Prediction of Obstructive Sleep Apnea Using Hypothalamic Radiomics and Machine Learning

To explore the potential of hypothalamic radiomics derived from T1-weighted magnetic resonance imaging (MRI) as an exploratory biomarker for predicting obstructive sleep apnea (OSA).

This study included 251 participants, comprising 127 OSA patients and 124 healthy controls (HCs), from two medical centers. Hypothalamic subunits were automatically segmented by a published deep convolutional neural network on 3D T1-weighted MRI. Radiomics features were extracted using PyRadiomics, including shape, first-order, texture, and wavelet features. Feature selection was performed using the Mann-Whitney U test, Pearson correlation, and LASSO regression. Seven classifiers were trained with three input types: clinical-only, radiomics-only, and radiomics-clinical. Model performance was evaluated using AUC, accuracy, precision, F1-score, and specificity in both internal and independent external validation cohorts. SHapley Additive exPlanations (SHAP) analysis was used to identify key predictive features.

A total of 4255 radiomics features were extracted, with 52 retained after feature selection. The radiomics-clinical Gradient Boosting Machine (GBM) model achieved the best performance, with an AUC of 0.808 in the internal validation cohort and 0.777 in the external validation cohort. Body mass index (BMI) was the most influential predictor, followed by radiomics features, notably the wavelet-LHL_firstorder_InterquartileRange_posterior from the posterior hypothalamic subunit.

Hypothalamic radiomics combined with clinical features offers a promising exploratory approach for predicting OSA. These findings highlight the potential of radiomics in identifying hypothalamic changes associated with OSA.