Prediction of obstructive sleep apnea using Fast Fourier Transform of overnight breath recordings

Abstract

The objective of this study is to address the problem of predicting the risk of obstructive sleep apnea (OSA) from overnight breath recordings collected by a subject using a smartphone or an iPhone. The dataset used in this study was collected at a health care facility and consists of breathing amplitudes of 42 subjects using the smart phone App ZeeAppnea. A total of four data mining multi-level classifiers are used on the Fast Fourier Transform (FFT) of each time series, and prediction accuracies are computed. The Random Forest (RF) and the Support Vector Machine (SVM) classifiers yielded the best results, with overall multi-level prediction accuracies of 93% and 90%, respectively; the overall multi-level prediction accuracy of manual interpretations of recordings was 55%. The binary overall accuracies for the severe OSA class were 98% (RF), 95% (SVM) and 69% (manual interpretations). Our results show that either RF or SVM can be used on the recordings obtained from ZeeAppnea instead of the time-consuming manual interpretation of charts of breathing amplitudes by medical personnel, as this would improve prediction accuracy and automate the process of this screening application.

Keywords

• Obstructive sleep apnea
• Machine learning
• Smartphone
• Application
• Screening