A Hybrid Feature Selection and Extraction Methods for Sleep Apnea Detection Using Bio-Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Sleep apnea (SA) increases risks for stroke and cardiovascular diseases.
• Polysomnography (PSG) is the standard for SA detection.
• Efficient and accurate SA detection methods are crucial.

Purpose of the Study:

• To perform feature selection from PSG signals for SA detection.
• To develop a Support Vector Machine (SVM) model for SA classification.
• To evaluate the performance of reduced feature sets in SA detection.

Main Methods:

• Utilized the Physionet Apnea Database for feature extraction (n=87) from ECG, SaO2, airflow, and respiratory signals.
• Employed rank-sum and ANOVA for feature significance analysis.
• Applied hill-climbing feature selection and k-fold cross-validation with SVM for classification.

Main Results:

• The top-five significant features achieved the best classification performance.
• SVM with a Linear kernel demonstrated high accuracy (AUC=95.23%, Sensitivity=94.29%, Specificity=96.17%).
• Feature subset selection reduced dimensionality and computational load.

Conclusions:

• Feature subsets derived from multiple bio-signals show strong potential for identifying SA patients.
• Optimized feature selection enhances the efficiency and accuracy of SA detection systems.
• This approach offers a promising method for clinical SA diagnosis.