Non-Contact, Simple Neonatal Monitoring by ... - MDPI

sensors Article

Non-Contact, Simple Neonatal Monitoring by Photoplethysmography Juan-Carlos Cobos-Torres 1, * , Mohamed Abderrahim 2 1 2 3

*

and José Martínez-Orgado 3

Postgraduate Subdirection, Catholic University of Cuenca, Cuenca 4X8V+5F, Ecuador Department of Systems Engineering and Automation, University Carlos III of Madrid, Leganes 28911, Spain; [email protected] Neonatology, San Carlos Clinical Hospital, Madrid 28040, Spain; [email protected] Correspondence: [email protected]; Tel.: +593-7-404-1654

Received: 15 October 2018; Accepted: 3 December 2018; Published: 10 December 2018







Abstract: This paper presents non-contact vital sign monitoring in neonates, based on image processing, where a standard color camera captures the plethysmographic signal and the heart and breathing rates are processed and estimated online. It is important that the measurements are taken in a non-invasive manner, which is imperceptible to the patient. Currently, many methods have been proposed for non-contact measurement. However, to the best of the authors’ knowledge, it has not been possible to identify methods with low computational costs and a high tolerance to artifacts. With the aim of improving contactless measurement results, the proposed method based on the computer vision technique is enhanced to overcome the mentioned drawbacks. The camera is attached to an incubator in the Neonatal Intensive Care Unit and a single area in the neonate’s diaphragm is monitored. Several factors are considered in the stages of image acquisition, as well as in the plethysmographic signal formation, pre-filtering and filtering. The pre-filter step uses numerical analysis techniques to reduce the signal offset. The proposed method decouples the breath rate from the frequency of sinus arrhythmia. This separation makes it possible to analyze independently any cardiac and respiratory dysrhythmias. Nine newborns were monitored with our proposed method. A Bland-Altman analysis of the data shows a close correlation of the heart rates measured with the two approaches (correlation coefficient of 0.94 for heart rate (HR) and 0.86 for breath rate (BR)) with an uncertainty of 4.2 bpm for HR and 4.9 for BR (k = 1). The comparison of our method and another non-contact method considered as a standard independent component analysis (ICA) showed lower central processing unit (CPU) usage for our method (75% less CPU usage). Keywords: image processing; neonatology; imaging photoplethysmography; vital signs; respiratory sinus arrhythmia

1. Introduction There are many different invasive and non-invasive methods of measurements in the context of neonatal pathology and therapeutic interventions. Among the non-invasive techniques for heart rate monitoring, the worldwide gold standard, is the electrocardiograph (ECG). However, recording the electrical potential generated by the heart requires proper electrode placement, which may interfere with the baby’s movements and interaction with parents and/or caretakers. Surprisingly, even today disadvantaged areas still lack quality electrodes, complicating their placement, causing skin lesions. Additionally, electrode misplacement may produce skin lesions or ischemia if wrapped around a limb. These disadvantages can lead to a work overload for health personnel who must revise the electrodes regularly [1].

Sensors 2018, 18, 4362; doi:10.3390/s18124362

www.mdpi.com/journal/sensors

Sensors 2018, 18, 4362

2 of 14

Another important measurement method is the finger pulse oximeter with a plethysmography wave; it measures heart rate (HR)and oxygen saturation in the blood but can cause problems like those caused by an ECG electrode including skin irritations. New oximeters also provide breath rate (BR) information linked to respiratory sinus arrhythmia. The third method for measuring heartbeat and breathing frequencies is based on the piezoelectric effect and involves placing a sensor on the abdominal area [2]. The above methods are the most commonly used for monitoring vital signs and require contact sensors to be placed on the patient, whereas new trends seek to allow non-contact monitorization. Newer methods for measuring heart and respiration rates include thermal imaging analysis [3,4], observation of the Doppler effect [5], Doppler-camera hybrid [6], and imaging photoplethysmography. The last technique analyzes changes in the color intensity and optical properties of a selected area of the skin that is affected by variations in blood flow or by pulsating vibrations of vital signs. Robust methods have been developed to cope with variations in light intensity but they can still be affected by the type of light source or flickering. These methods are called photoplethysmographic images (PPGI). One method that works under ambient lighting focuses on the human face but can only provide adequate results with stable illumination [7]. A more robust method was based on source separation through the application of an independent component analysis (ICA) [8]. It has a high computational cost and the original signals must be independent; something which does not happen if breath rate is measured based on sinus arrhythmia. Another relevant method is Eulerian magnification [9], which uses an extra layer of pre-processing that amplifies the redness of the person’s skin in order to measure the HR. This also carries a heavy computational cost and requires an initial measuring point to provide a more precise approximation. An HR extraction system was built and tested using video footage of three hospitalized neonates [9,10]. The system produced a good measurement of HR from real-world videos taken under good lighting conditions and when the subjects’ movements were smooth; nevertheless, the degree of agreement or concordance (e.g., the Bland-Altman method) between PPGI measurements and the pulse oximetry measurements was not provided. After reviewing the principal PPGI methods, we considered their suitability for monitoring in a Neonatal Intensive Care Unit (NICU) environment. This is important because possible areas or regions for measurements have not been determined. Additionally, newborns’ movements are gentle and subtle but they may also have spasms that generate artifacts or false measurements. Therefore, such movements can affect the accuracy of photoplethysmographic measurements. In addition, NICU lighting may be different in every hospital. Consequently, below we describe different PPGI methods that have been implemented in NICUs. The Mestha method [11] for continuous HR monitoring has a low accuracy in environments with excessive movement. This is due to the use of a high pass finite impulse response (FIR) filter to remove very low-frequency components. The FIR filters do not work properly with low frequencies and are not optimal given the large number of coefficients they require. Additionally, the published study performed measurements in an environment with excellent stable lighting (300 lux), so the results cannot demonstrate that the method would work properly under poor conditions. Another method based on an ICA analysis has been presented by Scalise [12] but is also characterized by a high computational cost. Moreover, in that report, the light conditions were stable and always under control according to the authors. The authors affirm that the camera was placed perpendicularly in front of the neonate’s face at a distance of 20 cm. However, this is not reflected in the example of the picture provided and some light reflections are visible on the transparent wall of the incubator. The method based on joint-time-frequency diagrams (JTFD) presented by Aarts [13] reports an accurate measurement of the plethysmographic signal but no details were given regarding signal processing or filtering. In addition, the method is not robust when faced with artifacts produced by motion and changes in illumination.

Sensors 2018, 18, 4362

3 of 14

The Wang method [14]REVIEW introduces a mathematical model. They analyze the relevant optical Sensors 2018, 18, x FOR PEER 3 of 14 and physiological properties of skin reflections. This is important in order to increase the understanding of The Wang methodbehind [14] introduces a mathematical model. They analyzemodel the relevant optical and the algorithmic principles PPGI. However, the PPGI mathematical is only based on the physiological properties of skin reflections. This is important in order to increase the understanding assumption of a single light source with a constant spectrum. of the algorithmic principles PPGI. However, the PPGI mathematical is only based on Finally, the method basedbehind discrete wavelet decomposition with a model periodicity-based voting the assumption of a single light source with a constant spectrum. scheme [15]. This method relies on the convolution of the original signal with FIR filter structures. Finally, the method based discrete wavelet decomposition with a periodicity-based voting As already explained, FIR filters do not work properly with low frequencies. scheme [15]. This method relies on the convolution of the original signal with FIR filter structures. As After reviewing number developed forfrequencies. PPGI, we have designed a robust and already explained,aFIR filters of do the not methods work properly with low efficient method that works under changing light conditions and in we thehave presence of moderate motion After reviewing a number of the methods developed for PPGI, designed a robust and artifacts. In addition, the system has a computationally lightweight algorithm that is appropriate for online efficient method that works under changing light conditions and in the presence of moderate motion artifacts.the In method addition,can the be system has a computationally lightweight algorithm is appropriate for use. Finally, implemented as part of a monitoring system inthat a NICU environment. online use. Finally, the method can be work implemented part a monitoring system in a NICU The method presented in the current is basedason theofanalysis of color intensity variations environment. in a given area on the newborn infant’s diaphragm. A pre-filtering step based on numerical analysis The method presented in the current work is based on the analysis of color intensity variations techniques was adopted to reduce the offset of the averaged signal. The proposed non-invasive in a given area on the newborn infant’s diaphragm. A pre-filtering step based on numerical analysis approach employs a standard camera and image processing to estimate the desired variables. techniques was adopted to reduce the offset of the averaged signal. The proposed non-invasive In addition, the method for BR measurement is decoupled from the sinus arrhythmia frequency. approach employs a standard camera and image processing to estimate the desired variables. In addition, the method for BR measurement is decoupled from the sinus arrhythmia frequency.

2. Materials and Methods

2. Materials Methods The overall and scheme adopted for the processing of the images and estimation of physiological variables The is illustrated in Figure 1. The method consists of images taking and the estimation images from a video camera, overall scheme adopted for the processing of the of physiological variables is them illustrated in Figure 1.an The method consists of taking the images from a video camera, pre-processing by selecting area of interest and separating each image into its color pre-processing them by of selecting anofarea of interest and separatingNext, each the image intoinitseach color components. The selection the area interest is done manually. pixels frame components. The selection of the area of interest is done manually. Next, the pixels in each frame are are averaged. Three stacks are generated with each averaged value and with a given length (last six averaged. Three stacks generated with each averaged value and with given lengthThe (lastfilter six has seconds of the video). Theseare signals are filtered while the stacks continue toabe updated. seconds of the video). These signals are filtered while the stacks continue to be updated. The filter two steps: an offset filter and narrow-band filters. The filtered signal is analyzed through a short-time has two steps: an offset filter and narrow-band filters. The filtered signal is analyzed through a shortFourier transform (STFT) (to find the frequency components of the signal). time Fourier transform (STFT) (to find the frequency components of the signal).

Figure 1. Sequence used for measurement of heartbeat and breathing.

Sensors 2018, 18, x FOR PEER REVIEW

Sensors 2018, 18, 4362

4 of 14

Figure 1. Sequence used for measurement of heartbeat and breathing.

4 of 14

The first segment in Figure 1 corresponds to the video capture in the area of analysis and the details of first this process 2.2. The segment presents signal for the The segmentare in given Figurein1 Section corresponds to second the video capture in the area offormation analysis and analysis. segment represents the2.2. steps filter the signalpresents image and then extract the details of The this third process are given in Section Thetosecond segment signal formation for frequencies of interest. A description of the processes involved in segments two and three is given in the analysis. The third segment represents the steps to filter the signal image then extract the Sections 2.3 of and 2.4. The segment summarizes the steps in forsegments signal processing and estimation frequencies interest. A fourth description of the processes involved two and three is given in of the physiological steps in the methodthe are explained further in the next of Sections 2.3 and 2.4. variables. The fourthThe segment summarizes steps for signal processing and sections estimation the paper. of the physiological variables. The steps in the method are explained further in the next sections of the paper. 2.1. Patient Population 2.1. Patient Population The study was performed in accordance with the Declaration of Helsinki regarding studies on Thesubjects study was in accordance with the Declaration of Helsinki regarding on human and performed was approved by the ethics committee of the Hospital Vicente Corralstudies Moscoso. human subjects and was approved by the ethics committee of the Hospital Vicente Corral Moscoso. The newborns who took part in this study, with the informed consent of their parents, were nine The newborns part in this study,weeks, with the informed their parents, were nine preterm infantswho withtook a gestation of 25–40 a weight of consent 500 g orofgreater, and without any preterm infantsPatient with a safety gestation of never 25–40 weeks, a weight of 500 the g ormedical greater, protocols and without any complications. was compromised because and/or complications. Patient safety was never compromised because the medical protocols and/or protective protective measures were not changed. Several recordings of each newborn were made and stored measures were not changed. To Several recordings of each newborn andofstored for later for later analysis in segments. explain the proposed method, datawere frommade only one the newborns analysis in used. segments. explain the proposed data from only one of the with newborns have have been All ofTo the newborns’ samples method, were used to validate the method the Blandbeen used. All of the newborns’ samples were newborns used to validate the method with theand Bland-Altman Altman technique. During the measurements, were inside the incubator the camera technique. the measurements, was placedDuring at an approximate distancenewborns of 50 cm.were inside the incubator and the camera was placed at an approximate distance of 50 cm. 2.2. Set-Up and Experimental Procedures 2.2. Set-Up and Experimental Procedures The experimental setup consisted of a digital camera placed outside the glass of the incubator Thetargeting experimental setup consisted of a digital placed the glass of the incubator directly the newborn and a second cameracamera focusing on theoutside vital signs monitor, to check and directly targeting the newborn and a second camera focusing on the vital signs monitor, to check and synchronize the variables being measured, as shown in Figure 2. synchronize the variables being measured, as shown in Figure 2. d d c

a

b

Figure 2. Setup of the experiment: (a) neonate, (b) Neonatal Intensive Care Unit (NICU Incubator, Figure 2. Setup of the experiment: (a) neonate, (b) Neonatal Intensive Care Unit (NICU Incubator, (c) (c) vital signs monitor, and (d) cameras. vital signs monitor, and (d) cameras.

Sensors 2018, 18, 4362

5 of 14

Measurements were made in the NICU of the University Hospital Puerta de Hierro-Majadahonda and the Hospital Vicente Corral Moscoso. The only light sources were the natural light from windows and skylights in the roof and/or artificial light from fluorescent lamps. The video sequences of the neonates were recorded with a digital camera. The resolution of the videos is 1920 × 1080 or 1280 × 720 pixels and the frame rate is 24 or 30 frames per second (fps). Image sequences were saved in AVI or MP4 format. The measurements corresponding to the vital signs were collected from the vital signs monitor. As a reference, the second camera recorded the vital signs from the monitor screen, heart rate and respiration rate. The two cameras were paired to start the recordings simultaneously. A simple script in Matlab analyzed the videos frame by frame. The script stored the readings at a frequency of 30 Hz. The values are averaged every second. The averaged value will be the reference value to be compared with our proposed method. The NICU monitor was Dräguer Infinity Delta. 2.3. Pre-Processing The images from the different video segments were exported to MATLAB® for post-processing and pre-filtering. The first step was to select the area of interest for analysis. The BR was visible in the abdominal and thoracic movements of newborns [16]. The abdominal area was chosen because it contains information reflecting the heartbeat and breath rates. These two vital signs cause changes in the intensity level of pixel brightness in the images. The measurements of these two signs are independent and unrelated to the sinus arrhythmia. The area in question was 40 × 40 pixels, from which the image was separated into three RGB-channels. The photoplethysmographic information contained in this area was sufficient to measure the vital signs. The input R-channel had higher intensities; it did not provide more plethysmographic information but it was important for analysis in poor light conditions. The RGB intensity value ranged from 0–255 for each of the RGB components of each pixel in the color image. Since not all of the pixels in the selected region of interest contained a variation in the brightness of the plethysmographic signal, the pixels were combined into a single average brightness value, Equations (1)–(3). " # h 1 w averageR = (1) ∑ Rxy wh x∑ =1 y =1 1 w averageB = wh x∑ =1 1 w averageG = wh x∑ =1

"

"

h

∑ Bxy

# (2)

y =1 h

∑ Gxy

# .

(3)

y =1

The steps of this process are summarized schematically in Figure 3. Additionally, as shown in Figure 4a, the image for the region of interest (ROI) analysis; Figure 4b the neonate in the incubator with a mark for the analysis area (red box); Figure 4c, the offset in the signal was very high, due to the illumination, pigmentation of the skin, and the distance from the camera. This offset was filtered and treated as part of the low-frequency components. The signal trend was calculated using numerical methods. First, the straight line that best adjusts to the signal was obtained by least squares. Then, the trend signal was generated through a moving average.

R G B

Sensors 2018, 18, 4362 Sensors 2018, 18, x FOR PEER REVIEW

RGB CHANNELS (40x40 pixels)

STACK OF AVERAGES Sliding window

FRAME AVERAGE

6 of 14 6 of 14

R G B

Area selected manually RGB

FRAME

STACK

CHANNELS OF AVERAGE Figure 3. Illustration of processing: area selection, separation of RGB-channels, an average of each (40x40 pixels) AVERAGES signal.

Additionally, as shown in Figure 4a, the image for the region of interest (ROI) analysis; Figure 4b the neonate in the incubator with a mark for the analysis area (red box); Figure 4c, the offset in the signal was very high, due to the illumination, pigmentation of the skin, and the distance from the Area selected camera. This offset was filtered and treated as part of the low-frequency components. The signal manually trend was calculated using numerical methods. First, the straight line that best adjusts to the signal Figure Illustrationofofprocessing: processing:area area selection, separation of RGB-channels, an average of Figure 3. 3. Illustration selection, separation of RGB-channels, an average of each was obtained by least squares. Then, the trend signal was generated through a moving average. each signal. signal.

Additionally, as shown in Figure 4a, the image for the region of interest (ROI) analysis; Figure 4b the neonate in the incubator with a mark for the analysis area (red box); Figure 4c, the offset in the signal was very high, due to the illumination, pigmentation of the skin, and the distance from the camera. This offset was filtered and treated as part of the low-frequency components. The signal trend was calculated using numerical methods. First, the straight line that best adjusts to the signal was obtained by least squares. Then, the trend signal was generated through a moving average.

Figure 4. Capture of the plethysmographic signal processing with Matlab, (a) region of interest (ROI), Figure 4. Capture of the plethysmographic signal processing with Matlab, (a) region of interest (ROI), (b) image analysis, and (c) graph average signals of the RGB-channels. (b) image analysis, and (c) graph average signals of the RGB-channels.

2.4. Post-Processing 2.4. Post-Processing The first second of the video segment was removed because the first moments are always unstable The first second of images the video segmentthan was normal). removedAfter because the first arethe always in a video recording (the are brighter calculating themoments average for ROI unstable in a video recording (the images arefirst brighter thangenerated. normal). After calculating for and each color channel, FIFO stacks (First in, out) were These stacks hadthe anaverage equivalent the ROI and each color channel, FIFO stacks (First in, first out) were generated. These stacks had an of 6 s of information. This period was selected to obtain a minimum of three heartbeats and three equivalent of 6 s of information. This period was selected to obtain a minimum of three heartbeats Figure 4. Capture the plethysmographic signal processing with Matlab, (a) were regionfiltered of interest (ROI), breaths. While a stackofcontinued to be updated, the low-frequency signals from the full and three breaths. While stack tothe beof updated, the the low-frequency signals were filtered (b) image analysis, anda(c) graphcontinued average on signals the RGB-channels. stack. The sources of these signals depend distance from camera as well as the lighting of from the full stack. The sources of these signals depend on the distance from the camera as well as the skin. These low frequencies were estimated using numerical analysis techniques to reduce the 2.4. Post-Processing the lighting of the skin. These low frequencies were estimated using numerical analysis techniques displacement signal. First, the least squares method was applied to find the linear function that best to reduce the displacement signal. First, the from least the squares method appliedthe to offset. find the linear fits the signal. This function subtracted original signalwas to reduce The first second of the was video segment was removed because the first moments areSecond, always function that best fits the signal. This function was subtracted from the original signal to reduce the large variations were smoothed byimages calculating quotients with the signal obtaining unstable in a video recording (the are brighter than normal). Afterdeviations. calculatingAfter the average for the signal, narrow band filters (type impulse response (IIR)) thepre-filtered ROI and each colortwo channel, FIFO stacks (First in,infinite first out) were generated. Theseattenuated stacks hadthe an frequencies outside band of interest. Thesewas bands are detailed in Table 1. equivalent of 6 s ofthe information. This period selected to obtain a minimum of three heartbeats and three breaths. While a stack continued to be updated, the low-frequency signals were filtered from the full stack. The sources of these signals depend on the distance from the camera as well as the lighting of the skin. These low frequencies were estimated using numerical analysis techniques to reduce the displacement signal. First, the least squares method was applied to find the linear function that best fits the signal. This function was subtracted from the original signal to reduce the

Sensors 2018, 18, x FOR PEER REVIEW

7 of 14

offset. Second, large variations were smoothed by calculating quotients with the signal deviations. After obtaining the pre-filtered signal, two narrow band filters (type infinite impulse response (IIR)) Sensors 2018, 18, 4362 attenuated the frequencies outside the band of interest. These bands are detailed in Table 1. 7 of 14 Table1. 1. Band Band of of interest. interest. Table Vital Signs Vital Signs HR BR

HR BR

Lower limit (bpm) 110 (1.83 Hz) 110 Hz) 30 (1.83 (0.50 Hz) 30 (0.50 Hz)

Lower Limit (bpm)

Upper limit Upper Limit (bpm) (bpm) 160 (2.67 Hz) 160 (2.67 Hz) 60 (1 Hz) 60 (1 Hz)

Hear rate (HR) and breath rate (BR) limits for healthy newborn infants [17].

Hear rate (HR) and breath rate (BR) limits for healthy newborn infants [17].

The method finds continuous the HR and BR measurements by applying STFT and peak The method finds continuous the HR and BR measurements by applying STFT and peak detection detection at small time intervals that coincide with the maximum 160 bpm for HR. This is equivalent at small time intervals that coincide with the maximum 160 bpm for HR. This is equivalent to a period to a period of 0.375 s, so a FIFO stack was created with the last samples from the signal. A delay was of 0.375 s, so a FIFO stack was created with the last samples from the signal. A delay was made so that made so that there were at least three peaks for the respiratory rate in the signal. Consequently, a sixthere were at least three peaks for the respiratory rate in the signal. Consequently, a six-second delay second delay at the start was found to produce a good accuracy for the estimation of the frequency at the start was found to produce a good accuracy for the estimation of the frequency and a continuous and a continuous reading in the detection of peaks in the signals. reading in the detection of peaks in the signals. The STFT was used to obtain the spectrum in one of the 6 s sliding-windows that the average The STFT was used to obtain the spectrum in one of the 6 s sliding-windows that the average signal runs (the sliding-window moved in 0.1 s steps). The spectrum of one of these sliding-windows signal runs (the sliding-window moved in 0.1 s steps). The spectrum of one of these sliding-windows is is shown in Figure 5. The identified frequency peaks can be noted in the filtered signal. The narrowshown in Figure 5. The identified frequency peaks can be noted in the filtered signal. The narrow-band band filters are the IIR. The Butterworth filter was chosen because the bandwidths required a lower filters are the IIR. The Butterworth filter was chosen because the bandwidths required a lower order. order. This meant that fewer calculations were needed and decreased the computational cost. This meant that fewer calculations were needed and decreased the computational cost.

Figure Spectrum analysis analysis of of the the three (a) unfiltered Figure 5. 5. Spectrum three RGB-channels, RGB-channels, (a) unfiltered spectrum spectrum and and (b) (b) filtered filtered spectrum, spectrum, aa sliding-windows sliding-windows of of 66ss(the (thedashed dashedlines linesare arethe thelimits limitsfor forthe theheart heartand andbreath breathrates). rates).

Finally, Finally, the the position position of of the the highest highest peak peak was was identified identified within within the the band band of of vital vital signs signs and and the the values were printed. If the value of any of the vital signs were close to any pre-set limit, an alarm values were printed. If the value of any of the vital signs were close to any pre-set limit, an alarm was was activated. activated. 3. Results 3. Results The validation and information on the performance and reliability of the system were analyzed The validation and information on the performance and reliability of the system were analyzed with the Bland-Altman method [18]. The differences between estimates from contact and remote with the Bland-Altman method [18]. The differences between estimates from contact and remote measurements were plotted against the averages of both systems for HR and BR, Figure 6a,b and measurements were plotted against the averages of both systems for HR and BR, Figure 6a,b and Table 2, respectively. In the HR case, the measurements were recorded every second. In contrast, the BR was recorded every three seconds, thus, an over correlation was avoided by the repetition of values. Means are represented by dotted lines; 95% limits of agreement (±1.96 media (SD)) are

Sensors 2018, 18, x FOR PEER REVIEW

8 of 14

Table 2, respectively. In the HR case, the measurements were recorded every second. In contrast, the 8 of 14 BR was recorded every three seconds, thus, an over correlation was avoided by the repetition of values. Means are represented by dotted lines; 95% limits of agreement (±1.96 media (SD)) are represented by by dashed 6. 6. Specifically, thethe mean biases were −1.5−bpm with represented dashed lines lineson onthe theplots plotsininFigure Figure Specifically, mean biases were 1.5 bpm 95% 95% limits of agreement −9.7 − to9.7 5.8tobpm for the −0.6 − bpm with with 95% 95% limits of agreement −9.2 with limits of agreement 5.8 bpm forHR the and HR and 0.6 bpm limits of agreement to 10.3 bpm for the BR. The standard deviation of the residuals was 4.5 bpm for HR and 4.97 bpm for −9.2 to 10.3 bpm for the BR. The standard deviation of the residuals was 4.5 bpm for HR and 4.97 bpm BR.BR. In addition, thethe estimated andand reference data were correlated. Respectively, 360 and 120 pairs of for In addition, estimated reference data were correlated. Respectively, 360 and 120 pairs measurements from nine videos forfor HR and BRBR are plotted inin Figure of measurements from nine videos HR and are plotted Figure6c,d. 6c,d. Sensors 2018, 18, 4362

15

10

BR_Monitor - BR_Camera (bpm)

HR_Monitor - HR_Camera (bpm)

15

+1.96 SD 6.8

5 0

Mean -1.5

-5 -1.96 SD -9.7

-10

+1.96 SD 10.3

10 5

Mean 0.6

0 -5

-1.96 SD -9.2

-10 -15

-15 100

110 120 130 140 150 Geometric mean of HR_Monitor and HR_Camera (bpm)

10

160

20 30 40 50 Geometric mean of BR_Monitor and BR_Camera (bpm)

(a)

60

(b)

160

55 50

150 BR_Camera (bpm)

HR_Camera (bpm)

45 140 130 120

40 35 30 25 20

110 15 100

10 100

110

120 130 140 HR_Monitor (bpm)

150

160

10

(c)

20

30 40 BR_Monitor (bpm)

50

60

(d)

Figure Figure6.6. Bland-Altman Bland-Altman plot plot showing showing the the level level of of agreement agreement between between (a) (a) HR HR monitor monitor and and HR HR with with the the camera, camera, (b) (b) BR BRmonitor monitorand andBR BRwith withthe thecamera. camera. Dispersion Dispersion diagram diagram showing showing the the relationship relationship between betweenthe themeasurements measurementsmade madewith withthe thecontact contact method method and and the the proposed proposedmethod method(c) (c) HR HR monitor monitor and andHR HRwith withthe thecamera, camera,(d) (d)BR BRmonitor monitorand andBR BRwith withthe thecamera. camera. Table 2. Information on the differences found in the measurements by the Bland-Altman method. Table 2. Information on the differences found in the measurements by the Bland-Altman method. HR (bpm) Sample size size Sample Arithmetic mean Arithmetic mean Standard deviation Standard deviation Lower limit Upper limitlimit Lower

Upper limit

HR (bpm) 360360 −1.4667 −1.4667 4.2135 4.2135 −9.7252 6.7918 −9.7252 6.7918

BR (bpm) BR (bpm) 120 120 0.5917 0.5917 4.9747 4.9747 − 9.1588 10.3421 −9.1588 10.3421

A Pearson’s coefficient was also calculated for the heart and breath rates, with respective ranges Afor Pearson’s was also calculated for the heart and breath withBR respective ranges of 0.94 HR andcoefficient 0.86 for BR, confirming the correlation between the tworates, HR and measurements. of 0.94 for HR and 0.86 for BR, confirming the correlation between the two HR and BR measurements. There was a slight systematic difference between the two measurements. The various monitoring There wasused a slight systematic difference between error the two measurements. various monitoring apparatus in the different rooms had different ranges (some had anThe accuracy of ± 3 percent apparatus used in the different rooms had different error ranges (some had an accuracy of ±3 according to their manufacturer). In addition, atypical values caused by strong movementspercent of the according to their area, manufacturer). In addition, atypical values byinstrong movements the baby’s abdominal shadows cast by medical personnel, orcaused changes the light intensity of were baby’s abdominal area, shadows cast by medical personnel, or changes in the light intensity were visible. The lighting in some rooms, which was a combination of artificial light (fluorescent lamps) and natural light (windows and skylights in the ceiling) resulted in different shadows.

Sensors 2018, 18, x FOR PEER REVIEW

9 of 14

visible. The lighting in some rooms, which was a combination of artificial light (fluorescent lamps) and natural light (windows and skylights in the ceiling) resulted in different shadows. SensorsThe 2018,degree 18, 4362 of agreement for the measurements was calculated using the measurements 9taken of 14 for 120 s, as shown in Figure 6c,d. The figures show that the results with our method were comparable with those obtained by the NICU monitor. Results are shown in Table 3. The degree of agreement for the measurements was calculated using the measurements taken for 120 s, as shown in Figure 6c,d. The figures show that the results with our method were comparable Table 3. Concordance correlation coefficients. with those obtained by the NICU monitor. Results are shown in Table 3. Parameters HR (bpm) BR (bpm) Table coefficients. Sample size3. Concordance correlation360 120 Concordance correlation coefficient 0.9342 0.7988 Parameters

HR (bpm)

95% Confidence interval Sample size

0.9198 360 to 0.9461

Concordance Pearson correlation coefficient 0.9342 ρ (precision) 0.9410 95% Confidence interval 0.9198 to 0.9461 Pearson $ (precision) 0.9410 Bias correction factor Cb (accuracy) 0.9927 Bias correction factor Cb (accuracy) 0.9927

BR (bpm)

0.7402 to120 0.8454 0.7988 0.8585 0.7402 to 0.8454 0.8585 0.9305 0.9305

CPU usage usage was was recorded recorded when when both both scripts scripts worked workedatatthe thesame sametime. time. Scripts Scripts analyzed analyzed and and CPU processedthe thesame same information. In order to show the lightweight of our algorithm, we have processed information. In order to show the lightweight of our algorithm, we have registered registered the CPU usage history through data comparison graph. The computer was an Intel(R) the CPU usage history through data comparison graph. The computer was an Intel(R) Core(TM) Core(TM) i7-4500U CPU (1.8 GHz, 4 MB cache). i7-4500U CPU (1.8 GHz, 4 MB cache). Theresults resultsof ofthis thisexercise exerciseare areshown shownin inFigure Figure77and andTable Table4.4.In In terms terms of of CPU CPU usage, usage, our our method method The used an average of about 3.00% while the other method used an average of 12.2%, which means that used an average of about 3.00% while the other method used an average of 12.2%, which means that our method requires almost 75% less of CPU usage. our method requires almost 75% less of CPU usage. 25

Median, % CPU usage

20

15

10

5

0 Moving Average

Fast_ICA

Figure 7. Performance comparison between our method and independent component analysis (ICA). Figure 7. Performance comparison between our method and independent component analysis Table 4. CPU usage comparison graph. (ICA). Parameters

Moving Average Fast_ICA Table 4. CPU usage comparison graph. Sample size 360 120 Moving Average Fast_ICA Lowest valueParameters (%) 2.0 8.0 Highest value (%) 5.1 360 Sample size 120 21.4 MedianLowest (%) 3.0 2.0 value (%) 8.0 12.2 25th percentile (%) 3.0 11.3 Highest value (%) 5.1 21.4 75th percentile (%) 3.1 13.4

Median (%) 3.0 12.2 25th percentile (%) 3.0 11.3 In addition, to test 75th the robustness our algorithm, 3.1 we implemented 13.4 the algorithm developed by percentileof (%) Poh et al. [8], which is considered as a standard in image-based photoplethysmography to perform a comparison. This to algorithm chosenof based on their high asthe reported by the authors. In addition, test the was robustness our algorithm, weperformance implemented algorithm developed To make an objective evaluation of each approach in terms of its HR estimation accuracy, both methods by Poh et al. [8], which is considered as a standard in image-based photoplethysmography to perform were executed to analyze the same ROI from the patient video. These measurements allow the two methods to be evaluated while they were operating in a changing environment (variations in light

Sensors 2018, 18, x FOR PEER REVIEW

10 of 14

a comparison. This algorithm was chosen based on their high performance as reported by the authors. To make an objective evaluation of each approach in terms of its HR estimation accuracy, both Sensors 2018, 18, 4362 of 14 methods were executed to analyze the same ROI from the patient video. These measurements 10 allow the two methods to be evaluated while they were operating in a changing environment (variations in light type, intensity the subject is interacting with another person for more 3 min). The type, intensity when thewhen subject is interacting with another person for more 3 min). The measurement measurement evaluated is evaluated inisFigure 8. in Figure 8.

Figure 8. Performance comparison between our method and ICA. Figure 8. Performance comparison between our method and ICA.

All the statistical analysis, tables and graphical methods have performed with the MedCalc® All the statistical analysis, tables and graphical methods have performed with the MedCalc® statistical software. statistical software. 4. Discussion 4. Discussion The analysis carried out by Verkruysse [7], citing Reference [19], indicates that the G-channel The analysis carried out by Verkruysse [7],which citing corresponds Reference [19], indicates that the absorption G-channel provides the strongest plethysmographic signal, to an oxyhemoglobin provides signal, which corresponds to an oxyhemoglobin absorption peak, butthe thestrongest study alsoplethysmographic indicates that RB-channels contain plethysmographic data. The abundance of peak, but the study also indicates that RB-channels The reddishness abundance capillaries, tissue perfusion, hemoglobin oxygenationcontain and skinplethysmographic thickness can causedata. rosacea, of capillaries, perfusion, hemoglobin and skin thickness can tissue causeperfusion rosacea, or paleness, or tissue a cyanotic tone, which are more oxygenation visible in white skinned people. Good reddishness or levels paleness, or a cyanotic are more avisible in white skinned with adequate of oxygenated redtone, bloodwhich cells produces pink color in the skin. people. Good tissueWhen perfusion with adequate levels oxygenated red blood cells produces a pink color in the analyzing the images, weoffound that the skin showed a pink hue in low light andskin. was When analyzing the images, we found that the skin showed a pink hue in low light and was pale pale with more light; however, the red component always showed a high-intensity level. This is with more light; red component always showed a high-intensity level. Thisthe is important important since however, light-darkthe alternations are a powerful and important signal to achieve temporal since light-dark alternations are a in powerful and important signal to weight achieve theand temporal organization of the circadian rhythm newborns as well as promote their gain growth organization of the[20]. circadian rhythm in newborns as well as promote theirsince weight gain and growth while hospitalized Therefore, work on R-channel analysis is essential, NICU illumination is while hospitalized [20]. Therefore, work on R-channel analysis is essential, since NICU illumination not constant. is notThe constant. information received from the camera is a reflection of hemoglobin oxygenation. Figure 9 Thethat information receivedintensity from the camera is a reflection of hemoglobin oxygenation. 9 shows shows the R-channel maintains a high-level independently of anyFigure variations in that the R-channel intensityinmaintains high-level independently of in any inwhen illumination. In illumination. In addition, Figure 4cahigher intensity is observed thevariations R-channel compared addition, in Figure is observed in the R-channel to GB-channels a to GB-channels for4cahigher good intensity level of patient oxygenation (about when twice compared the average intensities offor the good level ofwith patient oxygenation (about twice the average intensities ofwith the R-channel, with respect the R-channel, respect to the other two channels). When working a dim light source, thetoidea other two channels). When working dim lightmeasurement source, the idea to start with thegiven assumption of a is to start with the assumption of awith lessa accurate onis the G-channel the higher less accurateby measurement on the G-channel given higherofabsorption by thisindicates component. In this case, absorption this component. In this case, the the absence light reflection a low level of the absence of light reflection indicates a low level of patient oxygenation. patient oxygenation.

Sensors 2018, 18, 4362 Sensors 2018, 18, x FOR PEER REVIEW

11 of 14 11 of 14

Figure 9. 9. Comparison Comparison of of intensities intensities in in RGB-channels, RGB-channels, with with different different levels levels of of illumination. illumination. Figure

Nevertheless,all allofofthis thisdepends depends type of camera used. Specifically, the sensor andfilter the Nevertheless, onon thethe type of camera used. Specifically, the sensor and the filter installed each color channel. Theofmajority cameras morefrom information from [21] the installed for eachfor color channel. The majority cameras of obtain more obtain information the G-channel G-channel [21] because the human eye is more sensitive to this color and the cameras emulate this because the human eye is more sensitive to this color and the cameras emulate this sensitivity. In any case, sensitivity. In any case, due to thethe low computational the threewith channels can be presented processedhere. with due to the low computational cost, three channels cancost, be processed the method the method presented here. here show that non-contact supervision through computer vision The results presented The results presented hereinshow that non-contact through computer vision techniques techniques may be included future modalities forsupervision neonatal monitoring in hospital NICUs. Very may be included in future modalities for neonatal monitoring in hospital NICUs. Very high-risk high-risk patients requiring close specialized supervision, such as those with heart problems, would patients close specialized supervision, such as those with heart problems, would still need still needrequiring an ECG to collect the necessary information. an ECG to collect the necessary information. Infants who need special care and are in an incubator have not achieved control of their nervous Infants who need special and are with in an small incubator haveinnot control nervous system. Their movements arecare spasmodic tremors theachieved hands or legs. of Astheir the nervous system. Their movements are spasmodic with small tremors in the hands or legs. As the nervous system matures and muscle control improves, jolts and tremors change to gentler movements of the system and muscle control and tremors change gentler movements of the legs andmatures arms. These movements willimproves, not affect jolts the measurements of vitaltosigns by plethysmography. legs and arms.infants These have movements not affect the measurements of vital signs by plethysmography. More mature greaterwill psychomotor control and may make movements that the system More mature infants have greater psychomotor control and may make movements that thetracking system will not be able to filter properly. This could be corrected by implementing a movement will not be able to filter properly. This could be corrected by implementing a movement tracking system and combining it with the method proposed here. The method to reduce the offset signal is system and combining it with the method very efficient, and the whole system can be proposed run online.here. The method to reduce the offset signal is very Studies efficient,ofand the whole system can be run online. whose cause is still unknown have observed sudden infant death syndrome (SIDS), sudden infant death syndrome whose programs cause is stilltounknown observed that Studies apnea of may precede bradycardia [22]. (SIDS), The British control have for SIDS use that apnea may precede bradycardia [22]. The British programs to control for SIDS use transcutaneous transcutaneous oxygen monitors because they can decouple the cardiorespiratory variables on the oxygen monitors because they canmakes decouple the cardiorespiratory variablesoron the monitoring monitoring equipment [23,24]. This it possible to detect signs of hypoxia obstructive apnea. equipment [23,24]. This makes it possible to detect signs of hypoxia or obstructive apnea. Therefore, Therefore, it is important to emphasize that our procedure has the advantage that the measurement it is important to referring emphasize ourarrhythmia procedure and has the advantagethe thatcardiorespiratory the measurementvariables. is made is made without tothat sinus disassociates without and referring to sinus arrhythmia and the cardiorespiratory variables. Cardiac and Cardiac respiratory dysrhythmias can disassociates be analyzed independently. respiratory dysrhythmias can be analyzed independently. Figure 10 depicts a special case that we obtained during a measurement. The vital signs monitor Figurein10Matlab depictsshows a special case that obtained during signs developed an abrupt HRwe decrease (3–4 s) thataismeasurement. interpreted byThe thevital doctor as amonitor benign developed in Matlab shows an abrupt HR decrease (3–4 s) that is interpreted by the doctor as a benign paroxysmal supraventricular tachycardia. It is a relatively frequent phenomenon in children with no paroxysmal supraventricular tachycardia. It is a relatively frequent phenomenon in children with pathological history of any kind. Signs of heart failure should be ruled out especially in younger no pathological history of have any kind. failure should[25]. be ruled out especially in younger patients and in those who moreSigns hoursofofheart clinical evolution This heart rate arrhythmia was patients and in those who have more hours of clinical evolution [25]. This heart rate arrhythmia was not repeated in the measurements. not repeated in the measurements.

Sensors 2018, 18, 4362 Sensors 2018, 18, x FOR PEER REVIEW

12 of 14 12 of 14

Figure Figure 10. 10. Signal Signal Monitor Monitor development development in in Matlab Matlab vs. vs. NICU NICU monitor. monitor.

At At present, present, the the results results obtained obtained by by PPGI PPGI are are not not completely completely categorized; categorized; the the measurements measurements still still lack reliable validation protocols. What is noteworthy is that the measurement by lack reliable validation protocols. What is noteworthy is that the measurement by PPGI PPGI in in neonates neonates differs differs from from classic classic detection detection methods, methods, because because itit will will not not cause causediscomfort, discomfort,injuries, injuries,etc. etc. The reason for choosing the abdominal area as the analysis region is that The reason for choosing the abdominal area as the analysis region is that breathing breathing in in newborn newborn infants is visible in abdominal and thoracic movements, which makes it easy to distinguish. In addition, infants is visible in abdominal and thoracic movements, which makes it easy to distinguish. In measurement also depends on the viewing angle; different test runs showed that the best placement addition, measurement also depends on the viewing angle; different test runs showed that the best for the camera was to attach the transparent wall of the pointed pointed at the baby’s placement for the camera was ittoto attach it to the transparent wallincubator, of the incubator, at the diaphragm. This resulted in an optimum viewing angle and produced a better quality measurement baby’s diaphragm. This resulted in an optimum viewing angle and produced a better quality of physiological measurement ofvariables. physiological variables. Finally, it will be be necessary to relax the cut-off frequencies to allowto forallow a larger of physiologic Finally, it will necessary to relax the cut-off frequencies forrange a larger range of measurements. Our proposed method could be useful for sick newborns. The heart and breathing physiologic measurements. Our proposed method could be useful for sick newborns. The heart and rates couldrates go outside of outside these bounds. breathing could go of these bounds. 5. Conclusions 5. Conclusions This paper describes a methodology to register cardiac and respiratory rates from video recordings This paper describes a methodology to register cardiac and respiratory rates from video of the abdominal area of newborns in NICUs. To the best of our knowledge, this is the first recordings of the abdominal area of newborns in NICUs. To the best of our knowledge, this is the demonstration of a relatively low-cost method for neonatal physiological measurement that permits first demonstration of a relatively low-cost method for neonatal physiological measurement that an independent analysis of cardiac and respiratory rates; also, this method is imperceptible to the permits an independent analysis of cardiac and respiratory rates; also, this method is imperceptible patient. The structure and algorithm of this system have not yet been tested online within a NICU to the patient. The structure and algorithm of this system have not yet been tested online within a setting. We have worked with videos because NICU access is limited so as to avoid disturbing the NICU setting. We have worked with videos because NICU access is limited so as to avoid disturbing patients. In order to evaluate the robustness and lightweight of our algorithm, we have performed the patients. In order to evaluate the robustness and lightweight of our algorithm, we have performed experiments to compare it with one of the most cited existing image photoplethysmography methods. experiments to compare it with one of the most cited existing image photoplethysmography methods. The experiment proved that our method is more robust and computationally more efficient. The experiment proved that our method is more robust and computationally more efficient. The improvement of this system can be achieved. This method would not work in the case of poor The improvement of this system can be achieved. This method would not work in the case of lighting or darkness. The system could be improved by dedicated illumination and optimal lighting poor lighting or darkness. The system could be improved by dedicated illumination and optimal conditions that would decrease or avoid shadows. In addition, working in parallel with distance lighting conditions that would decrease or avoid shadows. In addition, working in parallel with measurement systems, such as the Kinect camera, is expected to result in better concordance among distance measurement systems, such as the Kinect camera, is expected to result in better concordance measurements and decrease in artifacts effects. among measurements and decrease in artifacts effects. On the other hand, with this approach, the authors have shown how easy it is to apply the On the other hand, with this approach, the authors have shown how easy it is to apply the system system in existing incubators. We believe it has a real potential to expand and improve access to in existing incubators. We believe it has a real potential to expand and improve access to health care. health care. In addition, this methodology should respond adequately to the increasing demand for In addition, this methodology should respond adequately to the increasing demand for home home hospitalization and telemedicine in the near future. Given the low cost and wide availability hospitalization and telemedicine in the near future. Given the low cost and wide availability of of portable devices with integrated cameras, this technology is also promising for basic residential portable devices with integrated cameras, this technology is also promising for basic residential monitoring of children by their parents. Of course, measurement must be made in alternative areas monitoring of children by their parents. Of course, measurement must be made in alternative areas of analysis, such as the child’s face to extract the HR. Moreover, it could be a practical solution for of analysis, such as the child’s face to extract the HR. Moreover, it could be a practical solution for measuring BR, but it needs to incorporate an adequate motion tracking feature. Although the scope of the work described in this document is limited to recovery of the heart and respiratory rates, many

Sensors 2018, 18, 4362

13 of 14

measuring BR, but it needs to incorporate an adequate motion tracking feature. Although the scope of the work described in this document is limited to recovery of the heart and respiratory rates, many other important physiological parameters can be measured using the same video images. Therefore, this line of research should be continued with the aim of developing a real-time and multi-node system for measuring physiological variables. In summary, our contributions are: 1. 2. 3. 4. 5.

By applying numerical analysis techniques, the authors have established a filter for reducing very low frequencies. This reduces the artifacts and the offset of the plethysmographic signal. Three color channels are used to analyze the plethysmographic signal. This allows a more robust estimation of HR and BR because its measurements can be interpolated. The developed algorithm is simple and of low computational cost, which makes it adequate for implementation in portable devices with low computing power. The sinus arrhythmia is not taken for the measurements by the method itself, which is important to supply an independent analysis of the cardiac and respiratory dysrhythmias rhythms. With all the above, a robust and complete system has been developed, which is adequate for monitoring neonates in a NICU environment.

Author Contributions: Conceptualization, J.-C.C.-T., M.A. and J.M.-O.; methodology, J.-C.C.-T., M.A. and J.M.-O.; software, J.-C.C.-T.; validation J.-C.C.-T. and M.A.; formal analysis, J.-C.C.-T. and M.A.; investigation, J.-C.C.-T. and M.A.; data curation, J.-C.C.-T., M.A. and J.M.-O.; writing—original draft preparation, J.-C.C.-T.; writing—review and editing, J.-C.C.-T., M.A. and J.M.-O.; visualization, J.-C.C.-T.; supervision, M.A.; project administration, J.-C.C.-T and M.A. Funding: This research received no external funding. Acknowledgments: The first author is indebted to the Ecuadorian state for the SENESCYT scholarship, which supports his studies. His gratitude also goes to the direct support of the laboratories of Universidad Carlos III of Madrid. Furthermore, the authors would like to thank the nursing staff at the Neonatal Unit of University Hospital Puerta de Hierro-Majadahonda for their cooperation during the recording of the videos. Conflicts of Interest: The authors declare no conflict of interest.

References 1.

2.

3. 4.

5.

6.

7.

Salas, G.; Satragno, D.; Bellani, P.; Quiroga, A.; Pérez, G.; Erpen, N.; Villalba, G.; Cortea, L.; Centeno, G.; Musante, R.; et al. Consenso sobre la monitorización del recién nacido internado. Parte 1: Monitorización no invasiva del recién nacido. Arch. Argent. Pediatr. 2013, 111, 353–359. [CrossRef] [PubMed] Aminian, K.; Thouvenin, X.; Robert, P.; Seydoux, J.; Girardier, L. A piezoelectric belt for cardiac pulse and respiration measurements on small mammals. In Proceedings of the 14th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Paris, France, 29 October–1 November 1992; pp. 2633–2634. [CrossRef] Abbas, K.; Heimann, K.; Jergus, K.; Orlikowsky, T.; Leonhardt, S. Neonatal non-contact respiratory monitoring based on real-time infrared thermography. BioMed. Eng. OnLine 2011, 10. [CrossRef] [PubMed] Ming, Y.; Liu, Q.; Turner, T.; Wu, Y. Vital sign estimation from passive thermal video. In Proceedings of the 2008 IEEE Conference on Computer Vision and Pattern Recognition, Anchorage, AK, USA, 23–28 June 2008. [CrossRef] Mikhelson, I.V.; Lee, P.; Bakhtiari, S.; Elmer, T.W., II; Katsaggelos, A.K.; Sahakian, A.V. Noncontact millimeter-wave real-time detection and tracking of heart rate on an ambulatory subject. IEEE Trans. Inf. Technol. Biomed. 2012, 16, 927–934. [CrossRef] [PubMed] Gu, C.; Wang, G.; Li, Y.; Inoue, T.; Li, C. A hybrid radar-camera sensing system with phase compensation for random body movement cancellation in doppler vital sign detection. IEEE Trans. Microw. Theory Tech. 2013, 61, 4678–4688. [CrossRef] Verkruysse, W.; Svaasand, L.O.; Nelson, J.S. Remote plethysmographic imaging using ambient light. Opt. Express 2008, 16, 21434–21445. [CrossRef] [PubMed]

Sensors 2018, 18, 4362

8. 9. 10. 11.

12.

13.

14. 15. 16.

17. 18. 19. 20.

21. 22. 23. 24. 25.

14 of 14

Poh, M.; McDuff, D.; Picard, R. Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Trans. Biomed. Eng. 2011, 58, 7–11. [CrossRef] [PubMed] Wu, H.; Rubinstein, M.; Shih, E.; Guttag, J.; Durand, F.; Freeman, W. Eulerian video magnification for revealing subtle changes in the world. ACM Trans. Graph. (TOG) 2012, 31. [CrossRef] Wu, H. Eulerian Video Processing and Medical Applications. Master’s Thesis, Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, Cambridge, MA, USA, June 2012. Mestha, L.; Kyal, S.; Xu, B.; Lewis, L.E.; Kumar, V. Towards continuous monitoring of pulse rate in neonatal intensive care unit with a webcam. In Proceedings of the 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, IL, USA, 26–30 August 2014; pp. 3817–3820. [CrossRef] Scalise, L.; Bernacchia, N.; Ercoli, I.; Marchionni, P. Heart rate measurement in neonatal patients using a webcamera. In Proceedings of the 2012 IEEE International Symposium on Medical Measurements and Applications Proceedings, Budapest, Hungary, 18–19 May 2012. Aarts, L.; Jeanne, V.; Cleary, J.P.; Lieber, C.; Nelson, J.S.; Oetomo, S.B.; Verkruysse, W. Non-contact heart rate monitoring utilizing camera photoplethysmography in the neonatal intensive care unit—A pilot study. Early Hum. Dev. 2013, 89, 943–948. [CrossRef] [PubMed] Wang, W.; den Brinker, B.; Stuijk, S.; de Haan, G. Algorithmic Principles of Remote-PPG. IEEE Trans. Biomed. Eng. 2017, 64, 1479–1491. [CrossRef] Hassan, M.A.; Malik, A.S.; Fofl, D.; Saad, N.; Meriaudeau, F. Novel health monitoring method using an RGB camera. Biomed. Opt. Express 2017, 8, 4838–4854. [CrossRef] Curzi-Dascalova, L. Phase relationships between thoracic and abdominal respiratory movement during sleep in 31–38 weeks CA normal infants comparison with full-term (39–41 weeks) Newborns. Neuropediatrics 1982, 13, 15–20. [CrossRef] Hatfield, N.T. Introductory Maternity and Pediatric Nursing; Lippincott Williams & Wilkins: Albuquerque, NM, USA, 2013; p. 215. Bland, J.; Altman, D. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986, 327, 307–310. [CrossRef] Van Kampen, E.; Zijlstra, W. Determination of hemoglobin and its derivatives. Adv. Clin. Chem. 1965, 8, 141–187. [CrossRef] [PubMed] Ángeles-Castellanos, M.; Vázquez, S.; Palma, M.; Ubaldo, L.; Cervantes, G.; Rojas-Granados, A.; Escobar, C. Desarrollo de los ritmos biológicos en el recién nacido. Cátedra especial (Dr. Ignacio Chávez). Revista de la Facultad de Medicina de la UNAM 2013, 56, 26–35. Jeon, G. Cost efficient multi-sensor interpretation. Int. J. Softw. Eng. Appl. 2015, 9, 113–122. Meny, R.; Carroll, J.L.; Carbone, M.T.; Kelly, D.H. Cardiorespiratory recordings from infants dying suddenly and unexpectedly at home. Pediatrics 1994, 93, 44–49. [PubMed] Poets, C.F.; Menyg, R.G.; Chobanian, M.R.; Bonofiglo, R.E. Gasping and other cardiorespiratory patterns during sudden infant deaths. Pediatr. Res. 1999, 45, 350–354. [CrossRef] [PubMed] Sridhar, R.; Thach, B.T.; Kelly, D.H.; Henslee, J.A. Characterization of successful and failed autoresuscitation in human infants, including those dying of SIDS. Pediatr. Pulmonol. 2003, 36, 113–122. [CrossRef] [PubMed] Balaguer, M.; Garcíaa, I.; Caritg, J.; Cambra, F.J.; Prada, F.; Palomaque, A. Supraventricular tachycardia in infants and children. An. Pediatr. 2007, 67, 133–138. [CrossRef] © 2018 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/).