Detection of sleep K-complexes using BOHB-tuned XGBoost based on multi-domain feature extraction and KMeansSMOTE

Abstract The K-complex, a distinctive brain wave pattern observed during sleep, holds significant importance in sleep research. Traditional manual annotation of electroence-phalogram (EEG) signals is time-consuming,labor-intensive and prones to errors. Although some studies have focused on the automatic detection of K-complexes, they still struggle to effectively address the issues caused by data imbalance. Therefore, the study demonstrates a method for detecting sleep K-complexs based on multi-domain feature extraction, KMeansSMOTE technology, and the combination of BOHB-Tuned XGBoost is proposed. There are four steps below. Firstly, the EEG data is preprocessed and 22 features are extracted. Then, the KMeansSMOTE oversampling technique is applied to generate samples for the minority class in the imbalanced data, aiming to achieve a balance between positive and negative samples. Next, XGBoost is employed for ranking feature importance and it is integrated with correlation analysis for feature selection. Finally, the BOHB-Tuned XGBoost model is utilized for K-complex detection using the selected features. Experimental results show that the model has an average accuracy of 96.60%, precision of 97.67%, sensitivity of 89.05%, specificity of 99.1%, and an F1 score of 92.99%. The method proposed in this study not only addresses the problem of data imbalance effectively but also demonstrates a more stable, high classification performance. This can assist sleep experts in analyzing EEG signals.