Patients who visited the otorhinolaryngology clinic between January 2023 and June 2024 and complained of snoring or other symptoms of OSA were enrolled in our study. Healthy controls were recruited from our hospital.

All enrolled participants met the following inclusion criteria: (a) over 18 years old, (b) examined by an otolaryngologist, (c) underwent a CT scan, and (d) patients with OSA additionally underwent polysomnography.

The exclusion criteria were as follows: (a) incomplete clinical data, (b) severe diseases, (c) inadequate CT scan images, (d) craniofacial anomalies, (e) dental implants causing CT artifacts; and (f) an overnight sleep duration of less than 6 h.

OSA was defined according to the American and Chinese guidelines using the apnea-hypopnea index (AHI), with a threshold of >5 events/h (23, 24). A classification model was evaluated, trained and tested on the features extracted from patients with OSA and healthy controls. Additionally, clinical parameters included the lowest and average oxygen saturation (SaO₂, percentage), body mass index (BMI, kg/m²).

The participant enrollment flowchart is presented in Figure 1A.

All participants underwent an upper airway CT scan (128-row, Siemens, Germany) in the supine position. During scanning, the participants were instructed to keep their mouths closed, maintain occlusion of the maxillary and mandibular teeth, and breathe normally (eupnea), while refraining from speaking, swallowing, or moving their bodies. The scanning range was extended from the roof of the nasal cavity to the supraglottic region, using a tube voltage of 120 kV, a matrix size of 512 × 512, and a layer thickness of 0.625–1.5 mm.

CT images were exported in Digital Imaging and Communications in Medicine (DICOM) format at a resolution of 300 dpi. The upper airway CT images were independently segmented by two experienced radiologists who were blinded to the patients' diagnoses. The regions of interest (ROIs) included the upper airway (AW), surrounding soft tissues (ST), and entire airway with soft tissues (EN). An experienced radiologist manually delineated the ROIs on each slice using open-source software 3D Slicer version 5.6.1 (https://www.slicer.org/).

The dataset was randomly divided into training and test sets in a 7:3 ratio. Cases in the training set were used to train the predictive model. Cases in the test set were used to independently assess the performance of the model. All images were resampled to a voxel size of 1 × 1 × 1 mm to standardize voxel spacing. An intraclass correlation coefficient (ICC) greater than 0.9 was considered robust.

A total of 1,834 radiomic features from the ROI images were extracted using Pyradiomics' in-house feature analysis program (version 3.0.1). Radiomic features can be categorized into three groups: (i) geometry, (ii) intensity, and (iii) texture. Geometric features describe the three-dimensional shape characteristics of the ROIs. Intensity features describe the first-order statistical distribution of voxel intensities within regions of interest (ROIs). Texture features characterize the spatial patterns, including second- and higher-order distributions of voxel intensities. Texture features are extracted using various methods, including the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). To obtain high-throughput features, various methods were applied using filters (wavelet, Laplace-Gauss, square root, logarithm, exponential, gradient transform, and local binary mode transform). Z-score normalization was applied to address the issue of varying scales in the manual radiomic features.

To extract high-throughput features, our study applied several techniques: non-linear intensity transformations to image voxels (Square, SquareRoot, Logarithm, Gradient, LBP3D, and Exponential); Gaussian Laplace filtering with sigma values of 1, 2, and 3; and eight wavelet transform algorithms (LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH) to compute first-order statistical and texture features.

The schematic diagram of the radiomics analysis workflow is presented in Figure 1B.

A Mann–Whitney U test was employed to screen for radiomic features. Only the radiomic features with a p-value < 0.05 were retained. An intraclass correlation coefficient (ICC) greater than 0.9 was considered robust. To ensure high repeatability, Spearman's rank correlation coefficient was computed between features; if the correlation coefficient between any two features exceeded 0.9, only one feature was retained. The minimum redundancy maximum relevance (mRMR) method was employed to select radiomic features. The least absolute shrinkage and selection operator (LASSO) regression model was applied to the discovery dataset for model construction. Depending on the regularization parameter λ, LASSO shrinks all regression coefficients toward zero and sets the coefficients of many irrelevant features exactly to zero. To determine an optimal λ, 10-fold cross-validation with a minimum error criterion was employed, whereby the final value of λ corresponded to the minimum cross-validation error (Supplementary Figure 1). The retained features with non-zero coefficients were used for regression model fitting and subsequently combined into the final model. The Python scikit-learn package was employed for LASSO regression modeling.

The final results of feature extraction are presented in Figure 2.

To address class imbalance between OSA patients and healthy controls in the training set, the Synthetic Minority Over-sampling Technique (SMOTE) was applied within each training fold of the cross-validation to generate synthetic samples for the minority class by interpolating between nearest-neighbor instances, thereby reducing the bias toward the majority class and improving model learning. Importantly, SMOTE was performed only on the training data to avoid information leakage, while the corresponding validation fold was kept unchanged. The final retained radiomic features from the three ROIs were used to build machine learning models separately using ten algorithms: logistic regression (LR), Naïve Bayes (NB), support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), extra trees (ET), eXtreme Gradient Boost (XGBoost), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), and multi-layer perceptron (MLP). These models were developed using 5-fold cross-validation. The performance was assessed through receiver operating characteristic (ROC) curves, along with additional evaluation metrics, including area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).

Subsequently, the most appropriate machine learning models, which consistently demonstrated optimal performance across all three ROIs, were further analyzed. Finally, the discrimination among the three machine learning models was evaluated through ROC analysis with AUC values and DeLong's test. Decision curve analysis (DCA) and calibration curves were employed to evaluate the clinical utility of the candidate models and identify the optimal radiomic model for OSA evaluation.

This study was approved by the Ethics Committee of the Seventh Affiliated Hospital of Sun Yat-sen University (KY-2025-385-01).

The Python statsmodels (version 0.13.2) package was used to perform statistical analysis. A P-value less than 0.05 was considered statistically significant. Continuous variables with a normal distribution were expressed as the mean ± standard deviation (SD) and analyzed using an independent samples t-test; continuous variables with a non-normal distribution were expressed as medians with the interquartile range [M (p25, p75)] and analyzed using the independent samples Mann-Whitney U-test; and categorical variables were expressed as frequencies (percentages) and analyzed using the χ² test. The correlations between radiomic features and clinical parameters were analyzed using the Pearson correlation coefficient, with statistical significance set at P < 0.05.

After screening, a total of 98 participants were finally enrolled in this study between January 2023 and June 2024. They were then randomly allocated to training and test sets. Participants were grouped as follows: healthy controls (n = 19), OSA (n = 79, AHI > 5 events/h), and 87 (88.8%) were male. The mean age of the HCs was 41.11 ± 14.26 years, their median BMI was 23.31 (22.56, 24.00) kg/m². All OSA patients underwent PSG, their mean age was 38.41 ± 7.76, their mean BMI was 27.96 ± 3.47 kg/m², the median AHI was 42.80 (23.30, 56.80) events/h, the median average SaO₂ was 94.00 (91.00, 95.00) %, and the median lowest SaO₂ was 75 (67.00, 82.00) %. Compared to healthy controls, patients with OSA had significantly higher weight and BMI in both the training and test sets.

The demographic and clinical parameters of the participants are summarized in Table 1.

A total of 818 AW, 661 ST, and 1,479 EN features were extracted using the Mann–Whitney U test (P < 0.05). After ranking with the mRMR method, the top 40 significant radiomic features were selected. Using LASSO regression, the key features comprised eight AW features, two ST features, and four EN features. The features were subsequently input into ten different machine learning algorithms to build predictive models. ROC analysis was performed to compare the performance of the machine learning algorithms based on radiomic features for model establishment. The NB algorithms achieved the best AUCs in the test set; their performance was confirmed in all three radiomic models, with AUCs ranging from 0.812 to 0.854 (Figure 3).

The ROC analysis of the training sets is shown in Supplementary Figure 2. More details on ML algorithms are provided in Supplementary Tables 1–3.

To identify the most accurate radiomic signature for OSA, three region-specific models (EN, AW, and ST) were evaluated using a naïve Bayes (NB) classifier in the independent test cohort. In this set, the EN model achieved the highest discrimination, with an area under the AUC of 0.854 (95% CI, 0.674–1.000), outperforming the AW model (AUC 0.819; 95% CI, 0.579–1.000), and the ST model (AUC 0.812; 95% CI, 0.627–0.998) (Figure 4A). DeLong's test confirmed that the AUC of EN model was significantly greater than those of both the AW and ST models (Table 2).

Calibration curves for all three radiomic models closely approximated the 45° reference line, indicating excellent agreement between predicted probabilities and observed OSA status (Figure 4B). Decision curve analysis further demonstrated that each model provided a net benefit over the “treat-all” and “treat-none” strategies. Notably, the EN model delivered the greatest net benefit across a wide range of threshold probabilities (Figure 4C).

The corresponding AUCs, calibration curves, and DCA in the training set are presented in Supplementary Figure 3.

The final selected radiomic features and clinical parameters (BMI, AHI, and lowest SaO₂) were regarded as variables. Associations among all variables were analyzed using Pearson correlation coefficient (P < 0.05).

Some radiomic features exhibit moderate correlations with clinical parameters, suggesting that they may reflect certain pathological changes in OSA. In contrast, for features exhibiting no or low correlations with clinical parameters, they may capture unique information that is complementary to what is not captured by traditional clinical measures, or clinical and radiomic features might operate independently, capturing distinct dimensions of OSA.

The correlation analysis of the test and training sets in the radiomic models is shown in Figure 5 and Supplementary Figure 4.