Results
The database search yielded 1414 articles, of which 1058 were left after removal of duplicates. For an overview of the selection process, see the PRISMA figure in Figure 1 . After assessment of relevance and eligibility, 14 studies were included. After checking the references and citations, three additional studies were included (E-text 3).
The methods researched in the included studies were bed-based methods, ultra-wideband (UWB) radar, Doppler radar, video, infrared (IR) cameras, garment-embedded with sensors, and sound analysis. Some studies used a combination of sensors. The main study characteristics are summarized in Table 1 . The accuracies of the researched techniques are summarized in Table 2 and the advantages and limitations of the techniques are summarized in Table3 .
Bed-based methods
The techniques that monitor breathing using sensors embedded in a bed were based on pressure sensors, vibration sensors, load cells, microphones, electromagnetic sensors, piezoelectric sensors, and electromechanical film. Arimoto et al.13, embedded these sensors in a sheet-like device, the SD-101(Kenzmedico co. Ltd., Saitama, Japan). Norman et al.14­,­15 and Collaro et al.16 evaluated the Sonomat (Sonomat, Balmain, Australia), a thin mattress overlay with embedded sensors and microphones. and So et al.17 used the TaidoSensor (Sumitomo Riko Company Limited, Nagoya, Japan), a rubber sheet made of piezoelectric material. Lee et al.18 embedded load cells in a bed frame and used their sensors to obtain a ballistocardiogram (BCG), from which the cardiac and respiration cycle, and thereby the RR, could be obtained. So et al.17Norman et al.14­,­15 and Collaro et al.16 aimed to detect apneas/hypopneas, and So et al.17 aimed to continuously measure RR and were also able to detect characteristic breathing patterns such as deep breathing and apneas. Lee et al.18 and Arimoto et al.13 used analytical software for automatic respiration monitoring, although the latter performed a manual correction since software was not yet available for measurements in children. Norman et al.14, 15researched the validity of the Sonomat (Sonomat) in children, and Collaro et al.16 aimed to do so for children with Down Syndrome. The Sonomat (Sonomat) contains four vibration sensors and two room sound microphones. Both Norman et al.14 and Collaro et al.16 used the combination of signals to differentiate between obstructive, central, and mixed apneas and compared these results with PSG.
UWB Radar
Kim et al.19, de Goederen et al.5, Huang et al.20, and Ziganshin et al.21 researched ultra-wideband (UWB) radar for respiration monitoring. Huang et al.20 and Ziganshin et al.21 aimed to detect apneas, while de Goederen et al.5 and Kim et al.19 used their system to measure RR. De Goederen et al.5 used the RR for sleep stage classification.
Doppler radar
Fox et al.22 used Doppler radar with analytical software for actimetry and compared their method with an actimetry watch.
Video
Al-Naji and Chahl23 used a video camera to monitor RR and used video magnification technique to do so.
IR camera
Both Al-Naji et al.24 and Bani Amer et al.25 used an IR camera in their research. Al-Naji et al.24 used the Kinect v2 (Microsoft, Redmond, WA) sensor, which has three optical sensors: an RGB camera, IR sensor, and IR projector, that provide an RGB image, IR image and depth image. These signals were used to measure RR. Simulated apneas were detected as well. Bani Amer et al.25 aimed to detect central, obstructive and mixed apneas. Both authors used software for signal analysis.
Garments
Two studies embedded sensors inside a garment; Gramse et al.26 in the MamaGoose (MMG) pajamas (Verhaert Design and Development, Kruibeke, Belgium), Ranta et al.27 in the NAPping PAnts (NAPPA) diaper cover (BABA Center, Helsinki, Finland). Both used an algorithm for RR monitoring, and Gramse et al.26 detected apneas based on visual examination of the abdominal and thoracic respiration signals. No reference standard was used for respiration signals.
Sound analysis
Emoto et al.28 used an artificial neural network to determine snoring/breathing episodes (SBEs) based on microphone recordings. Norman et al.15 compared the signals from the Sonomat (Sonomat) microphones with the mat as a whole. Breathing sounds and pattern prior to body movement were used to differentiate between spontaneous and respiration-induced body movement.
Quality assessment
Figure 2 shows the risk of bias and applicability concerns about each domain for the individual studies. In the figure, the colors green, yellow and red correspond to low, unclear and high risk of bias, respectively.
Eight studies included less than ten study participants. This, and the lack of information about the enrollment of study participants or exclusion criteria, increased the risk of bias in the patient selection.
Only Arimoto et al.13, Norman et al.14 and Collaro et al.16 compared their techniques with PSG and Ranta et al.27 compared their garment with capnography. These authors all described the process of patient sampling and of index and reference test interpretation.
Both Fox et al.22 and So et al.17included children and adults in their study population. In both of these studies, children <12 years old were the minority of the study population, which raises applicability concerns. These concerns also apply to the study by Kim et al.19 who evaluated their UWB radar device at the NICU. Although the neonates were clinically stable and full-term, this also raises concerns regarding applicability to our review question.