Aim: This work proposes an initial framework to detect non-apnea arousal based on statistical machine learning algorithms. Method: The proposed method is composed of feature extraction and classification steps. In the first step, features are extracted both in time and frequency domains for each frame obtained from oxygen saturation and respiratory airflow signals. Time domain features are obtained based on the evaluation of each frame’s mean, logarithm of energy and Shannon entropy statistical parameters. Frequency domain features, are constructed based on cepstrum coefficients of each frame. In the classification step, the data is classified into three classes such as non-apnea arousal, apnea arousal and non-arousal. For that purpose, kNN algorithm is utilized to obtain the preliminary multiple classification results since it has a low computational complexity compared to other classifiers.
Result: In initial experimental results, we have used 80% of the labeled data for training and the remaining 20% for the validation. We have obtained accuracy rate of 53% for three classes. Conclusion: We have obtained moderate initial results for classification task, and we are investigating new features and training strategies to increase the classification performance. For that, other available bio-signals of training dataset will be included in the feature extraction step to better model the non-apnea arousals. In addition, other classification approaches will be evaluated for alleviating the problem of class imbalance since only a couple percent of the training dataset frames involves the target arousal.