Towards Continuous CameraBased Respiration Monitoring in Infants
Abstract
:1. Introduction
2. Materials and Methods
2.1. Materials
2.1.1. Experimental Setup
2.1.2. Dataset
2.1.3. Manual Annotation
2.2. Method
2.2.1. Preprocessing
2.2.2. Motion Detection
 Gross Motion Detector: let $\mathbf{X}\left(n{T}_{s}\right)$ be the frames in the jth window, with $n=0+(j1)/{T}_{s},\phantom{\rule{0.166667em}{0ex}}1+(j1)/{T}_{s},\phantom{\rule{0.166667em}{0ex}}\dots ,\phantom{\rule{0.166667em}{0ex}}N+(j1)/{T}_{s}$, and $N=72$ samples, corresponding to the samples in the jth window with a sampling period ${T}_{s}=0.111$ s. The gross motion detector was based on the absolute value of the Difference of Frames (DOFs) in the jth window. More formally:$$\mathbf{D}\left(u{T}_{s}\right)=\left\frac{\partial \mathbf{X}\left(n{T}_{s}\right)}{\partial n}\right,$$$$\mathbf{MP}\left(u{T}_{s}\right)=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\phantom{\rule{4.pt}{0ex}}\mathbf{D}\left(u{T}_{s}\right)>th{r}_{1}\hfill \\ 0\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$$$th{r}_{1}=\frac{Range\left(\mathbf{X}\right)}{{f}_{1}},$$$$s\left(u{T}_{s}\right)=\frac{{\sum}_{\tilde{m}=1}^{\tilde{M}}{\sum}_{\tilde{l}=1}^{\tilde{L}}m{p}_{\tilde{m},\tilde{l}}\left(u{T}_{s}\right)}{\tilde{M}\xb7\tilde{L}}.$$Here, $m{p}_{\tilde{m},\tilde{l}}\left(u{T}_{s}\right)$ is an element of $\mathbf{MP}\left(u{T}_{s}\right)$ at the position $\tilde{m}$ and $\tilde{l}$.
 Motion Classification: the ratio of moving pixels $s\left(u{T}_{s}\right)$ was used to perform the classification between usable and unusable segments for RR detection. In particular, we aim at detecting the unusable moments, i.e., the ones containing type 1 motion. The main assumption is that type 1 is part of a more complex kind of motion, typical of infants’ crying motion. Therefore, the simplest way to detect it is to assume that type 1 motion will result in more moving pixels compared to any of the usable segments.To perform a classification between the two, a second threshold $th{r}_{2}$ was introduced, which was applied to the ratio of moving pixels $s\left(u{T}_{s}\right)$. The final classification was, therefore, performed on a windowbased fashion, i.e., each window was classified as containing type 1 motion, corresponding to 1, or usable, corresponding to 0.Since we used three cameras in the thermal setup, we applied this algorithm three times. For the RGB dataset this was not necessary, as there was only a single camera used. In the visible case the classification will be:$$Motio{n}_{vis}\left(j\right)=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\phantom{\rule{4.pt}{0ex}}\exists \phantom{\rule{0.277778em}{0ex}}\phantom{\rule{0.277778em}{0ex}}u\phantom{\rule{0.166667em}{0ex}}:\phantom{\rule{0.166667em}{0ex}}s\left(u{T}_{s}\right)\ge th{r}_{2}\hfill \\ 0\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$For the thermal case instead:$$Motio{n}_{th}\left(j\right)=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\phantom{\rule{4.pt}{0ex}}\exists \phantom{\rule{0.277778em}{0ex}}\phantom{\rule{0.277778em}{0ex}}u\phantom{\rule{0.166667em}{0ex}}:\phantom{\rule{0.166667em}{0ex}}({s}_{1}\left(u{T}_{s}\right)\ge th{r}_{2}\phantom{\rule{4.pt}{0ex}}\mathrm{OR}\phantom{\rule{4.pt}{0ex}}\hfill \\ & \phantom{\rule{2.em}{0ex}}\phantom{\rule{2.em}{0ex}}{s}_{2}\left(u{T}_{s}\right)\ge th{r}_{2}\phantom{\rule{4.pt}{0ex}}\mathrm{OR}\phantom{\rule{4.pt}{0ex}}\hfill \\ & \phantom{\rule{2.em}{0ex}}\phantom{\rule{2.em}{0ex}}{s}_{3}\left(u{T}_{s}\right)\ge th{r}_{2})\hfill \\ 0\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$${s}_{1}\left(u{T}_{s}\right)$, ${s}_{2}\left(u{T}_{s}\right)$, and ${s}_{3}\left(u{T}_{s}\right)$ are the ratios of moving pixels obtained from the three thermal views.
 Ground Truth: The ground truth used to evaluate the performance of our motion detector was obtained based on the manual annotations presented in Section 2.1.3. In particular, the ground truth was built using the sliding window approach. Each window was classified as excluded, as type 1 motion, or as usable. The condition used was the presence of at least a frame in the window which results in being true for one of those categories. The excluded class had the priority, if this was true for at least a frame in the window, the entire window was classified as excluded. If the latter was false then type 1 motion was taken into consideration in the same manner, and lastly if the two above were both false we classified the window as usable.
 Parameters Optimization: the factor ${f}_{1}$, for the moving pixels detection, and the threshold $th{r}_{2}$, for the motion classification, were optimized. A leaveonesubjectout crossvalidation was used to optimize the two parameters. The approach was chosen considering that environment changes, e.g., environment temperature, blankets type, and position, can influence the parameters values and therefore, the betweenbaby variability is more important than the withinbaby variability. The set of parameters that resulted in the highest balanced accuracy for each fold was considered as a candidate set. The final chosen set was the most selected candidate set. This metric was preferred compared to the classic accuracy due to the imbalance in our two classes (usable was more frequent than type 1 motion). The optimization was performed on the training and testing set, presented in Table 1. This set includes 9 babies and therefore, 9 folds were performed in the crossvalidation. Two sets of parameters were empirically chosen for the training and correspond to ${f}_{1}=[4;\phantom{\rule{0.166667em}{0ex}}5;\phantom{\rule{0.166667em}{0ex}}6;\phantom{\rule{0.166667em}{0ex}}7;\phantom{\rule{0.166667em}{0ex}}8;\phantom{\rule{0.166667em}{0ex}}9;\phantom{\rule{0.166667em}{0ex}}10;\phantom{\rule{0.166667em}{0ex}}11;\phantom{\rule{0.166667em}{0ex}}12]$ and $th{r}_{2}=[0.004;\phantom{\rule{0.166667em}{0ex}}0.005;\phantom{\rule{0.166667em}{0ex}}0.006;\phantom{\rule{0.166667em}{0ex}}0.007;\phantom{\rule{0.166667em}{0ex}}0.008;\phantom{\rule{0.166667em}{0ex}}0.09;\phantom{\rule{0.166667em}{0ex}}0.010;\phantom{\rule{0.166667em}{0ex}}0.011;\phantom{\rule{0.166667em}{0ex}}0.012]$. The most chosen set, used in the next steps, was ${f}_{1}=8$ and $th{r}_{2}=0.005$, more information on the results can be found in Section 3.
2.2.3. Respiration Rate Estimation
 PseudoPeriodicity: this first feature is based on the assumption that respiration can be considered a periodic signal. This feature was not changed compared to [26]. A differential filter was used to attenuate lowfrequencies resulting in filtered time domain signals called ${x}_{m,l}^{\prime}\left(n{T}_{s}\right)$. The signals were zeropadded, reaching a length equal to ${N}_{z}=120\xb7N$, and multiplied for an Hanning window. Afterwards, a 1D Discrete Fourier Transform (DFT) was used to estimate the spectrum called ${y}_{m,l}^{\prime}\left({f}_{k}\right)$ with $k=0,\phantom{\rule{0.166667em}{0ex}}1,\phantom{\rule{0.166667em}{0ex}}\dots ,\phantom{\rule{0.166667em}{0ex}}\frac{{N}_{z}}{2}1$ and ${f}_{k}=\frac{k}{{N}_{z}\xb7{T}_{s}}\phantom{\rule{0.166667em}{0ex}}Hz$. This feature consists of the calculation of the height of the normalized spectrum’s peak. More formally:$${q}_{m,l}={\displaystyle \frac{{\displaystyle \underset{0\le {f}_{k}\le \frac{({N}_{z}/21)}{{N}_{z}\xb7{T}_{s}}}{max}}\underset{}{(\mid {y}_{m,l}^{\prime}\left({f}_{k}\right)\mid )}}{\sqrt{{\displaystyle \sum _{{f}_{k}=0}^{\frac{({N}_{z}/21)}{{N}_{z}\xb7{T}_{s}}}}{\mid {y}_{m,l}^{\prime}\left({f}_{k}\right)\mid}^{2}}}}.$$Each ${q}_{m,l}$ represent the height of the peak of the spectrum of the pixel in position $(m,l)$, ${q}_{m,l}$ are elements of the first feature $\mathbf{Q}$.This feature is sensitive to the presence of type 2 motion. Regions moving due to this type of motion can generate a big variation in the pixels’ values (depending on the contrast). This variation can, therefore, produce a strong DC component, which will result in a high ${q}_{m,l}$. The combination with the other features allows us to obtain motion robustness, Figure 3 presents an example during a type 2 motion and the pseudoperiodicity feature is visible in Figure 3b.
 Respiration Rate Clusters (RR Clusters): this feature is based on the observation that respiration pixels are not isolated but grouped in clusters. To automatically identify the pixels of interest more accurately, modifications were introduced to this feature to improve the robustness to the presence of NNS, typical when the infant has the soother, and to cope with the presence of the respiration’s first harmonic. The frequencies corresponding to the local maxima of the spectrum ${y}_{m,l}^{\prime}\left({f}_{k}\right)$ were found and the properties of the harmonic were checked:$${\mathbf{h}}_{m,l}=\underset{li{m}_{1}<{f}_{k}<li{m}_{2}}{arg\phantom{\rule{0.166667em}{0ex}}localmax}(\mid {y}_{m,l}^{\prime}\left({f}_{k}\right)\mid ),$$$$r{r}_{m,l}=\left\{\begin{array}{cc}{h}_{m,l}\left(1\right)\hfill & \mathrm{if}\phantom{\rule{0.277778em}{0ex}}\exists \phantom{\rule{0.277778em}{0ex}}\widehat{z}>1:\mid {h}_{m,l}\left(\widehat{z}\right)2\xb7{h}_{m,l}\left(1\right)\mid <\frac{1}{N\xb7{T}_{s}}\phantom{\rule{0.277778em}{0ex}}AND\hfill \\ & ({y}_{m,l}\left({h}_{m,l}\left(\widehat{z}\right)\right)<{y}_{m,l}\left({h}_{m,l}\left(1\right)\right)\phantom{\rule{0.277778em}{0ex}}AND\hfill \\ & {y}_{m,l}^{\prime}\left({h}_{m,l}\left(\widehat{z}\right)\right)\ge {y}_{m,l}^{\prime}\left({h}_{m,l}\left(1\right)\right))\hfill \\ \underset{{f}_{k}}{arg\phantom{\rule{0.166667em}{0ex}}max}\left(\mid {y}_{m,l}^{\prime}\left({f}_{k}\right)\mid \right)\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$We have, therefore, estimated the main frequency component for each pixel. To avoid erroneous RR estimation caused by higher frequencies components, e.g., caused by NNS, the $r{r}_{m,l}$ that were higher than $li{m}_{2}$ were put to zero. Therefore:$${\widehat{rr}}_{m,l}=\left\{\begin{array}{cc}r{r}_{m,l}\hfill & \mathrm{if}\phantom{\rule{0.277778em}{0ex}}r{r}_{m,l}<li{m}_{2}\hfill \\ 0\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$The ${\widehat{rr}}_{m,l}$ are elements of $\widehat{\mathbf{RR}}$, an example is shown in Figure 3f. The nonlinear filter introduced in [26] was applied:$${w}_{m,l}={\displaystyle \frac{1}{9}}\sum _{r=1}^{3}\sum _{o=1}^{3}\left({\displaystyle \frac{1}{exp({\kappa}_{1}\xb7\mid {\widehat{rr}}_{m,l}{\widehat{rr}}_{r,o}\mid /{\widehat{rr}}_{m,l})}}\right),$$It should be noted that the ${\widehat{rr}}_{m,l}$ on which we imposed the value 0 in Equation (10), will not result in a high ${w}_{m,l}$, even if there are clusters of zeros in $\widehat{\mathbf{RR}}$. This is due to the equation of the filter that with ${\widehat{rr}}_{m,l}=0$ will produce NaNs (Not a Number). The same will happen for regions with type 2 motion, where the main frequency component is the DC. This property allowed to avoid type 2 motion regions in the pixel selection phase achieving motion robustness, an example is visible in Figure 3e.
 Gradient: this last feature is based on the assumption that respiration motion can be only visualized at edges. This feature has been modified to make it independent of the setup used:$${g}_{m,l}=\left\{\begin{array}{cc}1\hfill & \mathrm{if}\phantom{\rule{4.pt}{0ex}}\sqrt{{\left({\displaystyle \frac{\partial {\overline{a}}_{m,l}}{\partial m}}\right)}^{2}+{\left({\displaystyle \frac{\partial {\overline{a}}_{m,l}}{\partial l}}\right)}^{2}}>\frac{Range\left(\mathbf{A}\right)}{{\kappa}_{2}},\hfill \\ 0\hfill & \mathrm{otherwise},\hfill \end{array}\right.$$
2.3. Evaluation Metrics
3. Results
4. Discussion
5. Conclusions
Author Contributions
Funding
Institutional Review Board Statement
Informed Consent Statement
Data Availability Statement
Acknowledgments
Conflicts of Interest
References
 Fairchild, K.; Mohr, M.; PagetBrown, A.; Tabacaru, C.; Lake, D.; Delos, J.; Moorman, J.R.; Kattwinkel, J. Clinical associations of immature breathing in preterm infants: Part 1—central apnea. Pediatr. Res. 2016, 80, 21–27. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Baharestani, M.M. An overview of neonatal and pediatric wound care knowledge and considerations. Ostomy/Wound Manag. 2007, 53, 34–55. [Google Scholar]
 Di Fiore, J.M. Neonatal cardiorespiratory monitoring techniques. Semin. Neonatol. 2004, 9, 195–203. [Google Scholar] [CrossRef]
 Alinovi, D.; Ferrari, G.; Pisani, F.; Raheli, R. Respiratory rate monitoring by video processing using local motion magnification. In Proceedings of the 2018 26th European Signal Processing Conference (EUSIPCO), Rome, Italy, 3–7 September 2018; pp. 1780–1784. [Google Scholar]
 Sun, Y.; Wang, W.; Long, X.; Meftah, M.; Tan, T.; Shan, C.; Aarts, R.M.; de With, P.H.N. Respiration monitoring for premature neonates in NICU. Appl. Sci. 2019, 9, 5246. [Google Scholar] [CrossRef] [Green Version]
 Jorge, J.; Villarroel, M.; Chaichulee, S.; Guazzi, A.; Davis, S.; Green, G.; McCormick, K.; Tarassenko, L. Noncontact monitoring of respiration in the neonatal intensive care unit. In Proceedings of the 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017), Washington, DC, USA, 30 May–3 June 2017; pp. 286–293. [Google Scholar]
 Huang, X.; Sun, L.; Tian, T.; Huang, Z.; Clancy, E. Realtime noncontact infant respiratory monitoring using UWB radar. In Proceedings of the 2015 IEEE 16th International Conference on Communication Technology (ICCT), Hangzhou, China, 18–20 October 2015; pp. 493–496. [Google Scholar]
 Kim, J.D.; Lee, W.H.; Lee, Y.; Lee, H.J.; Cha, T.; Kim, S.H.; Song, K.M.; Lim, Y.H.; Cho, S.H.; Cho, S.H.; et al. Noncontact respiration monitoring using impulse radio ultrawideband radar in neonates. R. Soc. Open Sci. 2019, 6, 190149. [Google Scholar] [CrossRef] [Green Version]
 Mercuri, M.; Lorato, I.R.; Liu, Y.H.; Wieringa, F.; Van Hoof, C.; Torfs, T. Vitalsign monitoring and spatial tracking of multiple people using a contactless radarbased sensor. Nat. Electron. 2019, 2, 252–262. [Google Scholar] [CrossRef]
 Joshi, R.; Bierling, B.; Feijs, L.; van Pul, C.; Andriessen, P. Monitoring the respiratory rate of preterm infants using an ultrathin film sensor embedded in the bedding: A comparative feasibility study. Physiol. Meas. 2019, 40, 045003. [Google Scholar] [CrossRef]
 Bu, N.; Ueno, N.; Fukuda, O. Monitoring of respiration and heartbeat during sleep using a flexible piezoelectric film sensor and empirical mode decomposition. In Proceedings of the 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Lyon, France, 22–26 August 2007; pp. 1362–1366. [Google Scholar]
 Bekele, A.; Nizami, S.; Dosso, Y.S.; Aubertin, C.; Greenwood, K.; Harrold, J.; Green, J.R. Realtime neonatal respiratory rate estimation using a pressuresensitive mat. In Proceedings of the 2018 IEEE International Symposium on Medical Measurements and Applications (MeMeA), Rome, Italy, 11–13 June 2018; pp. 1–5. [Google Scholar]
 Abbas, A.K.; Heimann, K.; Jergus, K.; Orlikowsky, T.; Leonhardt, S. Neonatal noncontact respiratory monitoring based on realtime infrared thermography. Biomed. Eng. Online 2011, 10, 93. [Google Scholar] [CrossRef] [Green Version]
 Pereira, C.B.; Yu, X.; Czaplik, M.; Rossaint, R.; Blazek, V.; Leonhardt, S. Remote monitoring of breathing dynamics using infrared thermography. Biomed. Opt. Express 2015, 6, 4378–4394. [Google Scholar] [CrossRef]
 Scebba, G.; Da Poian, G.; Karlen, W. Multispectral Video Fusion for Noncontact Monitoring of Respiratory Rate and Apnea. IEEE Trans. Biomed. Eng. 2021, 68, 350–359. [Google Scholar] [CrossRef]
 Eichenwald, E.C. Apnea of prematurity. Pediatrics 2016, 137, e20153757. [Google Scholar] [CrossRef] [Green Version]
 Shao, D.; Liu, C.; Tsow, F. Noncontact Physiological Measurement Using a Camera: A Technical Review and Future Directions. ACS Sens. 2021, 6, 321–334. [Google Scholar] [CrossRef] [PubMed]
 Massaroni, C.; Nicolò, A.; Sacchetti, M.; Schena, E. Contactless Methods For Measuring Respiratory Rate: A Review. IEEE Sens. J. 2020. to be published. [Google Scholar] [CrossRef]
 Mercuri, E.; Pera, M.C.; Brogna, C. Neonatal hypotonia and neuromuscular conditions. In Handbook of Clinical Neurology; Elsevier: Amsterdam, The Netherlands, 2019; Volume 162, pp. 435–448. [Google Scholar]
 Mizrahi, E.M.; Clancy, R.R. Neonatal seizures: Earlyonset seizure syndromes and their consequences for development. Ment. Retard. Dev. Disabil. Res. Rev. 2000, 6, 229–241. [Google Scholar] [CrossRef]
 Lim, K.; Jiang, H.; Marshall, A.P.; Salmon, B.; Gale, T.J.; Dargaville, P.A. Predicting apnoeic events in preterm infants. Front. Pediatr. 2020, 8, 570. [Google Scholar] [CrossRef]
 Joshi, R.; Kommers, D.; Oosterwijk, L.; Feijs, L.; Van Pul, C.; Andriessen, P. Predicting Neonatal Sepsis Using Features of Heart Rate Variability, Respiratory Characteristics, and ECGDerived Estimates of Infant Motion. IEEE J. Biomed. Health Inform. 2019, 24, 681–692. [Google Scholar] [CrossRef]
 Alinovi, D.; Ferrari, G.; Pisani, F.; Raheli, R. Respiratory rate monitoring by maximum likelihood video processing. In Proceedings of the 2016 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), Limassol, Cyprus, 12–14 December 2016; pp. 172–177. [Google Scholar]
 Janssen, R.; Wang, W.; Moço, A.; de Haan, G. Videobased respiration monitoring with automatic region of interest detection. Physiol. Meas. 2015, 37, 100–114. [Google Scholar] [CrossRef] [Green Version]
 Villarroel, M.; Chaichulee, S.; Jorge, J.; Davis, S.; Green, G.; Arteta, C.; Zisserman, A.; McCormick, K.; Watkinson, P.; Tarassenko, L. Noncontact physiological monitoring of preterm infants in the Neonatal Intensive Care Unit. NPJ Digit. Med. 2019, 2, 1–18. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Lorato, I.; Stuijk, S.; Meftah, M.; Kommers, D.; Andriessen, P.; van Pul, C.; de Haan, G. Multicamera infrared thermography for infant respiration monitoring. Biomed. Opt. Express 2020, 11, 4848–4861. [Google Scholar] [CrossRef]
 Sun, Y.; Kommers, D.; Wang, W.; Joshi, R.; Shan, C.; Tan, T.; Aarts, R.M.; van Pul, C.; Andriessen, P.; de With, P.H. Automatic and continuous discomfort detection for premature infants in a NICU using videobased motion analysis. In Proceedings of the 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Berlin, Germany, 23–27 July 2019; pp. 5995–5999. [Google Scholar]
 Hafström, M.; Lundquist, C.; Lindecrantz, K.; Larsson, K.; Kjellmer, I. Recording nonnutritive sucking in the neonate. Description of an automatized system for analysis. Acta Paediatr. 1997, 86, 82–90. [Google Scholar] [CrossRef] [PubMed]
 Pineda, R.; Dewey, K.; Jacobsen, A.; Smith, J. Nonnutritive sucking in the preterm infant. Am. J. Perinatol. 2019, 36, 268–276. [Google Scholar] [PubMed]
 Patel, M.; Mohr, M.; Lake, D.; Delos, J.; Moorman, J.R.; Sinkin, R.A.; Kattwinkel, J.; Fairchild, K. Clinical associations with immature breathing in preterm infants: Part 2—periodic breathing. Pediatr. Res. 2016, 80, 28–34. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Mohr, M.A.; Fairchild, K.D.; Patel, M.; Sinkin, R.A.; Clark, M.T.; Moorman, J.R.; Lake, D.E.; Kattwinkel, J.; Delos, J.B. Quantification of periodic breathing in premature infants. Physiol. Meas. 2015, 36, 1415–1427. [Google Scholar] [CrossRef] [PubMed] [Green Version]
 Lorato, I.; Stuijk, S.; Meftah, M.; Verkruijsse, W.; de Haan, G. CameraBased OnLine Short Cessation of Breathing Detection. In Proceedings of the 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), Seoul, Korea, 27–28 October 2019; pp. 1656–1663. [Google Scholar]
 Lee, H.; Rusin, C.G.; Lake, D.E.; Clark, M.T.; Guin, L.; Smoot, T.J.; PagetBrown, A.O.; Vergales, B.D.; Kattwinkel, J.; Moorman, J.R.; et al. A new algorithm for detecting central apnea in neonates. Physiol. Meas. 2011, 33, 1–17. [Google Scholar] [CrossRef] [PubMed]
Infant  Video Type  Gestational Age (weeks + days)  Postnatal Age (days)  Sleeping Position  Duration (hours)  Set 

1  Thermal  26w 4d  59  Supine  2.98  T&T 
2  Thermal  38w 5d  3  Supine  2.74  T&T 
3  Thermal  34w 1d  16  Supine  2.93  T&T 
4  Thermal  26w 3d  59  Prone  3.16  T&T 
5  Thermal  39w  2  Lateral  3.05  T&T 
6  Thermal  40w 1d  6  Supine  2.95  T&T 
7  Thermal  40w 2d  1  Lateral  0.92  T&T 
8  RGB  36w  47  Supine  0.30  T&T 
9  RGB  30w  34  Supine and Lateral  0.57  T&T 
10  Thermal  26w 4d  77  Supine  2.94  V 
11  Thermal  26w 4d  77  Supine  2.97  V 
12  Thermal  33w 4d  5  Supine  2.97  V 
13  Thermal  34w 2d  9  Supine  2.87  V 
14  Thermal  32w 2d  11  Supine  2.96  V 
15  Thermal  35w 1d  8  Supine  2.94  V 
16  Thermal  38w 1d  2  Supine  3.00  V 
17  Thermal  27w 5d  16  Supine  2.96  V 
Annotation Labels  Subcategories and Details  

Included  (i) Infant activity 

(ii) NNS    
Excluded  (iii) Interventions  includes both parents and caregivers interventions 
(iv) Other 

Accuracy  Balanced Accuracy  Sensitivity  Specificity  

Training and testing set  88.22%  84.94%  80.30%  89.58% 
Validation Set  82.52%  77.89%  66.85%  88.93% 
Previous Version of Method [26]  Current Version of the Method  

Usable  NNS Only  Usable  NNS Only  
MAE (BPM)  4.54 ± 1.82  9.39 ± 3.68  3.55 ± 1.63  7.11 ± 4.15 
PT  68.59% ± 13.29%  4.59% ± 6.93%  68.59% ± 13.29%  4.59% ± 6.93% 
Infant  Usable Excluding NNS  Type 2 motion Only  Still Only  

MAE  RMSE  PR  PT  MAE  PT  MAE  PT  
Training and testing  1  1.86  3.34  83.61%  70.38%  1.57  27.92%  1.51  34.61% 
2  2.87  3.97  73.71%  40.60%  2.56  20.90%  2.64  13.02%  
3  6.30  8.09  39.44%  67.83%  6.32  39.23%  6.28  24.38%  
4  4.43  6.21  60.16%  72.75%  4.99  44.09%  2.49  20.39%  
5  5.04  7.61  56.44%  40.22%  4.84  29.24%  2.24  5.35%  
6  2.97  4.73  71.34%  66.74%  3.70  29.96%  1.94  31.69%  
7  2.80  4.15  72.08%  46.16%  2.57  30.28%  0.70  4.61%  
8  1.89  3.40  88.63%  89.71%  1.76  11.47%  1.91  77.84%  
9  1.62  2.70  85.55%  81.60%  2.88  24.16%  1.08  56.76%  
Average  3.31  4.91  70.11%  64.00%  3.47  28.58%  2.31  29.85%  
± sd  ± 1.61  ± 1.94  ± 15.84%  ± 17.82%  ± 1.62  ± 9.56%  ± 1.62  ± 24.22%  
Validation  10  4.46  6.62  61.41%  63.62%  5.52  34.40%  2.44  22.78% 
11  3.79  5.54  64.96%  55.55%  4.01  34.62%  2.27  12.29%  
12  6.23  7.98  38.98%  68.20%  5.98  33.70%  6.60  23.35%  
13  6.29  8.51  44.00%  69.53%  6.30  51.04%  3.59  6.13%  
14  6.89  9.56  47.37%  73.38%  7.35  44.73%  4.58  18.00%  
15  4.75  6.65  54.11%  78.86%  4.83  42.08%  4.39  26.81%  
16  4.09  5.73  60.97%  76.84%  4.39  28.92%  3.21  30.73%  
17  6.40  8.78  47.79%  71.22%  7.64  40.14%  3.15  19.60%  
Average  5.36  7.42  52.45%  69.65 %  5.75  38.71%  3.78  19.96%  
± sd  ± 1.21  ± 1.49  ± 9.35%  ± 7.47%  ± 1.32  ± 7.14%  ± 1.40  ± 7.90% 
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. 
© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).
Share and Cite
Lorato, I.; Stuijk, S.; Meftah, M.; Kommers, D.; Andriessen, P.; van Pul, C.; de Haan, G. Towards Continuous CameraBased Respiration Monitoring in Infants. Sensors 2021, 21, 2268. https://doi.org/10.3390/s21072268
Lorato I, Stuijk S, Meftah M, Kommers D, Andriessen P, van Pul C, de Haan G. Towards Continuous CameraBased Respiration Monitoring in Infants. Sensors. 2021; 21(7):2268. https://doi.org/10.3390/s21072268
Chicago/Turabian StyleLorato, Ilde, Sander Stuijk, Mohammed Meftah, Deedee Kommers, Peter Andriessen, Carola van Pul, and Gerard de Haan. 2021. "Towards Continuous CameraBased Respiration Monitoring in Infants" Sensors 21, no. 7: 2268. https://doi.org/10.3390/s21072268