Contactless detection of periodic leg movements during sleep: A 3D video pilot study

Abstract In clinical practice, the quality of polysomnographic recordings in children and patients with neurodegenerative diseases may be affected by sensor displacement and diminished total sleep time due to stress during the recording. In the present study, we investigated if contactless three‐dimensional (3D) detection of periodic leg movements during sleep was comparable to polysomnography. We prospectively studied a sleep laboratory cohort from two Austrian sleep laboratories. Periodic leg movements during sleep were classified according to the standards of the World Association of Sleep Medicine and served as ground truth. Leg movements including respiratory‐related events (A1) and excluding respiratory‐related events (A2 and A3) were presented as A1, A2 and A3. Three‐dimensional movement analysis was carried out using an algorithm developed by the Austrian Institute of Technology. Fifty‐two patients (22 female, mean age 52.2 ± 15.1 years) were included. Periodic leg movement during sleep indexes were significantly higher with 3D detection compared to polysomnography (33.3 [8.1–97.2] vs. 30.7 [2.9–91.9]: +9.1%, p = .0055/27.8 [4.5–86.2] vs. 24.2 [0.00–88.7]: +8.2%, p = .0154/31.8 [8.1–89.5] vs. 29.6 [2.4–91.1]: +8.9%, p = .0129). Contactless automatic 3D analysis has the potential to detect restlessness mirrored by periodic leg movements during sleep reliably and may especially be suited for children and the elderly.

have been demonstrated to be associated with increased sympathetic activity mirrored by increased heart rate and blood pressure (Pennestri, 2013;Siddiqui, 2007).
In recent years several automatic PLMS-detection techniques have been tested and validated successfully; for example, based on software, integrated into the polysomnography (PSG) system (Stefani, 2017), stand-alone devices Ferri, Rundo, et al., 2016) or actigraphy (Gschliesser, 2009). In a previous study (Garn, 2016), we showed that automatic contactless 3D movement analysis (3D) detected all leg movements recorded during PSG that had been annotated by a human scorer. Most recently, Weinreich et al. reported a high accuracy of a non-contact device in the detection of sleep-disordered breathing and PLMS (Weinreich, 2018).
In this prospective pilot study, we collected polysomnographic recordings of patients with PLMS and investigated if automatic 3D analysis was able to reliably detect PLMS. Patients and their polysomnographic recordings were excluded from the final analysis if clinical data were incomplete or technical problems (artifacts or loss of sensor information during a significant portion of the night) rendered the recording not useable for data extraction.

| Patient recruitment
The study was approved by the ethics committees of the Medical University of Vienna (EK-No. 1091 and the state of Upper Austria ). Written informed consent was obtained from all patients participating in the study.

| Video PSG
All subjects underwent at least one night of 8-hr video PSG according to the American Academy of Sleep Medicine (AASM) standards (Berri, 2018). In the case of more than one PSG recording in a patient, the second PSG was used for this study, un- in Linz; High Speed Dome AU-G65 in Vienna). Leg movements were recorded using surface electrodes placed longitudinally and symmetrically around the middle of the tibialis anterior muscle, 2-3 cm apart. For scoring of EMG activity, bipolar surface EMG was recorded with the low pass filter at 100 Hz, the high pass filter at 10 Hz and a sampling rate of 500 Hz. Amplification was set at 10 μV per mm. Impedance of surface EMG electrodes had to be lower than 10 kΩ. Two experienced somnologists (SS and MB) selected the final dataset.

| 3D analysis
The 3D sensor was mounted on the ceiling above the bed at a distance of about 1.8 metres (Figure 1a). Three-dimensional recordings were time-synchronized with the PSG by feeding a time synchronization signal from the clock of the 3D system into a separate channel of the PSG headbox. We used the time-of-flight (TOF) sensor (Schwarte, 2001) Microsoft Kinect One (Payne, 2014) for motion detection. This sensor emits weak, amplitude-modulated incoherent near-infrared (IR) light at a wavelength of 860 nanometers. Its radiation intensity is far below current safety standards (BS EN 14255-1 2005) for optical radiation. The surface of the scene reflects the light and this light is measured by a matrix of detector diodes. The electronic circuits behind each pixel provide both grey-scale values and the time-offlight of the light, which is used to compute the distance between sensor and reflecting surface. From the data of 512 × 424 pixels, our software computes a grey-scale video and a 3D depth image for each frame at 30 frames per second. Thereby, changes in the scene can reliably be detected. Temporal changes of the distances reflect movements, which can be assigned to specific body parts. Figure 1b shows how events were presented and compared in EMG, IR-video and 3D.

| Computerized scoring algorithm for PLMS detection and analysis
Three-dimensional and polysomnographic data were processed by Austrian Institute of Technology-developed software written in Python 3.4 for detecting and localizing movements in 3D and comparison of detections in 3D and EMG. Leg movements detected by 3D were also visually inspected by co-author MG to exclude noiserelated artefacts. Well aware of the necessity to include wakefulness and rapid eye movement (REM) sleep according to World Association of Sleep Medicine (WASM) criteria Ferri, Rundo, et al., 2016), we only analysed artefact-free non-REM sleep for this pilot study for the pragmatic reason that PLMS occur more frequently during non-REM sleep.
"Leg movements" in the EMG signals of the PSG recordings, which were carried out according to AASM standards (Berri, 2018) (see 2.2), were visually verified by authors MB and SS and served as the ground truth for the automatic quantification of 3D leg movements. Candidate leg movements in the PSG (CLM_ PSG) were selected according to WASM criteria Ferri, Rundo, et al., 2016), with a duration of 0.5-10 s for the unilateral and 0.5-15 s for the bilateral case. Leg movements detected in 3D with duration of 0.5-15 s were defined as CLM_3D.

| Statistical analysis
Normal distribution of data was tested with the Shapiro-Wilk test with a significance level of alpha = 0.05. Accordingly, mean or median values were calculated depending on whether the samples were normally distributed (=mean) or not (=median). Significant differences between 3D and PSG of the same measure (CLM, PLMS, PLMSI and PLMS_AI) were tested using the Wilcoxon signed-rank test as individual distributions that were not normally distributed. Again, we used a significance level of alpha = 0.05.
Differences between 3D and PSG were calculated as follows: the numerical difference of candidate leg movements (delta CLM) was calculated by subtracting candidate leg movements detect by PSG (CLM_PSG) from candidate leg movements detect by 3D (CLM_3D).
The numerical difference of periodic leg movements (delta PLMS) was calculated by subtracting periodic leg movements detected by PSG (PLMS_PSG) from periodic leg movements detect by 3D (PLMS_3D).
The difference of the periodic leg movement index (delta PLMSI) was calculated by subtracting the PSG-derived periodic leg movement index (PLMSI_PSG) from the 3D-derived periodic leg movements index (PLMSI_3D).
The difference of the periodic leg movement arousal index (delta PLMS_AI) was calculated by subtracting the PSG-derived periodic leg movement arousal index (PLMS_AI_PSG) from the 3D-derived periodic leg movement arousal index (PLMS_AI_3D).
Correlation analyses used Spearman's rank-order correlations coefficient to test for non-correlation. The test provides a correlation coefficient (r) and a p-value. We tested the correlation of resulting delta metrics: the numerical difference of candidate leg movements (delta CLM), the numerical difference of periodic leg movements (delta PLMS), the difference of the periodic leg movement index (delta PLMSI), the difference of the periodic leg movement arousal index (delta PLMS_AI) and the following PSG parameters: sleep efficiency (SE), arousal index (AI), the minimum O 2 saturation (min-SaO 2 ), the mean O 2 saturation (meanSaO 2 ) and the apnea-hypopnea index (AHI).

Commonly used classification scores derived from PSG and 3D
annotations measured the performance of the 3D method. PSG annotations served as the ground truth values. True positive counts (TP) were defined as events where a 3D movement overlaps with a TA EMG activity in the PSG. However, if two separated 3D movements were detected during the TA EMG activity in the PSG, only the first 3D movement was scored as TP, whereas the second was scored as a false-positive count (FP). FP counts also included 3D movements not overlapping with TA EMG activity in the PSG. False-negative counts (FN) presented TA EMG activity in the PSG not overlapping with any movement detected by the 3D camera/algorithm.

| Patients
Complete 3D and video PSG recordings of 65 consecutive patients were collected for the study. After careful review of the PSG data, 13 (20%) patients were excluded from the study because they did not have a sufficient number of leg movements during the PSG to analyse (see  disorder (n = 1), periodic limb movement disorder (n = 1) and no sleep disorder according to ICSD-III (n = 8). Descriptive polysomnographic data of these 52 patients are given in Table 1.  In the whole sample, 91.0/90.5/90.7% of the CLM_PSG were also detected in 3D (true-positive rate, CLM_TPR) and only 9.0/9.5/9.4% of the CLM_PSG were missed by 3D (false-negative rate, CLM_FNR). Thus, the positive predictive value (CLM_PPV) of 3D was 68.7/68.4/67.8% for CLM_PSG.
Correlation analysis did not show any significant correlation between the difference of PLMS_AI_PSG and PLMS_AI_3D (delta PLMS_AI) and the PSG parameters mentioned in 2.4. Figure 2 shows the total number of CLM_PSG (blue columns) und CLM_3D (green columns) with respect to their IMI. Based on recent findings by Ferri et al., ( , (2017b, we divided the IMI into three categories (i.e., IMI <10 s, IMI 10-90 s and IMI >90 s). According to WASM criteria , 3D detected significantly more CLM (CLM_3D) with an IMI of <10 s than the PSG (

| D ISCUSS I ON
We found that automatic 3D video analysis yielded an even higher number of PLMS than PSG, especially in the short and medium IMI range. Intriguingly, and given the background of the ongoing discussion of the relationship between arousals and PLMS (Figorilli, 2017), 3D video analysis was able to detect more PLMS associated with arousals.
In clinical practice we accept PLMS associated with arousals to be clinically significant, especially if the patient complains of non-restorative sleep and other causes of sleep disruption have been ruled out. Nevertheless, the term "clinical relevance" must be interpreted with caution. Manconi et al. (2008) showed that in patients with RLS only PLMS with an IMI between 10 and 90 s respond to dopaminergic treatment. PLMS with a short IMI of <10 s might not reflect "true" PLMS in the strictest sense, but it has also been shown that cardiac activation may be even more pronounced when induced or associated short interval leg movements (Ferri et al., , 2017b. Recently, Hooper (2018) performed a systematic video analysis of PLMS in a clinical population and proposed that the magnitude of movements should be considered as an additional factor related to the clinical significance of PLMS. Due to the exploratory nature of our study, we did not perform a systematic analysis of the composition of 3D-detected PLMS, but the contactless approach enabled us to detect leg movements involving more proximal leg muscles than TA. Leg movements detected only in 3D but not in PSG were visually verified. Still, a certain percentage of leg movements could only be detected by the TA EMG. Applying machine learning algorithms would most likely improve the detection rate of these movements in 3D.
Although PSG undoubtedly remains the reference standard for sleep studies, our contactless 3D video analysis may be a valuable tool to study motor restlessness in special patient groups such as children and the elderly, who are both prone to anxiety and sensor displacement during a polysomnography in a restricted environment. Most recently, Del Rosso, (2018) proposed "restless sleep disorder" as a new diagnostic category, which presents with "restless sleep" or motor behaviours involving large muscle groups and consisting of frequent repositioning, moving bed sheets or even falling out of bed. Given the age range of our sample, we are not able to draw firm conclusions regarding the applicability of 3D detection of body and/or leg movements in children. Both sleep laboratories participating in the current study usually perform PSG in adults; thus we plan to cooperate with the respective paediatrics departments in future studies.
We studied a relatively small and heterogeneous sample in this pilot study. Compared to RLS cohorts (Hornyak, 2007) the PLMS arousal indexes of our patients were low, which most likely reflects the substantial proportion of sleep apnea patients in our sample, who have been reported to show lower PLMS arousal indexes (O'Brien, 2009).
A limitation of our pilot study is the fact that we only analysed non-REM sleep. This was for the pragmatic reason that PLMS occur more frequently during non-REM sleep and also to avoid any REMsleep-associated movements and artefacts. In a further study both non-REM and REM sleep have to be carefully analysed to comply with the full diagnostic PSG criteria of PLMS.
In conclusion, our pilot study showed that automatic 3D video analysis of PLMS works well in a sleep laboratory setting and may be a promising tool to screen patient groups such as children and the elderly for motor restlessness during sleep in a contactless fashion.

This work was supported by the Austrian Research Promotion
Agency (FFG), project ID 860159.