA prediction model performance and feature importance across different datasets. (A–C) Receiver Operating Characteristic (ROC) curves of OSA prediction models using (A) all data, (B) least data, and (C) selected data. Model performance was evaluated using Extreme Gradient Boosting (XGBoost), a Light Gradient Boosting Machine (LightGBM), and Random Forest classifiers. The AUC values are indicated in the legends. (D–F) SHAP (SHapley Additive exPlanations) value plots indicating the relative importance of input features used by the Extreme Gradient Boosting model for (D) all data, (E) least data, and (F) selected data. The features were color-coded to distinguish between demographic and lifestyle variables and wearable-derived digital phenotypes.

A feature importance analysis conducted using SHAP provided insights into the key predictors influencing the decisions of the model (Fig. 3D–F). In the all dataset (Fig. 3D), BMI was the most influential feature, followed by the SOS-Q index, sex, and age. Notably, several wearable device features were ranked highly, including maximum exercise time during the daytime on weekdays (max_exercise_week_day), maximum exercise time during the nighttime on weekdays (max_exercise_week_night), resting heart rate (rest_HR), intradaily variability on holidays (IV_holiday).

In the selected data (Fig. 3F), BMI remained the highest contributor, with sex, age, and ISI scores playing significant roles. Notably, wearable device features such as maximum exercise time during the daytime on weekdays (max_exercise_week_day), maximum heartrate during nighttime on holidays (max_HR_holdiay_night), resting heartrate (rest_HR), second heart rate recovery (HR_second_recovery_phase), intradaily variability on weekdays (IV_week) and various heartrate or step data occupied a substantial portion of the top-ranked variables. This highlights the unique and significant contributions of digital phenotyping features to predictive models.

Calibration curves (Figure S1 in supplementary information) showed that the models generally exhibited high consistency between observed and predicted probabilities but tended to underestimate. That trend is also clearly evident in the confusion matrix (Figure S2–S4 in supplementary information).

The PR curve shows balanced performance when using the all dataset or a selected data, but performance is somewhat weaker when using the least dataset. This indicates that while patients can be assessed to some extent based on external factors (appearance etc.) alone, the digital phenotype can provide additional information for identifying individuals at risk for OSA.

Decision curve analysis (DCA) to evaluate the clinical utility of OSA prediction models (Supplementary Figure S6 in supplementary information) showed that most models provided consistently higher net benefit compared to the baseline strategy of “treat all” or “none”, suggesting potential clinical utility.

Discussion

Principal findings

The results of this study demonstrate the potential utility of integrated digital phenotyping derived from consumer-grade wearable devices, smartphone applications, and clinical scales for scalable OSA risk screening. Rather than replacing PSG-based diagnosis, this approach may help identify individuals at elevated risk who warrant further formal sleep evaluation, thereby addressing an important gap in access to early assessment. The results can be summarized across three domains: statistical comparisons of baseline characteristics, machine-learning model performance, and interpretability of key features.

In the analysis of baseline characteristics, individuals in the high-risk OSA group exhibited several clinically relevant differences compared with those in the low-risk group. For instance, consistent with the established literature, the high-risk group showed significantly higher BMI and older age, both well-known risk factors for OSA, compared to the low-risk group³⁶ (Supplementary Table S3 in Supplementary Information).

Wearable-derived heart rate metrics revealed that the high-risk group had a higher mesor (adjusted mean heart rate) and an elevated mean heart rate during the daytime and nighttime on weekdays and holidays compared to the low-risk group. This is consistent with the understanding that OSA is associated with increased sympathetic nervous system activity and cardiovascular burden. The finding that the second heart rate recovery was significantly smaller in the high-risk group was particularly noteworthy. Heart rate recovery after physical activity is a crucial indicator of autonomic nervous system function and primarily reflects parasympathetic reactivation and sympathetic withdrawal. Therefore, attenuated heart rate recovery in patients with OSA suggests impaired autonomic control characterized by sympathetic overactivity and/or reduced vagal tone, which is a common physiological consequence of intermittent hypoxia and sleep fragmentation in OSA³⁷. Our findings underscore that heart rate recovery, acting as a digital biomarker derived from consumer wearables, effectively reflects the physiological stress imposed by OSA³⁸.

Furthermore, the high-risk group showed reduced REM sleep duration on holidays, and lower daily step counts and distances traveled compared to the low-risk group. Reduced REM sleep in OSA is a clinically recognized phenomenon that is often associated with more severe apneic events during this sleep stage and is linked to metabolic and neurocognitive dysfunction^39,40. This reduction in REM sleep may be attributed to OSA; however, further research is required to confirm this finding. Additionally, the lower daytime physical activity levels observed in the high-risk group may reflect increased daytime somnolence and fatigue, which are common symptoms of OSA that can lead to reduced engagement in exercise and daily activities. This is consistent with previous studies which have found that patients with OSA tend to be less physically active than healthy individuals, likely due to the excessive daytime fatigue associated with untreated OSA⁴¹. This sedentary pattern could contribute to a vicious cycle: fatigue leads to inactivity, which in turn can promote weight gain and higher BMI, further exacerbating the severity of OSA. Our data, showing higher BMI and lower activity in the high-risk group, are consistent with this feedback loop. Breaking this cycle through interventions, such as exercise programs, may be beneficial, as regular physical activity has been shown to improve OSA outcomes and daytime energy levels^42,43.

A notable finding was the significantly lower SOS-Q score in the high-risk OSA group than that in the low-risk group. The SOS-Q measures smartphone dependency, with higher scores indicating greater dependency. Thus, our results indicate that the high-risk OSA group had significantly lower smartphone dependency, which is contradictory to some general observations linking increased screen time to poor sleep quality^44,45. This discrepancy may be due several plausible interpretations. First, individuals with significant sleep disturbances, such as those with OSA, may experience profound daytime fatigue and reduced energy, leading to less engagement in stimulating activities, including extensive smartphone use. Second, some individuals who recognize their sleep problems might actively reduce smartphone use as a conscious effort to improve sleep hygiene, which could reflect lower SOS-Q scores. Consequently, this finding warrants further investigation in larger and diverse cohorts to elucidate the complex interplay between sleep disorders, daytime functioning, and digital device usage patterns. Additionally, this finding indicates that not all factors associated with poor sleep in the general population (such as high screen time contributing to insomnia) will directly apply in a clinical OSA population; sometimes, severe OSA might overshadow or alter those typical behavioral associations. Therefore, future studies could benefit from longitudinal monitoring to determine whether SOS-Q scores change after OSA treatment (e.g., evaluating if device use increases once patients are less fatigued or if they maintain reduced screen habits).

In the machine learning analysis, the predictive performance across different data configurations underscored the additive value and efficiency of digital phenotyping. While models trained solely on conventional clinical variables (BMI, Sex, Age – “Least data”) achieved a respectable AUC of up to 0.88, incorporating a broader range of features from wearable devices, smartphone apps, and self-reported questionnaires (“All data”) significantly improved performance, with XGB reaching an AUC of 0.94. This finding demonstrates that digital phenotyping provides substantial complementary insights beyond traditional risk factors. Notably, models trained using a selected feature set (via Mutual Information) achieved the highest overall performance, with RF attaining an AUC of 0.92 and XGB with an AUC of 0.91. This highlights the power of efficient feature selection in reducing computational complexity and enhancing predictive accuracy, a finding consistent with other multimodal approaches in biomedical informatics⁴⁶. This superior performance, accompanied by a reduced yet highly informative feature set, reinforces the feasibility of developing lean yet highly effective screening tools.

Comparison to prior work

Model interpretability through SHAP analysis further elucidated the contributions of individual features. As expected, traditional risk factors, such as BMI, sex, age, and ISI score, were consistently among the most influential predictors across all data configurations. In addition, several novel wearable-derived features, including heart rate recovery, resting heart rate, maximum exercise time, and various circadian rhythm parameters (e.g., rest_HR and IV_holiday), emerged as significant contributing factors. The prominence of these features, which are not captured in conventional diagnostic workflows, emphasizes the unique insights provided by continuous real-world digital monitoring. While some of these findings are consistent with those of previous machine learning studies on OSA risk classification⁴⁷, the contribution of physical activity recovery dynamics and fine-grained circadian behavior, in particular, represents a relatively underexplored area in OSA prediction, making this study methodologically distinctive and clinically innovative.

Strengths and limitations

Consequently, these findings have significant clinical implications. Our results demonstrated that a meaningful prediction of OSA risk can be achieved using only readily available consumer-grade wearable devices and smartphone data, even in the absence of a gold-standard PSG or wearable oxygen saturation (SpO2) sensor. Thus, the ability to leverage minimal commonly available sensor values for screening OSA risk offers a highly scalable and low-burden solution for early identification. This is crucial for guiding individuals towards formal diagnostic evaluation (e.g., PSG or home sleep apnea testing) and facilitating timely intervention, especially in primary care settings or large-scale public health initiatives where PSG accessibility is limited, ultimately reducing the burden of undiagnosed OSA and its associated comorbidities.

Despite these promising results, this study has some limitations that must be acknowledged.

First, the use of the Berlin Questionnaire rather than PSG-confirmed OSA outcomes introduces the possibility of misclassification bias. Although PSG-based labeling would be preferable for diagnostic validation, the present study was designed to evaluate a scalable screening framework rather than a definitive diagnostic modell⁴⁸. In this context, the Berlin Questionnaire served as a pragmatic and clinically recognized screening anchor for OSA risk. Unlike one-time questionnaire administration, multimodal digital phenotyping may enable repeated and low-friction risk flagging in daily life. Nevertheless, our findings should not be interpreted as a replacement for PSG-based diagnosis, and future studies using PSG-confirmed OSA outcomes in larger and more diverse cohorts are needed to establish clinical validity and generalizability.

Second, missing data from wearable devices and self-reported measures were relatively common. Although advanced multimodal imputation strategies were applied to balance data retention with biological plausibility, these procedures may attenuate the true individual variability or introduce residual distortion. Third, internal validation was conducted on a relatively small subset (n = 50), which limits generalizability. Furthermore, the study population was recruited from a specific region in South Korea; therefore, cultural or lifestyle factors may have influenced the digital phenotyping patterns, necessitating validation in more diverse cohorts.

In conclusion, this study suggests that integrated digital phenotyping using consumer-grade wearable devices, smartphone-based measures, and clinical scales may support scalable pre-screening for OSA risk in real-world settings. The identified feature patterns, including heart rate dynamics and activity-related markers, highlight the potential value of everyday digital signals for early risk stratification. However, these findings should be interpreted in light of the questionnaire-based outcome definition and internal validation design, and further validation against PSG-confirmed OSA outcomes is warranted.

Conclusion

This study successfully demonstrated the feasibility of a scalable and low-burden approach for OSA risk screening. By integrating digital phenotyping data from widely accessible consumer-grade wearable devices, smartphone applications, and self-reported clinical scales, our machine learning models achieved high predictive performance (AUC of up to 0.94) without the need for gold-standard PSG.

Our findings underscore the significance of novel digital biomarkers such as heart rate recovery and circadian rhythm indicators, along with traditional risk factors such as BMI and ISI scores, in contributing to robust OSA risk prediction. These insights, derived from continuous real-world data, offer a unique perspective beyond that of conventional clinical assessments.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

OSA

Obstructive sleep apnea

PSG

Polysomnography

SpO2

Peripheral oxygen saturation

PPG

Photoplethysmography

BMI

Body mass index

REM

Rapid eye movement

CR

Circadian rhythm

L5

The least active 5-hour period

M10

The most active 10-h period

RA

Relative amplitude

IS

Intradaily stability

IV

Intradaily variability

HR

Heart rate

IRLS

International restless legs syndrome study group rating scale

ISI

Insomnia severity index

AUDIT-K

Korean version alcohol use disorder identification test

SOS-Q

Smartphone overuse screening questionnaire

PHQ

Patient Health Questionnaire

GAD

Generalized anxiety disorder

DBAS

Dysfunctional beliefs and attitudes about sleep

K-BRIAN

Korean version biological rhythms interview of assessment in neuropsychiatry

STAI-S

State-trait anxiety inventory

MESOR

Midline estimating statistic of rhythm

SMOTE

Synthetic minority oversampling technique

Author contributions

T.C. (tcheong@korea.ac.kr) and C.H.C. (david0203@korea.ac.kr) are the corresponding authors responsible for the data and materials, manuscript submission, peer review, publication process, authorship details, and ethics committee approval. H.K. performed data preprocessing and machine learning analysis, and authored the Methodology, Results, and Conclusion sections of the original manuscript. S.K. interpreted the results and authored the Introduction and Discussion sections of the original manuscript. H. K. and S. K. contributed equally to this study. J.W.Y., B.N.K., S.P.P., and H.J.L. were responsible for the data collection and approval of the medical research procedures. C.H.C. managed and supervised the project and provided guidance for interpreting the research findings. T.C. conceptualized the methodology for this problem. C.H.C. and T.C. reviewed and edited the manuscript.

Funding

This work was supported by the National Research Foundation (NRF) of Korea grants funded by the Ministry of Science and Information and Communications Technology (MSIT); Government of Korea (NRF-2021R1A5A8032895 and NRF-2022M3C1B6080866); ); and a grant from the Information and Communications Promotion Fund through the National IT Industry Promotion Agency (NIPA; H0601-24-1017), funded by the Ministry of Science and Information and Communications Technology (MSIT), Republic of Korea.

Data availability

The data supporting the findings of this study are available from the corresponding authors upon reasonable request. The data is not publicly available because of privacy and ethical restrictions. The analysis code is publicly available at: [https://github.com/subway75/Prediction-OSA-risk-group]

Declarations

Competing interests

The authors declare no competing interests.

Human ethics and consent to participate

All study procedures were reviewed and approved by the Institutional Review Board (IRB) of Korea University Anam Hospital (IRB No. 2022AN0587). Written informed consent was obtained from all the participants.

Consent for publication

Not applicable.