INTRODUCTION
Obstructive sleep apnea (OSA) is a respiratory disorder characterized by recurrent episodes of partial or complete obstruction of the upper airway during sleep, leading to intermittent hypoxemia, sleep fragmentation, and sustained sympathetic activation. (1,2) Its prevalence ranges from 9% to 38% in the general adult population, (3) with a higher incidence in men and in overweight or obese individuals. (4)
In the context of hypertension (HTN), OSA has gained particular relevance as it is a recognized cause of secondary hypertension and is associated with poorer blood pressure control, increased arterial stiffness, cardiac remodeling, and a higher risk of cardiovascular events. (5,6) Previous studies have shown that up to 30–50% of patients with HTN may have OSA, and this proportion is even higher in the subgroup with resistant hypertension. (7,8)
Despite its high prevalence, OSA remains an underdiagnosed condition, especially in the outpatient setting, where polysomnography, considered the gold standard for its diagnosis,is not always feasible due to limitations in availability, cost, or accessibility. (9) This highlights the need for specific screening tools for this population, capable of optimizing the selection of candidates for confirmatory studies and rationalizing the use of diagnostic resources.
In addition to polysomnography, ambulatory respiratory polygraphy has established itself as a valid diagnostic alternative in patients with high clinical suspicion, particularly in the outpatient setting. However, its performance may be limited in patients with cardiovascular comorbidities, such as hypertension, in whom clinical complexity may require more comprehensive studies.
Although validated screening scales exist for the general population, such as STOP-BANG or NoSAS, (6,7) their performance may vary in hypertensive patients, where the incorporation of echocardiographic parameters, such as diastolic dysfunction or cardiac morphological characteristics, could provide additional predictive value.
On this basis, the objective of this study was to develop and internally validate a clinical predictive screening model for OSA in patients with hypertension, using routine clinical and echocardiographic variables to generate a tool applicable in outpatient practice and adapted to this highrisk population.
OBJECTIVE
To develop and validate a clinical predictive model based on routine clinical and echocardiographic variables to estimate the risk of OSA in patients with hypertension.
METHODS
Study design and population
A retrospective observational study was conducted in a cohort of patients diagnosed with hypertension (HTN) treated in the outpatient Cardiology Department of a private institution in the Autonomous City of Buenos Aires, Argentina from 2017 to 2022. Adult patients over 18 years of age with a diagnosis of HTN who underwent a complete clinical evaluation and echocardiographic study were consecutively included. In the context of routine clinical practice, a subgroup of patients was evaluated for obstructive sleep apnea (OSA) using polysomnography based on clinical criteria and availability. Patients with incomplete data on the variables of interest were excluded, as were those with inconclusive or technically inadequate sleep studies. Additionally, patients with severe chronic respiratory diseases, advanced heart failure, or conditions that could interfere with the correct interpretation of sleep studies were excluded.
Definitions
Hypertension was defined according to current guidelines (10,11) and based on ambulatory blood pressure monitoring (ABPM) values: nighttime blood pressure (resting period) ≥120/70 mmHg, daytime blood pressure (active period) ≥135/85 mmHg, and mean 24-h blood pressure ≥130/80 mmHg. A nondipper pattern was considered present when the nocturnal reduction in systolic blood pressure was <10% compared to daytime values. Ambulatory blood pressure was measured with a validated oscillometric device (MEDITECH® ABPM-04 model, Easy ABPM software version 1.1.2.3) (12). The cuff size was selected according to the patient’s brachial circumference and the cuff was placed on the nondominant arm. The active period was defined as the time between 8:00 a.m. and 10:00 p.m. in which measurements were taken every 20 minutes, and the passive period, as the time between 10:00 p.m. and 8:00 a.m. in which measurements were taken every 30 minutes. This information was supplemented with the patient’s personal diary to corroborate the periods of activity and rest.
OSA, previously known as obstructive sleep apneahypopnea syndrome (OSAHS), was defined according to the current guidelines of the American Academy of Sleep Medicine (AASM) at the time of patient enrollment. (13) OSA was defined as the presence of repeated episodes of complete or partial upper airway obstruction during sleep detected by polysomnography and quantified by the apneahypopnea index (AHI). In the analyzed cohort, the diagnosis of OSA was exclusively determined by polysomnography, considered the gold standard for diagnostic confirmation. OSA was diagnosed with an AHI ≥5 events/hour, and its severity was classified as mild (5-14), moderate (15-29), or severe (≥30). Hypopnea was defined as a ≥30% airflow reduction for at least 10 seconds, associated with ≥3% oxygen desaturation and/or an arousal defined as a transient microarousal detected by polysomnography, which briefly interrupts the continuity of sleep.
Left ventricular (LV) diastolic dysfunction was defined according to current guidelines, (14,15) by applying the algorithm of the American Society of Echocardiography (ASE) and the European Association of Cardiovascular Imaging (EACVI), which integrates echocardiographic parameters to establish diastolic dysfunction and estimate LV filling pressure. In patients with sinus rhythm, diastolic dysfunction was considered when at least two of the following criteria were met: mean E/e′ ratio >14, septal e′ <7 cm/s or lateral e′ <10 cm/s, left atrial volume index >34 ml/m², or tricuspid regurgitation velocity >2.8 m/s. Severity was categorized according to this algorithm as grade I (impaired relaxation with normal filling pressure), grade II (pseudonormal pattern with elevated filling pressure), or grade III (restrictive pattern with elevated filling pressure). For the predictive model, diastolic dysfunction was considered a dichotomous variable (presence/absence), regardless of the grade of severity, to facilitate its use as a rapid screening tool.
Statistical analysis
Cohort description
Continuous variables were expressed as mean and standard deviation or median and interquartile range, depending on their distribution, and categorical variables as percentages. Comparisons between groups were performed using Student’s ttest or the Mann-Whitney U test for continuous variables and the chisquare test for categorical variables. Quantitative variables were categorized according to clinically relevant cutoff points or explored using generalized additive models and visual analyses to determine potential clinically significant inflection points, which were then used for categorization. Body mass index (BMI) was calculated as weight in kilograms divided by height in square meters. Age was analyzed as both a continuous and a categorical variable (stratified into <45, 45–65, and >65 years).
Selection of OSA predictors
Clinical and echocardiographic variables selected a priori based on their pathophysiological relevance and prior evidence were considered. Clinical variables included demographic characteristics, traditional cardiovascular risk factors (such as hypertension, diabetes, dyslipidemia, and smoking), and anthropometric parameters. Echocardiographic variables included structural and functional measures of the left ventricle, including assessment of diastolic and systolic function. To identify predictors, two complementary selection methods were used: the LASSO (Least Absolute Shrinkage and Selection Operator) penalized regression and the stepwise selection guided by the Akaike Information Criterion (AIC), which prioritized a parsimonious and clinically interpretable approach. Each method was run within a bootstrap resampling scheme with replacement (1000 iterations, with the frequency of variable inclusion recorded in each one, with the selection proportion and its 95% binomial confidence interval). Since both methods were applied with the same number of iterations, they were assigned equal weighting (1/2 each) in the calculation of a combined stability score. Predefined cutoff points (≥75% selection in each method) were established to define consensus, and those variables that simultaneously exceeded these thresholds in both methods were considered robust. This combined approach was adopted to improve robustness in predictor selection and reduce the risk of overfitting.
Multivariate analysis
The variables with the highest consistency (high frequency and/or agreement across methods) and those considered clinically relevant were included in multivariate logistic regression models to estimate the risk of OSA. Given the low number of events, a parsimonious approach was prioritized to avoid overfitting. Likewise, the inclusion of highly correlated or redundant variables was avoided to reduce model instability and improve its generalizability. The number of events per variable remained close to the recommended limits; therefore, the results should be interpreted with caution regarding the model’s stability. The results were expressed as odds ratios (OR) with their 95% confidence intervals (95% CI). The model’s performance and discriminatory power were assessed with the area under the ROC curve (AUC), and standard diagnostic metrics such as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy were calculated over 1000 bootstrap iterations to estimate their stability.
Model calibration was verified using the Hosmer-Lemeshow test and calibration curves. All analyses were performed using RStudio software (v4.4.0), and a pvalue <0.05 was considered statistically significant.
Ethical considerations
The study was conducted in accordance with the principles of the Declaration of Helsinki. (16) Given the retrospective nature of the analysis and the use of anonymized data, individual informed consent was not required.
RESULTS
General description of the population
A total of 1795 patients diagnosed with hypertension and complete complementary studies were included, of whom 88 (4.9% ) had a confirmed diagnosis of OSA via polysomnography. The mean age of the cohort was 59.3 ± 14.2 years, with a predominance of females (53%). In the OSA group, a higher proportion of males was observed (63.6% vs. 46.2%; p=0.002) and the mean age did not show statistically significant differences (56.9 ± 11.9 vs. 59.4 ± 14.3 years; p=0.115). Regarding cardiovascular risk factors, in the OSA group, the percentage of patients with diabetes mellitus was similar (15.9% vs. 17.0%; p=0.902) but the body mass index was higher (32.9 ± 6.5 vs. 29.3 ± 5.5 kg/m²; p<0.001). Likewise, the prevalence of diastolic
dysfunction was significantly higher in the OSA group (51.1% vs. 34.7%; p=0.002). Regarding cardiovascular events during followup, no statistically significant differences were observed between the two groups; the remaining baseline characteristics are presented in Table 1.
Variable selection using machine learning algorithms
In the variable selection analyses, both methods, the stepwise and LASSO, demonstrated high stability in identifying predictors of OSA. The selection frequency was plotted graphically, showing that variables such as age, sex, BMI, diastolic dysfunction, and cardiovascular history remained among the most consistently selected. The comparative map between the two methods revealed a group of variables that simultaneously exceeded the consensus thresholds (≥75% selection in both stepwise and LASSO methods), thus reinforcing their robustness as predictors. In this graph, the shaded area represents strict consensus between methods, while the size of each point reflects the weighted score (a 50/50 average of the proportions in both algorithms), allowing us to appreciate not only the consensus between methods but also the degree of stability of each variable. In contrast, other variables appeared more sporadically or were methoddependent, suggesting lower stability. These findings are summarized in Figure 1, which shows a Selection Concordance Map (LASSO - stepwise). In the comparative analysis, a small subset of variables showed a high frequency of simultaneous selection in both methods (consensus zone), which are the most stable predictors. The remaining exhibited low frequencies or were dependent on a single approach, which reflects lower consistency. In line with these findings and to prioritize model parsimony, diabetes mellitus was not included in the final model. Although it showed an association in the bivariate analysis, this could be explained by its multicollinearity with BMI and other cardiometabolic factors. Additionally, the limited number of events in the cohort constrained the number of variables that could be included in the multivariate model; therefore, its exclusion should be interpreted as a methodological decision aimed at preserving the model’s stability rather than as a lack of clinical relevance.
Fig. 1
Stability map: LASSO vs. stepwise
Size = weighted score (50% LASSO / 50% STEP). Shaded region = consensus (75% / 75% thresholds).
AF: atrial fibrillation; BMI: body mass index; DBP: diastolic blood pressure; DLP; dyslipidemia; HDL: highdensity lipoprotein; LA: left atrium; MACE: major adverse cardiovascular events.
Multivariate analysis: predictors of OSA
The presence of diastolic dysfunction, gender, age between 45 and 65 years, and BMI were associated with the presence of OSA. Other variables, such as a history of diabetes, coronary artery disease, or biochemical parameters, did not show a statistically significant association (Table 1). In the multivariable analysis, the independent predictors of OSA were diastolic dysfunction (OR 2.39; 95% CI 1.54-3.74; p<0.001), male gender (OR 2.12; 95% CI 1.35-3.38; p=0.001), age between 45-65 years (OR 1.71; 95% CI 1.10-2.66; p=0.017), BMI (OR per 1 kg/m² 1.08; 95% CI 1.05-1.11; p<0.001).
Final prediction model
The final model showed an AUC of 0.70 (95% CI 0.59-0.81), 71% sensitivity, 55% specificity, 8% PPV and 98% NPV. Calibration—assessed using the Hosmer-Lemeshow test— showed no significant differences (p>0.05), which indicated a good fit. The diagnostic performance of the model is presented in Figure 2.
Fig. 2
Performance of the OSA predictive model
ROC curve obtained with 1,000 bootstrap iterations. Median AUC: 0.70 (95% CI: 0.59-0.81). On the right, model performance with 71% sensitivity, 55% specificity, 98% negative predictive value (NPV), 8% positive predictive value (PPV), and 56% overall accuracy.
OSA: obstructive sleep apnea
DISCUSSION
Our findings can be summarized in three main points. First, we observed a low prevalence of OSA in our cohort of hypertensive patients, which likely reflects a selection bias, given that polysomnography was not performed systematically. Second, we identified that routine clinical and echocardiographic variables, easily obtainable in daily practice—such as gender, middle age (45–65 years), BMI, and diastolic dysfunction—were independently associated with OSA. Finally, we developed and internally validated an exploratory clinical model for OSA. The model demonstrated moderate discriminatory power consistent with tools based on simple clinical variables and acceptable sensitivity along with high negative predictive value, which suggests potential utility as a complementary tool to guide indication for polysomnography in outpatient practice.
Regarding prevalence, our results contrast with those reported in the literature. Globally, it is estimated that more than 936 million adults have OSA at any degree, and approximately 425 million cases are moderatetosevere. (3) In hypertensive patients, the reported prevalence ranges from 30% to 50%, and in the subgroup with resistant hypertension it ranges approximately from 70% to 80%. (1,2,17,18) These differences reinforce the idea that in our cohort, the prevalence is likely underestimated, consistent with observations in other clinical settings where the indication for sleep studies depends on clinical suspicion. (19,20)
Regarding associated variables, left ventricular diastolic dysfunction was the most consistent echocardiographic finding in our model. The intermittent hypoxia and nocturnal pressure overload characteristic of OSA promote myocardial and atrial remodeling, which is associated with an increased risk of atrial fibrillation and disease progression (5,15,21,22). Likewise, male sex, middle age (45-65 years), and elevated BMI are widely validated predictors in screening scales such as STOP-Bang and NoSAS. (6,7)
Although tools such as STOP-Bang and NoSAS have demonstrated good performance in the general population, their applicability in hypertensive patients may be limited as they do not consider cardiac structural variables. In this regard, our model incorporates diastolic dysfunction as a marker of cardiovascular remodeling, which could add a potentially relevant pathophysiological dimension in this specific population. This approach could represent an advantage over tools based exclusively on clinical and anthropometric variables, particularly in populations with high cardiovascular risk, such as hypertensive patients. However, the absence of a direct comparison with these scores in our cohort constitutes a limitation; therefore, future studies should further evaluate the relative performance and incremental utility of this model compared to widely validated screening tools.
Table 1
Baseline characteristics
| Variable | Overall | No OSA | OSA | p |
|---|---|---|---|---|
| n | 1795 | 1707 | 88 | |
| Female, n (%) | 951 (53.0) | 919 (53.8) | 32 (36.4) | 0.002 |
| Age (years), mean (SD) | 59.25 (14.2) | 59.37 (14.3) | 56.86 (11.9) | 0.107 |
| BMI, kg/m² , mean (SD) | 29.50 (6.0) | 29.32 (5.9) | 32.90 (6.5) | <0.001 |
| Daytime SBP, mmHg, mean (SD) | 139.52 (15.6) | 139.36 (15.6) | 142.47 (16.1) | 0.069 |
| Nighttime SBP, mmHg, mean (SD) | 126.27 (18.2) | 126.05 (18.1) | 130.56 (20.6) | 0.024 |
| LA area, cm², mean (SD) | 24.06 (8.8) | 24.12 (8.9) | 22.83 (7.6) | 0.179 |
| IVS, mm, mean (SD) | 11.18 (1.7) | 11.17 (1.7) | 11.37 (1.6) | 0.278 |
| Total cholesterol, mg/dL, mean (SD) | 181.11 (41.0) | 181.10 (40.7) | 181.38 (47.5) | 0.951 |
| LDL, mg/dL, mean (SD) | 105.76 (36.1) | 105.86 (36.1) | 103.67 (37.7) | 0.579 |
| HDL, mg/dL, mean (SD) | 49.89 (16.8) | 50.14 (16.9) | 45.07 (13.3) | 0.006 |
| ClCr, mL/min, mean (SD) | 110.34 (49.9) | 109.72 (49.5) | 122.38 (55.9) | 0.020 |
| Diabetes mellitus, n (%) | 304 (16.9) | 290 (17.0) | 14 (15.9) | 0.906 |
| Dyslipidemia, n (%) | 559 (31.1) | 518 (30.3) | 41 (46.6) | 0.002 |
| Statins, n (%) | 520 (29.0) | 486 (28.5) | 34 (38.6) | 0.054 |
| Diastolic dysfunction, n (%) | 637 (35.5) | 592 (34.7) | 45 (51.1) | 0.002 |
| LVH, n (%) | 333 (18.5) | 312 (18.3) | 21 (23.9) | 0.240 |
| Treatment, n (%) | 1,401 (78.6) | 1,329 (78.4) | 72 (83.7) | 0.295 |
| Treatment with 3 or more drugs, n (%) | 252 (14.0) | 237 (13.9) | 15 (17.0) | 0.500 |
| Age 45-60 years, n (%) | 612 (34.1) | 570 (33.4) | 42 (47.7) | 0.008 |
| CrCl <50 ml/min, n (%) | 111 (6.2) | 109 (6.4) | 2 (2.3) | 0.182 |
| Albuminuria >20 mg/day, n (%) | 508 (28.3) | 486 (28.5) | 22 (25.0) | 0.560 |
| HTN stratum, n (%) | 0.397 | |||
| Controlled HTN | 653 (36.4) | 624 (36.6) | 29 (33.0) | |
| Isolated systolic HTN | 469 (26.1) | 450 (26.4) | 19 (21.6) | |
| Isolated diastolic HTN | 105 (5.8) | 100 (5.9) | 5 (5.7) | |
| Systolic and diastolic HTN | 568 (31.6) | 533 (31.2) | 35 (39.8) | |
| Followup, months, mean (SD) | 43.16 (24.1) | 42.98 (24.1) | 46.62 (24.8) | 0.167 |
| MACE, n (%) | 162 (9.0) | 158 (9.3) | 4 (4.5) | 0.189 |
| Death, n (%) | 53 (3.7) | 51 (3.8) | 2 (2.3) | 0.684 |
| Treatment, n (%) | 1,401 (78.6) | 1329 (78.4) | 72 (83.7) | 0.295 |
BMI, body mass index; CrCl, creatinine clearance; HDL, highdensity lipoproteins; HTN, hypertension; IVS, interventricular septum; LA, left atrium; LDL, lowdensity lipoproteins; LVH, left ventricular hypertrophy; MACE, major adverse cardiovascular events; OSA, obstructive sleep apnea; SBP, systolic blood pressure; SD, standard deviation.
Although diabetes showed an association in the bivariate analysis, it did not remain a predictor in the final model, which could be explained by its multicollinearity with other cardiometabolic variables, particularly BMI. Additionally, the limited number of events in the cohort constrained the number of variables that could be included in the multivariate model; therefore, its exclusion should be interpreted as a methodological decision aimed at preserving the stability of the model rather than as a lack of clinical relevance. (23,24)
From a methodological standpoint, the combined use of two variable selection strategies (LASSO and AIC stepwise) within an internal validation framework using bootstrapping made it possible to identify clinically relevant predictors with good stability in order to reduce the overfitting risk. (9,25) Unlike other scores developed in the general population, this model was derived from a cohort of hypertensive patients, a highrisk group with a high prevalence of OSA (1-3, 22,23,26), which could enhance its applicability in outpatient clinical practice. The clinical implications of detecting OSA in hypertensive patients are significant. Untreated OSA contributes to the progression of hypertension and is associated with an increased risk of cardiovascular events. (15,27) In this regard, the American Heart Association has emphasized the importance of considering screening strategies in highrisk subgroups, such as those with resistant hypertension. (24) Similarly, interventions such as continuous positive airway pressure (CPAP) have demonstrated a reduction in blood pressure, especially in patients with resistant hypertension or an altered nocturnal blood pressure profile. (23,28)
Finally, among the limitations of our study, its retrospective and singlecenter design should be noted, which could restrict the generalizability of the results. The low prevalence of OSA in the cohort, along with the small number of events, may have limited statistical power and affected the stability of the model, with a potential risk of overfitting in the selection of predictors. However, the implementation of variable selection techniques and internal validation using bootstrapping sought to mitigate this effect. Likewise, the low prevalence of OSA may have affected the diagnostic metrics, particularly the positive predictive value. The high negative predictive value observed should be interpreted in this context, given that prevalence directly impacts its magnitude. In this scenario, the model may prove more useful for ruling out the disease than for confirming it, although its performance should be evaluated in populations with higher prevalence.
Since the indication for polysomnography was not systematic but rather guided by clinical suspicion, the analyzed cohort represents a selected subgroup of patients, which introduces potential selection and verification bias. Consequently, the results should be interpreted in the context of a population with an increased pretest probability of OSA. Future studies should validate this model in larger multicenter cohorts, compare it with existing screening tools, and explore the incorporation of biomarkers and parameters derived from nocturnal blood pressure and left atrial strain, in order to optimize its performance. (14,15) Taken together, these findings provide preliminary evidence regarding the value of a clinical screening model in this population, which needs further confirmation in prospective studies and other cohorts.
CONCLUSION
The developed clinical model, based on simple variables that are easily obtained in outpatient practice, demonstrated moderate discriminatory power and high negative predictive value, which suggests its utility as a screening tool to rule out OSA in hypertensive patients. Its implementation could optimize screening, rationalize the indication for polysomnography, and promote early detection of a highly prevalent and frequently underdiagnosed condition. However, external validation in different settings and populations is required before widespread clinical adoption.