Objective: Even though studies using machine learning on sleep-wake states have been performed, studies in various conditions are still necessary. This study aimed to examine the performance of the prediction model of locomotor activities on sleep-wake states using machine learning algorithms.
Methods: The processed data using moving average of locomotor activities were used as predicting features. The sleep-wake states were used as true labels. The prediction models were established by machine learning classifiers such as support vector machine with radial basis function (SVM-RBF), linear discriminant analysis (LDA), naïve Bayes, and random forest (RF). The prediction model was evaluated by a six-fold cross validation.
Results: The SVM-RBF and RF showed acceptable performance within a window of moving average from 480 to 1,200 seconds. The highest accuracy (0.869) was shown by the RF at the interval of 480 seconds. Meanwhile, the highest area under the curve (0.939) was shown by LDA at the interval of 870 seconds.
Conclusion: This study suggested that the prediction model on sleep-wake state using machine learning could show an improvement of the model performance when using moving average with raw data. The prediction model using locomotor activity can be useful in research on sleep-wake state.
Keywords: Animal model; Locomotor activity; Machine learning; Moving average; Sleep.