A method for the automatic detection of arousals during sleep.

Abstract

A method for the automatic detection of arousals in digital polysomnographic recordings is described. The computer program analyzed two EEG and one EMG derivations marking variable length segments as arousals. The processing of EEG data started from the wavelet transform, which characterizes the signal in the time-frequency domain, and resulted in a set of indices used to discriminate possible arousal segments. Transient increases in muscle activity were also identified, while a multichannel and context sensitive analysis allowed arousal detection. Out of 11 overnight recordings, 3 were used as the training set and 8 as the program testing set. In the first stage of the study two EEG experts inspected the tracings independently to score arousals. They then reviewed all recordings and jointly examined each event for validation, both those scored by themselves and those scored by the computer. A reference set of definite arousals (1125 in the testing set) and a number of uncertain events (266) were thus obtained. The sensitivity of the automatic system (88.1%) was higher than that of the human experts (72.4 and 78.4%) while the selectivity was lower (74.5% for the automatic system, 83.0 and 82.0% for the experts). This suggested that automatic detection, followed by an expert's validation, may render the analysis of arousals more widely feasible as well as support the study of arousal features.