Predictive modeling of obstructive sleep apnea using pharyngeal MRI radiomics and clinical data.

Abstract

Study objectives: This study aims to assess the predictive performance of models combining pharyngeal MRI radiomics and clinical data for distinguishing severe and non-severe obstructive sleep apnea (OSA).

Methods: A total of 106 patients were included in the study, with 48 patients having an AHI < 30 events/h and 58 patients having an AHI ≥ 30 events/h. Radiomics features were extracted from MRI images. After applying Minimum Redundancy and Maximum Relevance and Lasso with Cross-Validation for dimensionality reduction, radiomics models were developed using logistic regression (LR), support vector machine (SVM), random forest (RF), and gradient boosting machine (GBM). Age and BMI were used as clinical features to construct a combined model with radiomics features. The performance of the models was evaluated using F1 scores and the area under the Receiver Operating Characteristic curve (AUC).

Results: A total of 129 radiomics features were extracted from MRI images. Following dimensionality reduction and feature selection, two radiomics features with significant predictive value were identified. The combined model, incorporating SVM (AUC=0.78, F1=0.75), RF (AUC=0.78, F1=0.74), GBM (AUC=0.79, F1=0.75), and LR (AUC=0.82, F1=0.80), demonstrated superior performance compared to models based solely on radiomics features. The radiomics-only models included SVM (AUC=0.76, F1=0.72), RF (AUC=0.73, F1=0.67), GBM (AUC=0.76, F1=0.73), and LR (AUC=0.78, F1=0.76). Among the combined models, LR achieved the highest predictive accuracy and classification performance.

Conclusions: The combined model, integrating radiomics features with clinical characteristics, demonstrates a superior ability to distinguish between severe and non-severe OSA. This approach offers a non-invasive and effective new perspective for clinical decision-making.