Feature Selection Algorithm based on Random Forest applied to Sleep Apnea Detection.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Signal Processing

Background:

• Feature selection is crucial in machine learning, particularly for biomedical data analysis.
• Existing random forest (RF) feature importance methods can be improved for better model interpretability and reduced feature sets.
• Automatic detection of sleep apnea using physiological signals like oxygen saturation (SpO2) requires robust feature selection.

Purpose of the Study:

• To introduce a new feature selection method based on out-of-bag (OOB) Cohen kappa values in random forest classifiers.
• To enhance RF feature selection by incorporating Cohen kappa, addressing correlated features, and ensuring patient-independent validation.
• To apply and evaluate this method for the automatic detection of sleep apnea from SpO2 signals.

Main Methods:

• Developed a feature selection technique modifying RF predictor importance by using Cohen kappa values instead of classification error.
• Incorporated a factor to mitigate correlated features and adapted OOB sample selection for patient-independent validation.
• Applied the method to SpO2 signals for sleep apnea classification, selecting features from an initial set of 286.

Main Results:

• Identified an optimal feature set of 3 parameters for sleep apnea classification, a reduction from 6 features in previous studies.
• Achieved improved model interpretability due to the reduced feature set.
• Observed a slight decrease in overall performance but maintained clinical screening performance.

Conclusions:

• The proposed feature selection method offers significant improvements over standard RF feature selection techniques.
• This approach effectively reduces the number of features, enhancing classifier interpretability.
• The method demonstrates potential for reliable sleep apnea detection with fewer, more meaningful parameters.