A quantization of deep belief networks for long short-term memory in sleep stage detection

Abstract

The study proposes the use of Deep Belief Networks (DBN) as quantization for inputs from Long Short-Term Memory (LSTM), called qLSTM. The system was tested using the sleep apnea data set to evaluate the combination. The target in the dataset is the sleep stage classification. Specifically, the study evaluates several variations of length of sequence data in LSTM. The result obtained that the optimal condition reached when the length is 25 with F-measure of 75.22%. These results are also compared with non-sequence classifiers namely Naive Bayes, Bayesian Networks, Bagging, and Multilayer Perceptron. All of the four non-sequence classifiers used only achieve F-measure below 60% except Bagging with F-measure of 64.20%. The other hand, the study also compares qLSTM and DBN HMM but qLSTM has higher of precision and F-measure than DBN HMM. It can conclude that the combination of DBN and LSTM quantization has a higher performance than the non-sequence classifiers and DBN HMM used.