Non-contact monitoring of respiration in the neonatal intensive care unit

Joao Jorge1, Mauricio Villarroel1, Sitthichok Chaichulee1, Alessandro Guazzi1, Sara Davis2,Gabrielle Green2, Kenny McCormick2 and Lionel Tarassenko2

1 Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford, UK2 Neonatal Unit, John Radcliffe Hospital, Oxford University Hospitals Trust, UK

Abstract— An abnormal respiratory rhythm is an earlyindicator of physiological deterioration. It is of critical impor-tance in the clinical management of critically-ill or prematureinfants, for whom apnoea of prematurity is a major concern.Nevertheless, respiratory signals are still largely disregardedin neonatal intensive care units due to the high prevalence ofnoise and high false alarm rates in conventional monitoring.To address this, we present a novel method for the extractionof respiration from camera-based measurements taken fromthe top-view of an incubator. A total of 107 events from 30neonatal admissions were annotated by three clinical reviewersas either true cessations of breathing (physiologically relevant)or false (artefact-related). The events were divided into twoindependent groups for training and validation and our al-gorithm was trained to classify true cessations. We achieveda good classification performance with 9 out of 10 cessationsand 7 out of 10 artefactual events correctly identified in thetraining set, and with 7 out of 10 cessations and 34 out of 44artefactual events correctly identified in the out-of-sample testset. A reduction in false alarm rate of 77.3% was achieved.

I. INTRODUCTION

A. Remote physiological measurements

Respiratory rate is an important indicator of physiologicaldistress. Conditions such as cardiopulmonary arrest [1],sudden infant death syndrome [2] and other conditions whichlead to changes in the arterial partial pressure of blood gases[3] are often preceded by changes in this vital sign.

In clinical practice, the respiratory function of sponta-neously breathing patients is usually monitored either in-termittently, by manual counting, or continuously, usinginvasive monitoring techniques available in intensive careunits . Flow sensors applied to the nose and/or mouth ordevices that detect chest wall excursions applied to the chestor abdominal area have also been used [4] although not inclinical practice.

In an effort to reduce patient discomfort and make res-piratory measurements more ubiquitous, several methodshave been proposed which do not require contact sensorsto be attached to the patient, such as thermal imaging [5],microwave-based [6] and radar-based [7] methods. Although

JJ acknowledges the RCUK Digital Economy Programme, grant numberEP/G036861/1 (Oxford Centre for Doctoral Training in Healthcare Inno-vation) and the Fundacao para a Ciencia e Tecnologia, Portugal, doctoralgrant SFRH/BD/85158/2012. MV was supported by the Oxford Centre ofExcellence in Medical Engineering funded by the Wellcome Trust andEPSRC under grant number WT88877/Z/09/Z. SD, GG and KMcC weresupported by the NIHR Biomedical Research Centre Programme, Oxford.

technically these are non-contact approaches, they requirespecialist instrumentation and complex set up procedures.

Under the paradigm of remote sensing, the latest decadehas seen the emergence of numerous approaches to themeasurement of vital signs using visible-light video cam-eras. With considerable overlap, most approaches to extractrespiration from visible light imaging can be classified intothose based on either a photoplethysmographic [8]–[15] orcomputer vision [16]–[23] approach to signal extraction andanalysis.

B. Remote monitoring of neonates

The immaturity of the brain mechanisms governing therespiratory rhythm in neonates can result in irregular breath-ing patterns with prolonged pauses (apnoeas). These periodspose a clinical problem particularly in infants of gestationalage below 36 weeks, for whom apnoea of prematurity is acause of recurrent episodes of severe oxygen desaturationoften associated with adverse clinical outcomes [24], [25].

The automatic monitoring of respiration during neonatalsleep is, therefore, of crucial importance both in (a) intensivecare units with high patient-to-staff ratios and in (b) homeenvironments.

Conventional contact devices present clear shortcomingsin these scenarios. Not only do they often require cumber-some equipment but they can also cause infant stress andeven pain due to damage to their fragile skin. Respirationmonitoring in neonatal units is generally done by impedancepneumography (IP). IP is a convenient method if patientsare already monitored by electrocardiography (ECG), butis prone to inaccurate readings due to a number of factorsincluding poor ECG electrode placement, motion artefact,and physiologic events which cause thoracic movementsunrelated to breathing (such as coughing, or crying) [26],[27].

Although several studies in visible-light monitoring ofneonates have so far been advanced, most have been con-ducted for the extraction of heart rate using camera monitors[14], [15], [28]–[31], with few approaches aimed at measur-ing respiratory parameters [31]–[33] in this population.

C. Contributions

In this paper, we describe a novel non-contact means ofrespiratory monitoring and assess its performance for thepurpose of detecting cessation of breathing events (COBEs)in a dataset of spontaneously breathing premature infants.978-1-5090-4023-0/17/$31.00 c©2017 IEEE

Our method includes the following innovations: firstly, skinsegmentation is used in order to localise the subject andselect a region-of-interest (ROI). Secondly, cumulative framedifferences are used to boost the signal-to-noise ratio (SNR)of subtle breathing motions over the ROI. Lastly, we performactivity analysis for each time window to detect the presenceof high-frequency motion, and so exclude these windowsfrom further analysis.

The rest of this paper is organised as follows: Section II-Aintroduces the dataset and II-B the proposed methodology.The experimental results are reported in Section III anddiscussed in Section IV. The main conclusions and directionsfor future work are presented in Section V.

II. MATERIALS AND METHODS

A. Data collection

Data for this study was acquired during an observationalstudy in the Neonatal Intensive Care Unit (NICU) at the JohnRadcliffe Hospital (MONITOR Study - REC: 13/SC/0597).This study involved the monitoring of 30 preterm infantsof less than 37 weeks of corrected postmenstrual age. Eachpreterm infant was recorded under ambient light for up tofour consecutive days.

We obtained a dataset of (a) RGB video footage acquiredusing a digital video camera positioned over the incubatorinside which study infants were nursed (Giraffe Omnibed R©

incubator, General Electric, Fairfield, USA), and (b) conven-tional vital sign data collected concurrently by the patientmonitor as part of routine care. This dataset is supplementedby (c) clinical annotations as described in Section II-C. Themonitoring equipment is shown in Figure 1.

Video acquisition

We used a 3 CCD (Sony ICX274AL R©, Sony, Tokyo,Japan) digital camera (JAI AT-200C R©, JAI, Glostrup, Den-mark) interfaced via Camera Link to a framegrabber card(Microenable IV AD4-CL R©, Silicon Software, Mannheim,Germany). This device was set to acquire 24-bit true colourimages (8-bit per colour) at a pre-set rate of 20.3 framesper second and at a resolution of 1628 × 1236 pixels(with an individual pixel size of 4.4 microns square). Theframe buffers were processed using a Spartan R© FPGA (FieldProgrammable Gate Array) board (Xilinx, San Jose, USA)installed in a 1.90 GHz Intel R© processor workstation with4 × 8 GB RDIMM (registered memory module) and 8 × 4TB SATA disks running in-house software under a Fedora20 Linux operating system.

Patient monitor and sensors

All conventionally-monitored signals were saved on anIntelliVue R© MX800 patient monitor (Philips, Amsterdam,Netherlands) and relayed to a separate work station, astandard PC with a 4 TB disk and 8 GB RAM underWindows 7 (iXELLENCE GMBH, ixTrend R©) via R-232serial transmission.

Fig. 1. Data acquisition apparatus in the Oxford Neonatal Unit: (a)monitoring equipment (left) imaging a mannequin inside the study incubator(centre) and standard patient monitor (right).

Using proprietary software, the infant’s vital signs werederived from the collected signals and reported at a rate of 1estimate per second; heart rate (HR) was derived from 3-leadECG, RR from bipolar impedance pneumogram (IP) signals(both collected using the set of neonatal chest electrodes pro-vided with the monitor at 250 Hz and 62.5 Hz, respectively)or, in the case of spO2 from the photoplethysmographic(PPG) signal acquired by a SET LNCS Neo R© pulse oximeterprobe (Masimo Corporation, Irvine, USA) sampled at 125Hz.

B. Methodology

We propose a new method for the measurement of respi-ration from neonatal video data, which we incorporate intoa COBE detection system.

First, skin pixels are identified in each frame using a skinclassifier and the region-of-interest (ROI) for motion analysisis selected. The purpose of motion analysis is two-fold; weseek to (1) detect breathing movements and estimate theirfrequency and (2) produce a binary signal quality index (SQI)for these estimates based on the level of activity on the videosequence. Finally, the information provided by these twosignals is combined into a COBE classifier (as shown inFigure 2).

Selection of ROI

The pixels in each frame were clustered into skin and non-skin classes based on their a* chroma value using a 2-classclassifier based on Gaussian Mixture models applied to adownsampled a* image. The smallest bounding box whichcontains the largest contiguous skin region in each frame isselected as the ROI. Figure 3 provides some examples of theROIs obtained using this method.

Motion analysis

1) for activity detection: A straightforward way to detectmotion noise inside the incubator, such as that caused byspurious patient movement, nursing interventions or maternal

Fig. 2. Flow diagram illustrating the algorithm for COBE detection in the NICU dataset. The quantities in the diagram are introduced in the main text.

Fig. 3. Selection of ROI for different body poses. The smallest boundingbox containing the segmented skin region is shown in white.

touch, is to monitor the high frequency content in raw pixelintensities.

In this analysis, we averaged the value of the blue colourchannel over the ROI area and processed this signal with ahigh-pass filter (3rd order Butterworth; frequency cut-off at1.6 Hz) designed to attenuate breathing motion as its cut-offfrequency approximates the maximum of the neonatal rangeof respiratory frequencies (100 breaths/min).

The blue channel was selected due to the low penetrationdepth of light in skin over this wavelength range [34].As blue light is more affected by specular reflection thanother colour channels, the intensity signal collected at thiswavelength is much less modulated by the blood-volumepulse, thus rendering it a good proxy for subject motion.

A simple frame-wise measure of activity was obtained byadding the filtered pixel intensities for each frame k. Lastly,this series was filtered using a moving 5-point average filterto achieve a smoother description of activity at the time ofeach frame, a(tk).

2) for breathing detection: During normal sleep, breath-ing causes subtle movements of the thorax and abdomen(and, to a lesser extent, the entire body). These are observableas a periodical variations in the temporal differences betweenconsecutive frames in continuous image data:

∆Ik(i, j) = |Ik(i, j)− Ik−1(i, j)| (1)

where k is the frame index, Ik is the array of intensityvalues over that frame (or, in our case, the blue colourcomponent) and i and j are spatial coordinates.

This approach has been applied successfully in [17], [18]to extract respiratory frequencies of subjects under controlledenvironments for illumination. By thresholding ∆Ik, one canobtain a difference mask. This mask might include some non-subject pixels (e.g. in image regions affected by shadows orother changes in illumination) as well as exclude movingregions with low contrast. To address this, we used thetemporal information over a time interval longer than thesampling period by accumulating the frame differences overthe past N frames, as given in Eq. 2.

ANk (i, j) =

k−N+1∑l=k

∆Ik(i, j) (2)

A reasonable value of N is that which corresponds to theaverage duration of the upslope of the respiratory waveform,a time interval typically associated with inspiration. Toexploit the intrinsic parameters of the system, we relate thisquantity to the respiratory time constant (τe), a property ofthe respiratory system analogous to the time constant of RCelectrical circuits [35]; one time constant is defined as thetime it takes the lung alveoli (capacitor) to discharge 63%of tidal volume (electrical charge) through the respiratoryairways (resistance).

For a premature infant with a lung compliance of 0.005L/cm H2O and an airway resistance of 30 cm H2O/Ls [36],τe is then 0.15 s [36]. Hence, it takes about 0.45 s (3τe) for95% of tidal volume to be exhaled.

In spontaneously breathing infants, there is a ratio ofapproximately 1:2 between the inspiration and expiratoryconstants [37]:

Time, t (min)

0 1 2 3 4 5

IP(a.u)

-5

0

5 X

Time, t (min)

0 1 2 3 4 5

RR

(breaths/min)

0

50

100

Time, t (min)

0 1 2 3 4 5

HR

(beats/min)

50

100

150

200

Time, t (min)

0 1 2 3 4 5

spO

2

(%)

60

80

100

Fig. 4. Segment of IP, RR, HR and spO2 tracings during infant motion(a2). The spO2 tracing shows a small noise-related desaturation at t = 3min. The HR tracing shows no signs of bradycardia. High amounts of noiseare visible in the IP signal before the desaturation (X; from t = 2 min to t =3 min). The horizontal line at RR = 20 breaths/min illustrates the criterionused in Section II-C for the detection of potential COBEs.

τe ≈ 2τi (3)

Thus, at an effective frame rate of 20.3 Hz, the value of Ncorresponding to 3 inspiratory constants is roughly N3τi =5 frames. A camera-based respiratory signal φk was thencomputed as

φk =∑

i,j∈ROIAN3τ

k (i, j) (4)

Finally, respiratory rate RRφ was estimated from 10 swindows of φ (with 1 s overlap) using a 5th order AR modelin the manner described in [11] adjusted to the neonatalrespiratory range.

C. Construction of a data set for validation

For the creation of training and validation subsets, we havegathered a set of 107 time epochs (of 300 s) comprisingepisodes of cessations of breathing (positive events) andinstances of normal breathing (negative events). These wereextracted from a dataset of 30 NICU stays (comprising atotal of 455 monitoring hours) as described below.

In severe cases, episodes of neonatal apnoea can last for 20seconds or longer [38]. The following was then consideredto be a necessary condition for detecting COBEs:

RRIP < 20 breaths/min for a period of at least 20 seconds,

Time, t (min)

0 1 2 3 4 5

IP(a.u)

-2

0

2Y

Time, t (min)

0 1 2 3 4 5

RR

(breaths/min)

0

50

100

Time, t (min)

0 1 2 3 4 5

HR

(beats/min)

50

100

150

200

Time, t (min)

0 1 2 3 4 5

spO

2

(%)

60

80

100

Fig. 5. Segment of IP, RR, HR and spO2 tracings during an apnoeic event in(b1). A cessation of breathing lasting 40 s can be observed in the IP signalfrom t = 2 min (Y), followed by bradycardia around 30 s later; arterialoxygen saturation falls to values near 60% at t = 3 min. The horizontal lineat RR = 20 breaths/min illustrates the criterion used in Section II-C forthe detection of potential COBEs.

where RRIP is is the respiratory rate estimated (at 1 Hz)by the patient monitor using bipolar IP.

The set Ωc of events detected by applying the criterionabove to the dataset was reviewed by a team of threereviewers (including a consultant neonatologist, a clinicalresearch fellow and a neonatal nurse) to ascertain whethereach occurrence was artefactual or an actual COBEs, and (ifso) apnoea-related. For each event, reviewers decided amongthree options: (a) the event was caused by noise or artefactsin the raw signals (b) the event was caused by a cessationof breathing (positive event), or (c) the event was not causedby cessation of breathing nor by artefacts in the raw signals(i.e. RRIP accurately reported a low respiratory rate value).

The (a) set was subdivided into (a2) motion artefacts and(a1) IP artefacts (i.e. related to suboptimal probe placementor other acquisition artefacts which cause the IP signal todeviate from true respiratory effort). The (b) set was subdi-vided into cessations of breathing with (b1) and without (b2)an associated apnoea (i.e. with desaturation and bradycardia).Examples of signal segments for scenarios a2 and b1 aregiven in Figures 4 and 5, respectively.

D. Classification of COBEs

To detect COBEs, RRφ values for times with a(t) > 0were discarded due to the presence of non-breathing motionin the video stream. RRφ was obtained by linearly interpo-lating RRφ over the time intervals between the remaining

Fig. 6. Number of events under each of the labels defined in Section II-C after manual annotation of the collected vital signs and video footage for theset of potential COBEs, Ωc.

segments where these were less than ∆t∗ apart. ∆t∗ is thedetection threshold, a parameter of the classifier. For eachevent, a COBE was deemed to have occurred if

RRφ < 20 breaths/min for a period of at least 20 seconds,

We used ROC analysis to determine ∆t∗ on the trainingset.

III. RESULTS

The clinical reviewers annotated a total of |Ωc| = 107potential COBEs. Figure 6 summarises the findings of thisannotation process. The analysis of potential COBEs bythe clinical reviewers revealed that 82 out of 107 (76.6%)such occurrences were due to the presence of artefacts inthe IP signal (11 due to motion and 71 due to poor signalacquisition). Of the remaining 25 artefact-free events, only 20(28.6% of potential COBEs) corresponded to true cessationsof breathing.

The manually classified events were then distributed atrandom between test and training sets in a manner whichpreserved the balance between the number of elements ineach subclass (Table I). In practice, only a subset of 10negative events was used in each experiment on the trainingset to ensure an equal number of positive and negative eventsin this set.

By following the methodology described, it was possibleto classify the true cessations of breathing. The performance

TABLE IDISTRIBUTION OF ANNOTATED EVENTS IN TRAINING AND TEST SETS,

NUMBER OF EVENTS.

Events, classes Train set Test set Total

IP artifacts, a1 36 35 71Motion artifacts, a2 5 6 11Apneic COBE, b1 8 9 17Non-apneic COBE, b2 2 1 3Low RR, c 2 3 5Total, Ωc 53 54 107

Positives, P = a1 ∪ a2 ∪ c 10 10 20Negatives, N = b1 ∪ b2 43 44 87Total, Ωc 53 54 107

TABLE IIPERFORMANCE IN COBE DETECTION. NUMBER OF CORRECTLY

IDENTIFIED EVENTS FOR EACH CLASS. NUMBERS ARE GIVEN FOR EACH

SUBCLASS UNDER THE TEST SET.

Train set Test set

P N P N(/10) (/10) (/10) (/44)

9 7 7 34b1 b2 a1 a2 c(/9) (/1) (/35) (/6) (/3)

6 1 26 6 2

Time, t (s)

0 20 40 60 80 100 120

Activity,

a

(a.u.)

0

0.5

Time, s (s)

0 20 40 60 80 100 120

φ

(a.u)

-2

0

2

Time, t (s)

0 20 40 60 80 100 120

RR

φ

(breaths/min)

0

50

100

Time, t (s)

0 20 40 60 80 100 120

IP(a.u)

-5

0

5

Time, t (s)

0 20 40 60 80 100 120

RR

IP

(breaths/min)

0

50

100

Fig. 7. Segment of activity, camera-derived respiratory signal (φ), IP and respiratory rates (RRφ

and RRIP ) during an apnoeic event. The second plotillustrates the application of the criterion defined in Section II-D for the deletion of segments corrupted by motion artefact as shown in the activity plota(t). The intervals between motion events, where a(t) is very low, correspond to periods of quiet sleep. A cessation of respiration (marked by both lowa(t) and RR

φ< 20 breaths/min) can be observed in the camera-derived respiratory signal from t = 65 s to t = 105 s.

of the proposed method for COBE detection on the trainingand test sets is reported in Table II.

The majority of IP artefacts (a1) in the test set werecorrectly identified as negative events (26 out of 35 events)and all the movement-related artefacts (a2) in the sameset were correctly identified (6 events). Regarding positiveevents (b1 and b2), 7 out of 10 COBEs in the test set weredetected. In conjunction, these results yield a reduction inthe false alarm rate of 77.3% on the test set.

Figure 7 shows activity a(t) and respiratory signals duringa clinically validated episode of neonatal apnoea. We observethat motion events manifest as perturbations in the activitysignal a(t). After motion events, marked by a(t) > 0 (shownas grey-shaded areas in the top panel) are excluded, the lowerintensity changes in the camera-derived signal due to therespiratory cycle become apparent (as illustrated by the φplot).

During the apnoeic episode, RRφ (shown in the thirdpanel of the same Figure) drops from a resting rate of50 breaths/min to assume values below the threshold setin Section II-D for the detection of true cessations (20

breaths/min) for a period of approximately 40 seconds,therefore generating a positive detection.

IV. DISCUSSION

Methods for measuring motion based on temporal differ-encing can adapt rapidly to changes in background, envi-ronment lighting or spurious motion. However, the smallamplitude of abdominal and chest movements associatedwith neonatal breathing (typically in the order of a few mil-limetres) means that these can be imperceptible between con-secutive frames. We successfully address this issue throughthe use of accumulated temporal differences. The relativelylow computational cost of computing temporal differenceswhen compared with other techniques for estimating motionbased on optical flow makes our approach ideally suited forreal-time implementation.

A resting respiratory rate below 20 breaths/min is abnor-mal in premature infants (Figure 8). Such low rates couldindicate a progression into an apnoeic event, and therefore,they should prompt immediate clinical attention. Our studyin the Oxford NICU revealed that a high humber of these oc-

Respiratory rate, RR (breaths/min)

0 50 100

No.data

points

×104

0

2

4

6

8

Fig. 8. Histogram of respiratory rate values derived by the patient monitorfrom IP for the duration of the study (approximately 455 hours). + is thedata mean, the boxplot bounds the 25% and 75% quartiles and the whiskersbound the 9% and 91% of the data.

currences (76.6%) are due to the presence of artefacts in theIP signal, the standard modality for monitoring respirationin neonatal units. This poses a clear problem as high falsealarm rates may desensitize NICU staff and reduce reactionspeed by alarm fatigue [39].

Through the analysis of the video signals acquired fromthe inside of the cot, we classified the potential cessation ofbreathing events in the dataset. By identifying where high-frequency movement occurs, it was possible to distinguishbetween subtle body movements associated with the respi-ratory effort from spurious motion inside the cot. Furtheranalysis of the segments of low activity allowed us to classifybreathing versus non-breathing behaviour.

A meaningful reduction in the number artefact-relatedevents of was achieved. All the movement-related artefacts(a2) and the majority of IP artefacts (a1) in an out-of-sampletest set were correctly identified as artefactual detections.Given that the detection of true cessations is paramount inpatient care, it is of critical importance that any increasespecificity is not achieved at the expense of the capacityto detect positive events (b1 and b2). In the work presentedhere, 7 out of 10 COBEs in the test set were detected. Oncloser inspection, 2 of these cases were found to correspondto periods when the monitored infant was undergoing photo-therapy. This procedure involves prolonged exposure to LEDlight predominantly in the blue region of the spectrum (460- 490 nm) to an irradiation level of 30 to 40 µW/cm2/nm.This causes bright regions in the acquired images to becomesaturated, thus decreasing SNR over these regions and de-pressing the a(t) signal.

V. CONCLUSIONS AND FUTURE WORKIn this paper, we presented a novel technique to obtain

activity and respiratory signals from neonatal infants nursedinside an incubator through the temporal processing of theacquired video. Our approach combines these two signals toobtain a more reliable respiratory signal and identify periodsof its cessation.

Our results show that the method presented can success-fully distinguish between artefactual and true decreases inRR in a range of critically low values for this variable (RR

< 20 breaths/min). Although sample size was small, thefeasibility of our approach based on motion analysis wasclearly demonstrated.

Conventional neonatal monitors could, therefore, benefitgreatly from the collection of video signals. Specifically,the implementation of video-based COBE detection systemscould potentially help reduce the high rate of false apnoeaalarms in these units. Hence, we believe our findings shouldmotivate extensive validation of such techniques.

In this work, we have used a simple colour-based skindetector to segment the ROI. To build in the capability forcontinuous monitoring over longer periods of time, thereis a need for more robust subject segmentation. Methodswhich include additional features such as texture, and opticalflow with high level features such as pose and knowledgeof anatomical structures could be explored. Recent develop-ments in deep learning research have provided a frameworkfor embedding visual features within convolutional neuralnetworks to yield highly accurate classifiers [40].

Finally, the possibility to extract additional vital signsusing the same camera-based device would also offer signif-icant advantages. In particular, the ability to obtain camera-derived estimates of HR and spO2, in addition to RR, duringepisodes of bradycardia and desaturation, would leveragethe system presented here from a COBE detector to anapnoea detector, thereby paving the way for truly non-contactsurveillance of apnoea-related cessations of breathing.

VI. ACKNOWLEDGMENTSThe authors gratefully acknowledge Sheula Barlow and

Sharon Baron for their contribution to the clinical annotationof COBEs. We would like thank the FG’17 reviewers fortheir comments.

REFERENCES

[1] J. F. Fieselmann, M. S. Hendryx, C. M. Helms, and D. S. Wakefield,“Respiratory rate predicts cardiopulmonary arrest for internal medicineinpatients,” Journal of General Internal Medicine, vol. 8, no. 7,pp. 354–360, 1993.

[2] A. Steinschneider, “Prolonged apnea and the sudden infant deathsyndrome: clinical and laboratory observations,” Pediatrics, vol. 50,no. 4, p. 646, 1972.

[3] M. A. Cretikos, R. Bellomo, K. Hillman, J. Chen, S. Finfer, andA. Flabouris, “Respiratory rate: the neglected vital sign.,” MedicalJournal of Australia, vol. 188, no. 11, pp. 657–659, 2008.

[4] M. Folke, L. Cernerud, M. Ekstrom, and B. Hok, “Critical reviewof non-invasive respiratory monitoring in medical care.,” Medical &Biological Engineering & Computing, vol. 41, no. 4, pp. 377–83, 2003.

[5] M. Garbey, N. Sun, A. Merla, and I. Pavlidis, “Contact-Free Mea-surement of Cardiac Pulse Based on the Analysis of Thermal Im-agery,” IEEE Transactions on Biomedical Engineering, vol. 54, no. 8,pp. 1418–1426, 2007.

[6] O. Baltag, “Microwave Doppler transducer for noninvasive monitoringof the cardiorespiratory activity,” in IEEE Transactions on Magnetics,vol. 44, pp. 4484–4487, 2008.

[7] C. Li, J. Cummings, J. Lam, E. Graves, and W. Wu, “Radar remotemonitoring of vital signs,” IEEE Microwave Magazine, vol. 10, no. 1,pp. 47–56, 2009.

[8] F. P. Wieringa, F. Mastik, and A. F. W. van der Steen, “Contactlessmultiple wavelength photoplethysmographic imaging: a first step to-ward ”SpO2 camera” technology.,” Annals of Biomedical Engineering,vol. 33, no. 8, pp. 1034–41, 2005.

[9] C. Takano and Y. Ohta, “Heart rate measurement based on a time-lapseimage.,” Medical Engineering Physics, vol. 29, no. 8, pp. 853–857,2007.

[10] W. Verkruysse, L. O. Svaasand, and J. S. Nelson, “Remote plethysmo-graphic imaging using ambient light,” Optics Express, vol. 16, no. 26,pp. 21434–21445, 2008.

[11] L. Tarassenko, M. Villarroel, A. Guazzi, J. Jorge, D. a. Clifton, andC. Pugh, “Non-contact video-based vital sign monitoring using am-bient light and auto-regressive models.,” Physiological Measurement,vol. 35, no. 5, pp. 807–31, 2014.

[12] M.-z. Poh, D. J. McDuff, and R. W. Picard, “Advancements in Noncon-tact, Multiparameter Physiological Measurements Using a Webcam,”IEEE Trans Biomed Eng, vol. 58, no. 1, pp. 7–11, 2011.

[13] F. Bousefsaf, C. Maaoui, and A. Pruski, “Continuous wavelet filteringon webcam photoplethysmographic signals to remotely assess theinstantaneous heart rate,” Biomedical Signal Processing and Control,vol. 8, no. 6, pp. 568–574, 2013.

[14] L. K. Mestha, S. Kyal, B. Xu, and L. Edward, “Towards continuousmonitoring of pulse rate in neonatal Intensive care unit with a web-cam,” Institute of Electrical and Electronics Engineers Inc., pp. 3817–3820, 2014.

[15] S. Fernando and G. D. Haan, “Feasibility of contactless pulse ratemonitoring of neonates using Google Glass,” in Mobile Health, vol. 1,pp. 3–7, 2015.

[16] K. Nakajima, A. Osa, and H. Miike, “A method for measuringrespiration and physical activity in bed by optical flow analysis,” inEngineering in Medicine and Biology Society, 1997. Proceedings ofthe 19th Annual International Conference of the IEEE, vol. 2054,pp. 2054–2057, 1997.

[17] K. S. Tan, R. Saatchi, H. Elphick, and D. Burke, “Real-time vi-sion based respiration monitoring system,” in IEEE Transactions onBiomedical Engineering, vol. 53, pp. 770–774, 2010.

[18] Y. Bai, W. Li, and C.-h. Yeh, “Design and implementation of anembedded monitor system for body breath detection by using imageprocessing methods,” 2010 Digest of Technical Papers InternationalConference on Consumer Electronics, pp. 2–4, 2010.

[19] M. Yu, J. Liou, and S. Kuo, “Noncontact respiratory measurementof volume change using depth camera,” in 34th Annual InternationalConference of the IEEE EMBS, pp. 2371–2374, 2012.

[20] N. Burba and M. Bolas, “Unobtrusive measurement of subtle non-verbal behaviors with the Microsoft Kinect,” IEEE VR Workshop onAmbient Information Technologies, 2012.

[21] J. Xia and R. A. Siochi, “A real-time respiratory motion monitoringsystem using KINECT: proof of concept.,” Medical physics, vol. 39,no. 5, pp. 2682–5, 2012.

[22] T. Lukac, J. Pucik, and L. Chrenko, “Contactless Recognition of Res-piration Phases Using Web Camera,” in 24th International ConferenceRadioelektronika, pp. 1–4, 2014.

[23] M. H. Li, A. Yadollahi, and B. Taati, “A non-contact vision-basedsystem for respiratory rate estimation,” IEEE Engineering in Medicineand Biology Society, vol. 2014, pp. 2119–2122, 2014.

[24] R. J. Martin and C. G. Wilson, “What to do about apnea of prematu-rity?,” Journal of Applied Physiology, vol. 107, no. 4, pp. 1015–1016,2009.

[25] C. F. Poets, “Interventions for apnoea of prematurity: A personalview,” Acta Paediatrica, International Journal of Paediatrics, vol. 99,no. 2, pp. 172–177, 2010.

[26] T. H. Warburton D, Stark AR, “Apnea monitor failure in infants withupper airway obstruction,” Pediatrics, vol. 60, no. 5, pp. 742–4, 1977.

[27] R. T. Brouillette, A. S. Morrow, D. E. Weese-Mayer, and C. E. Hunt,“Comparison of respiratory inductive plethysmography and thoracicimpedance for apnea monitoring,” The Journal of Pediatrics, vol. 111,no. 3, pp. 377–383, 1987.

[28] L. Scalise and N. Bernacchia, “Heart rate measurement in neonatal pa-tients using a webcamera,” IEEE International Symposium on MedicalMeasurements and Applications Proceedings, pp. 1–4, 2012.

[29] L. Aarts, V. Jeanne, J. P. Cleary, C. Lieber, J. S. Nelson, S. BambangOetomo, and W. Verkruysse, “Non-contact heart rate monitoringutilizing camera photoplethysmography in the neonatal intensive care

unit - A pilot study.,” Early Human Development, vol. 89, no. 12,pp. 943–8, 2013.

[30] J. H. Klaessens, M. van den Born, A. van der Veen, J. Sikkens-van deKraats, F. a. van den Dungen, and R. M. Verdaasdonk, “Developmentof a baby friendly non-contact method for measuring vital signs: Firstresults of clinical measurements in an open incubator at a neonatalintensive care unit,” Advanced Biomedical and Clinical DiagnosticSystems XII SPIE Proceedings, vol. 8935, p. 89351P, February 2014.

[31] M. Villarroel, A. Guazzi, J. Jorge, S. Davis, P. Watkinson, G. Green,K. McCormick, and L. Tarassenko, “Continuous non-contact vital signmonitoring in the neonatal intensive care unit,” Healthcare TechnologyLetters, vol. 1, no. 3, pp. 87–91, 2014.

[32] R. Janssen, W. Wang, A. Moco, and G. de Haan, “Video-basedrespiration monitoring with automatic region of interest detection,”Physiological Measurement, vol. 37, no. 1, pp. 100–114, 2016.

[33] M. Gastel, S. Stuijk, and H. G, “Robust respiration detection fromremote photoplethysmography,” Biomedical Optics Express, vol. 7,no. 12, pp. 107–120, 2016.

[34] W. J. Cui, L. E. Ostrander, and B. Y. Lee, “In vivo reflectance of bloodand tissue as a function of light wavelength.,” IEEE transactions onbiomedical engineering, vol. 37, no. 6, pp. 632–9, 1990.

[35] A. S. Sedra and K. C. Smith, Microelectronic Circuits. OxfordUniversity Press, 2004.

[36] E. Bancalari, “Inadvertent positive end-expiratory pressure duringmechanical ventilation,” The Journal of Pediatrics, vol. 108, no. 4,pp. 567–569, 1986.

[37] J. P. Goldsmith and E. H. Karotkin, “Introduction To Assisted Venti-lation,” Assisted Ventilation of the Neonate, pp. 1–14, 2003.

[38] E. C. Eichenwald, “Apnea of Prematurity,” Pediatrics, vol. 137, no. 1,pp. e2–e7, 2016.

[39] V. Monasterio, F. Burgess, and G. D. Clifford, “Robust classificationof neonatal apnoea-related desaturations,” Physiological Measurement,vol. 33, no. 9, pp. 1503–1516, 2012.

[40] S. Chaichulee, M. Villarroel, J. Jorge, C. Arteta, G. Green, K. Mc-cormick, A. Zisserman, and L. Tarassenko, “Multi-task ConvolutionalNeural Network for Patient Detection and Skin Segmentation inContinuous Non-contact Vital Sign Monitoring (under review),” inIEEE Face and Gesture Recognition, 2017.