# Clinical validation of a contactless respiration rate monitor

### Study design

This paper covers studies validating RR with a contactless sleep monitor (CSM) against the thoracic effort belt (TREB) of the polysomnography (PSG) setup, in two different clinics (Klinik Lengg, Switzerland and Ruhrlandklinik, Germany).

In a monocentric, prospective, open-label non-randomized validation study at Klinik Lengg, the CSM was validated against PSG over 20 min on resting participants and during sleep (Study 1).

Additionally, we analyzed a dataset obtained as part of a larger study at Ruhrlandklinik, where the participants underwent a single overnight recording in the sleep laboratory with a PSG setup and the Sleepiz device (Study 2).

### Participants

In Study 1, patients suspected to suffer from sleep-related disorders or chronic cardiac, respiratory, or neuromuscular disorders were recruited from the neurorehabilitation ward of Klinik Lengg. The subjects included in Study 2 were patients of the Ruhrlandklinik, with a suspected or diagnosed sleep-related breathing disorder.

In both studies, the patients were ≥ 18 years old, and could consent to electrophysiological routine assessment (PSG) in writing. Exclusion criteria were previous enrolment into the current study, being a study investigator or one of their relatives, employees, and other dependent persons, having an implanted electrical device, pregnancy and inability to follow the study procedures, e.g. due to language problems or psychological disorders.

### Sleepiz One+ description

The Sleepiz One+ is a contactless respiration monitor whose operating principle is continuous-wave Doppler radar. The device is placed on the bedside table or a stand beside the user’s bed and points at the user’s thorax from a 40–50 cm distance. From this position, it transmits a continuous electromagnetic signal at a fixed frequency of 24 GHz (output power < 18 dBm) which is reflected at the user’s thorax, while being minimally influenced by clothing or blankets. With a beam aperture of 80° in horizontal- and 34° in vertical plane, the Sleepiz One+’s field of view from the specified position covers the user’s whole thorax-abdomen region. The reflected signal is received and pre-processed using a quadrature receiver29, resulting in two output channels BI(t) and BQ(t) which contain information on the movement in the whole thorax-abdomen region as distance changes of the user relative to the measurement device24,30 (Fig. 1).

The Sleepiz One+ data is transmitted to a cloud server, where it is analyzed using proprietary software which identifies and extracts rhythmic thorax and abdomen movements related to breathing motions to provide RR estimates, whose validation is the topic of this work.

The Sleepiz One+ is able to operate with two people in the same bed, but records only the respiration rate of the person closest to the device. Furthermore, multiple devices can operate in the same room, as long as their field of views do not intersect. Within this study, all participants were recorded with no other people in the bed, and with one device per room.

### Experimental setup

#### Study 1

Each participant was monitored for one night simultaneously with the Sleepiz One+ and a full PSG setup (SOMNOtouch RESP, Somnomedics) including electrocardiogram, electroencephalogram, airflow (nasal cannula), respiratory effort (thorax and abdomen belts) and pulse oximetry. The Sleepiz One+ device was placed beside the bed at a distance of 50 cm pointing at the participant’s thorax (Fig. 2). The person included in the image has provided informed consent to publish the image. All enrolled participants underwent the same procedure.

#### Study 2

Each participant underwent an overnight recording in the sleep laboratory with a full PSG setup (Nox A1, Nox Medical), including a TREB), which was the channel used for the performance assessment. The Sleepiz One+ was placed beside the bed at a distance of 50 cm pointing at the participant’s thorax.

### Statistical analysis

#### Data preprocessing

Data from each device, namely the TREBs of the PSG systems and the Sleepiz device, were exported and converted to a Python dictionary containing its raw data, measurement start time, and sampling rate. Next, devices were time-synchronized with each other. Finally, quality assessment was performed. The signals were visually assessed to identify artifacts, i.e. regions that are not interpretable, due to the patient moving or the sensors being positioned suboptimally, leading to a signal-to-noise ratio insufficient for analysis. These were removed from further analysis. The criteria used to exclude participants for analysis are specified below.

#### Study 1

For each patient, the 20-min analysis window to be included for instantaneous RR performance assessment was selected by identifying a window of 20 min of the recording, where the person was calmly lying in bed, within the first 3 h after the recording setup had been completed. The window was selected based on the following three conditions: (1) the reference device had started recording, (2) the Sleepiz device had started recording, and (3) less than 10 min of total duration where either the patient was moving, as detected by the Sleepiz device, or there were artifacts in RR estimates from the TREB.

The purpose of these conditions was to ensure the correctness of the reference instantaneous RRs and that a sufficient portion of the 20 min could be used for analysis (i.e., it was movement, breath irregularities and artifact-free), also considering that these conditions were most likely to be met in the first 3 h of sleep.

If no analysis window fulfilling all three conditions was identified in the first three hours of the recording, as determined by the Sleepiz device start time, the patient was excluded from the analysis. No sleep stage information was used for the selection of the analysis windows.

Participants were excluded from the RR statistics analysis if either the TREB or the Sleepiz data presented artifacts for 60% or more of the time-points between analysis start and end times. The start time would be the time of device which started to record later of the two and end time would the time of the device ended earlier of the two.

#### Study 2

Exclusion criteria were the same as for Study 1 for RR statistics analysis.

#### Computation of instantaneous respiratory rate

We define instantaneous RR as the number of breaths in a time window of 60 s (epoch), where consecutive time windows have an overlap of 55 s, i.e., they present a 5 s offset.

Instantaneous RR was computed for the data from the TREB through the following procedure:

• The signals were smoothed with a 7th order Butterworth low-pass filter with a cut-off frequency of 0.6 Hz. The filter cut-off was chosen based on the assumption of a maximum RR of 35Brpm.

• The inhalation times were identified by performing a quantile normalization and finding maxima in the resulting signal fulfilling properties related to the peak prominence and minimum distance between peaks. The identified inhalation times were visually reviewed to ensure their correctness.

• The instantaneous RR was derived by taking the inverse of the median time between inhalations.

The computed instantaneous RR from TREB inhalation times was not post-processed through any other means. Exhale times were not used for the RR computation due to challenges in precisely identifying the time of complete exhale; the exhalation curve is frequently flatter than the inhalation curve, and contains multiple peaks (Fig. 3, top row).

RR was computed for the Sleepiz One+ data following a similar procedure. The patient's chest movement was first computed from the raw data from the Sleepiz One+, consisting of channels BI(t) and BQ(t), using an arctangent demodulation algorithm. The chest movement was then filtered and normalized, and the inhalation and exhalation times were identified from the resulting signal. This information was used to compute a probability density over RRs in the range of 5–35 Brpm. In other words, for each RR within the range of 5 to 35 Brpm, the probability of the patient having that particular RR, given the observed inhalation and exhalation times, was computed. The probability density was used to obtain a RR per epoch, as well as to estimate the prediction confidence, i.e., the confidence of the algorithm in the RR prediction made being correct. If the confidence in the prediction is below a threshold, no RR is provided. It is worth noting that times where the patient is tossing and turning are automatically detected by the software and removed from the analysis, as these can lead to errors in the computation of RR.

#### Computation of respiratory rate statistics

RR statistics consisted of the minimum, average, and maximum RR over the recording. To remove the influence of outliers on these metrics, we computed them using the data percentiles, i.e. through the 10%, 50%, and 90% percentiles of instantaneous RR.

#### Performance evaluation metrics

Instantaneous RR performance was assessed through the mean absolute error (MAE), and accuracy: the fraction of time-points where the estimates provided by the reference device and the Sleepiz device were within 3 Brpm.

The agreement between reference and Sleepiz device for RR statistics was evaluated by the absolute error (AE).

#### Sample size

The minimum number of participants required for statistically significant outcomes of instantaneous RR accuracy was calculated using Eq. (1):

$${n}_{min}=\frac{\left(Z\alpha \right)2 \cdot {S}^{2}}{d2}$$

(1)

where Za determines the confidence level of the estimate (significance level), S is the sample standard deviation, and d is the minimal effect of interest. The significance level was set to 95%, corresponding to a value of Za to 1.64. The sample standard deviation was set to 5.2%, based on results from a previous pilot study. Finally, the minimal effect of interest was set to 2%. From these, we obtain a sample size of 19 participants. We included more than said number of participants in each center.

### Ethical approval

The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. Both included studies were conducted in accordance with the Declaration of Helsinki, the guidelines of Good Clinical Practice (GCP) issued by ICH, as well as the European Regulation on medical devices 2017/745, ISO 14155 and ISO 14971, and the Swiss Law and Swiss regulatory authority’s requirements (Study 1) or the German Law and German regulatory authority’s requirements (Study 2), respectively. Both studies were approved by the local ethics committees (Study 1: BASEC-Nr. 2020-00455, Cantonal Ethics Committee Zurich; Study 2: 19-8961-BO, Ethics committee of the medical faculty of the University of Duisburg-Essen) and written informed consent was taken from all individual participants.