A Novel Approach to Assess Sleep-Related Rhythmic Movement Disorder in Children Using Automatic 3D Analysis

Abstract

Background: Unlike other episodic sleep disorders in childhood, there are no agreed severity indices for rhythmic movement disorder. While movements can be characterized in detail by polysomnography, in our experience most children inhibit rhythmic movement during polysomnography. Actigraphy and home video allow assessment in the child's own environment, but both have limitations. Standard actigraphy analysis algorithms fail to differentiate rhythmic movements from other movements. Manual annotation of 2D video is time consuming. We aimed to develop a sensitive, reliable method to detect and quantify rhythmic movements using marker free and automatic 3D video analysis. Method: Patients with rhythmic movement disorder (n = 6, 4 male) between age 5 and 14 years (M: 9.0 years, SD: 4.2 years) spent three nights in the sleep laboratory as part of a feasibility study (https://ichgcp.net/clinical-trials-registry/NCT03528096). 2D and 3D video data recorded during the adaptation and baseline nights were analyzed. One ceiling-mounted camera captured 3D depth images, while another recorded 2D video. We developed algorithms to analyze the characteristics of rhythmic movements and built a classifier to distinguish between rhythmic and non-rhythmic movements based on 3D video data alone. Data from 3D automated analysis were compared to manual 2D video annotations to assess algorithm performance. Novel indices were developed, specifically the rhythmic movement index, frequency index, and duration index, to better characterize severity of rhythmic movement disorder in children. Result: Automatic 3D video analysis demonstrated high levels of agreement with the manual approach indicated by a Cohen's kappa >0.9 and F1-score >0.9. We also demonstrated how rhythmic movement assessment can be improved using newly introduced indices illustrated with plots for ease of visualization. Conclusion: 3D video technology is widely available and can be readily integrated into sleep laboratory settings. Our automatic 3D video analysis algorithm yields reliable quantitative information about rhythmic movements, reducing the burden of manual scoring. Furthermore, we propose novel rhythmic movement disorder severity indices that offer a means to standardize measurement of this disorder in both clinical and research practice. The significance of the results is limited due to the nature of a feasibility study and its small number of samples. A larger follow up study is needed to confirm presented results.

Keywords: 3D video; Jacinto capita nocturna; automated; contactless; diagnostic tool; rhythmic movement disorder; sleep.

Copyright © 2019 Gall, Kohn, Wiesmeyr, van Sluijs, Wilhelm, Rondei, Jäger, Achermann, Landolt, Jenni, Riener, Garn and Hill.

Figures

Figure 1

Sleep laboratory setup. The participant was placed in a bed with a 3D camera (Kinect v2) mounted on the ceiling. The camera was directed towards the participant with the field of view covering the bed area. The recorded video was transmitted to a fan-less mini-PC located in the same room. This PC further transmitted the data to a notebook located in the control room.

Figure 2

3D algorithm processing pipeline. The chosen example shows a child lying in a bed where the head is indicated in (A and B) by the white circle. (A) The pre-processed depth image stores the distance between camera and bed for each pixel in color coded millimeters. The example shows a person lying in bed covered with a blanket. A white circle highlights the participant’s head. (B) The motion map indicates movement. Here, a movement augmentation in the head region is present (yellow area). (C) The 3D signal (blue) shows a deflection for this time interval and a movement is annotated (red) where the 3D signal is augmented. Represented is a movement sequence where the first part reflects the participant changing from a lying position to a RM position on all fours followed by RMs. (D) The FFT spectrogram shows the frequencies over time. Relative strength of the magnitude is indicated by the colorbar. The RM part shows a narrower frequency band around 2Hz (right of green line). In contrast, the transitional movement shows a blurred spectrum (left of green line).

Figure 3

Minimum pause duration influence on evaluation measures. Indicated are the F1-score and Cohen’s kappa for different values of the minimum pause duration applied to 3D video and visual-manual method. Durations >10 s do not result in further improvements.

Figure 4

RM time of night distribution plot for all recordings. Each data point shows how many RM episodes occur over all recordings in time intervals of 30 min. RMs occur predominantly in the beginning of the night and the start of the second half of the night. Automatic 3D finds high agreement with data obtained by manual annotations.

Figure 5

Duration distribution plot for all recordings. Represented is the count of detected RMs with a duration between intervals of 1 min. Noticeable is the high count of durations between 0 and 1 min. The data point for automatic 3D in this first interval exceeds the plot window, but its value is indicated by the arrow.

Figure 6

RM duration distribution overnight. The plot shows the mean duration of RMs per hour of the recorded nights for the automatic 3D and manual 2D method. Error bars indicate the standard error of the mean. RM durations indicate high variety for each of the periods. Periods 1-2 and 2-3 include only a single RM each for the manual data, where the standard deviations equal zero.

Figure 7

Fragmentation plot of all records. Illustrates the distribution of time elapsed between onset of consecutive RMs presented with a convenience interval width of 10 min. 3D reports higher fragmentation than the manual approach as the majority of time periods are less than 10 min. This data point lies outside of the plot window, but its value is indicated by the arrow.

Figure 8

RM time of night distribution plots according to Table 3. The three examples show automatic 3D annotations (lower, red) and manual 2D annotations (upper, cyan) over time. RM annotations are indicated with a colored rectangle. Periods without annotations stay blank. Case A only shows few RMs in the beginning of the night, hardly recognizable on the full night scaled plot. Case B exhibits the most severe form of RMD, while case C exhibits mild symptoms compared to B.

Figure 9

Frequency distribution plot of an individual record (Case B). The mean frequency of an episode is plotted over the time in hours. For this case, frequency remains relatively stable over the night.