An examination of inductive learning algorithms for the classification of sleep signals.

To determine their usefulness in sleep stage scoring, nine inductive learning algorithms have been tested against the sleep signals of 161 subjects representing more than 130,000 sleep events. The performance of each algorithm has been examined relative to the number of somnologist-supplied events, the simplicity of the induced rules, the percentage of all events correctly classified and the percentage of classified-events correctly classified. The last category is especially important in building reliable systems in a medical domain, where it is better for an event to be labeled "unknown" rather than incorrectly labeled. Algorithms showing the best overall performance are C4, MDL, and AIMS, generating the simplest rules, with a very high overall accuracy. PRG, a more conservative classifier, has a significantly higher accuracy on events that it is able to classify. COBWEB and the Nearest Neighbor method had marginally higher accuracy when the number of somnologist-supplied events is limited.