Combining machine learning models for the automatic detection of EEG arousals

Electroencephalographic (EEG) arousals are related to sleep fragmentation and the consequent daytime sleepiness, and are usually detected by visual inspection of sleep polysomnographic (PSG) recordings. As this is a time-consuming task, automatic processes are required. A method using signal processing and machine learning models is presented. Using signal processing techniques, after a first step of signal conditioning, abrupt frequency changes in two EEG derivations and amplitude events in one submental electromyogram are identified. These events are grouped if they occur at the same time, using the epoch segmentation for that purpose. A set of features (that includes Hjorth's Parameters and the Sleep Stage), is extracted from each group and used as input for several machine learning models. With a first dataset of 20 PSG recordings, six models are configured and compared: Fisher's Linear Discriminant, Support Vector Machines, Artificial Neural Networks, Classification Trees, k-Nearest Neighbors, and Naive Bayes. The best models, in terms of the classification error and the capabilities to detect EEG arousals, were used to build two different combined approaches. The first approach follows the Shortliffe and Buchanan's certainty factors model and the second follows a linear combination. Conducting experiments on 26 PSG recordings, a sensitivity of 0.78 and a specificity of 0.89 with an error of 0.12 was achieved using the first approach, and a sensitivity of 0.81 and a specificity of 0.88 with an error of 0.13 was achieved using the second approach. Both approaches improved the performance over the individual models. These results were also compared to two well-known ensemble methods: Random Forest and k-Nearest Neighbor Ensemble. Again, the combined approaches showed the best performance.