Application of various machine learning techniques to predict obstructive sleep apnea syndrome severity

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Sleep Medicine

Background:

• Obstructive sleep apnea syndrome (OSAS) incidence is rising globally.
• Polysomnography (PSG), the traditional diagnostic method, has limitations.
• There is a growing need for effective OSAS screening tools.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting OSAS severity.
• To explore both supervised and unsupervised learning techniques for OSAS classification.
• To identify potential AI-driven alternatives to conventional OSAS diagnosis.

Main Methods:

• Utilized a dataset of 4014 patients.
• Applied unsupervised learning (clustering: hierarchical agglomerative, K-means, bisecting K-means, Gaussian mixture model).
• Employed supervised learning (classification: XGBoost, LightGBM, CatBoost, Random Forest) with feature engineering.

Main Results:

• Achieved high classification accuracy for OSAS severity prediction.
• Accuracy rates were 88% for AHI ≥ 5, 88% for AHI ≥ 15, and 91% for AHI ≥ 30.
• Demonstrated the efficacy of gradient boost-based models in OSAS severity assessment.

Conclusions:

• Machine learning holds significant potential for predicting OSAS severity.
• AI-powered tools can complement or enhance traditional OSAS screening methods.
• Further research can leverage these findings for improved OSAS diagnosis and management.