Abstract
Hibernating bears show remarkable metabolic suppression. Their decline in core body temperature (T (b) ) is moderate (from 38°C to 30-35°C), but their metabolism declines as much as 75%. To understand the role of sleep in this hypometabolic state, we recorded biotelemetrically EEG, EOG and EMG data over 3500 days from 16 captive American black bears in and out of hibernation under semi-natural conditions. This data set is too large to score manually for Wake, REM- and NREM sleep, so we tested two machine learning classifiers: (1) Somnotate trained on multiple one-day recordings, and (2) Somnivore, trained on a small subset from each recording. As automated scoring methods have not been applied to hibernating species before, a major concern is the effect changing brain temperature has on the EEG and on the machine learning based detection. Therefore, we selected reference data using consensus by 3 manual sleep scorers from each of 6 bears, two one-day recordings at the highest and lowest body temperatures during hibernation when T (b) was oscillating in multiday cycles, and a non-hibernating one-day recording in summer. Somnotate results were excellent when trained separately for hibernating and non-hibernating data. Training Somnotate separately for high and low T (b) within hibernation did not improve results further. Sleep times in hibernation were about 2x that in summer for both automated scores and manual scores (p<0.0001). There were no significant differences in occupancy of vigilance states between automated and manual scores in hibernation (p>0.05), but a small overestimate of sleep time in summer (p<0.05). Both applications yielded F-measures against manual scores in the 0.90-0.98 range. Outliers in the 0.67-0.88 range were correlated between the two applications, indicating that specific files are more challenging to annotate. We conclude that both applications have accuracies approaching that of manual scorers when trained on high quality data.