Characterizing rhythmic wheel-turning behavioral patterns in cockroaches Rhyparobia maderae using machine learning

Abstract

Organisms must adapt to environmental changes occurring across multiple time scales, with endogenous multiscale clocks coordinating physiology and behavior with recurring environmental rhythms, including the dominant 24-hour cycle and faster ultradian rhythms. The Madeira cockroach (Rhyparobia maderae) provides a suitable model for investigating such multiscale temporal organization. Here, locomotor activity was recorded in running-wheel experiments under constant darkness. While the endogenous circadian clock produces a clearly visible 24-hour rhythm, it remains unknown whether locomotor behavior also exhibits temporal patterns at additional time scales. These temporal patterns cannot be found by classical frequency analysis, as they are veiled by higher harmonics of the circadian rhythm which are in the same frequency range. Unsupervised machine learning methods such as K-Means clustering, self-organizing maps and Gaussian mixture are used in search for fast ultradian rhythms possibly linked to circadian cycles in locomotor activity. Prior to applying these methods, data metrics are defined which characterize bouts of activity (called activity impulses) compared to periods of reduced activity. A stochastic pattern was found in these activity metrics which characterizes the time distance between activity impulses. Across all approaches, a consistent ultradian rhythm of approximately one hour was identified in the timing of the activity maxima. This rhythm was mainly detected during the subjective night, suggesting circadian control, and appears to consist of two components with periods of approximately 40 minutes and 1.5 hours. The method proposed in this paper is applied to two cockroach groups with different levels of activity, and is generalizable to diverse datasets occurring in the form of a time series with a dominant rhythm.