Neonatal Video Database and Annotations for Vital Sign Extraction and

Monitoring

Hussein Sharafeddin

, Lama Charafeddine

, Jamila Khalaf

, Ibrahim Kanj

and Fadi A Zaraket

Lebanese University, Beirut, Lebanon

American University of Beirut, Beirut, Lebanon

Keywords:

Video Database, Neonatal Monitoring Data Set, Noninvasive Monitoring.

Abstract:

Background: The end goal of this project is to detect early signs of physiological disorders in term and

preterm babies at the Neonatal Intensive Care Unit using real time camera-based non-contact vital signs mon-

itoring technology. The contact sensors technology currently in use might cause stress, pain, and damage to

the fragile skin of extremely preterm infants. Realization of the proposed camera based method might com-

plement and eventually replace current technology. Non-invasive early detection of heart rate variability might

allow earlier intervention, improve outcome, and decrease hospital stay. This study constructed a curated set

of videos annotated with accurate and reliable measurements of the monitored vital parameters such as heart

and respiratory rates so that further analysis of the curated data set lead towards the end goal. Body: The

data collection process included 56 total hours of recording in 127 videos of 27 enlisted neonates. The video

annotations include (1) vital signs acquired from bedside patient monitors at second based intervals, (2) the

neonate state of health entered and manually reviewed by a healthcare provider, (3) region of interest in video

frames for heart rate detection extracted semi-automatically, and (4) the anonymized and clipped region of

interest videos. Conclusion: The paper presents a curated data set of 127 video recordings of deidentiﬁed

neonate foreheads annotated with vital signs, and health state in XML format. The paper also presents a utility

study that shows accurate results in estimating the heart rate of term and preterm neonates. We hypothesize

that the data set we collected is beneﬁcial for improving state of the art monitoring techniques. Its timely

dissemination may help lead to techniques that detect anomalies earlier, hence, leading to earlier treatment

and improved outcome.

1 INTRODUCTION

Current medical and technological advancements are

increasingly leading to higher survival rates of term

and premature neonates. Of all live births at least

ten percent need some intervention after birth to help

them adapt to the extra-uterine life and close to one

percent will need further management because of dif-

ferent conditions and illnesses. Typically, all in-

fants admitted to the neonatal intensive care unit

(NICU) require continuous cardiopulmonary moni-

toring which is essential for early detection of signs

of illness and for tracking changes in physiologic

state (Liu et al., 2012; Lozano et al., 2012). Since the

earliest sign of physiologic disturbances in neonates

is a change in heart rate (HR) followed by the res-

piratory rate (RR), it is important to have a reliable

continuous monitoring system that permits early de-

tection and prompt intervention as early as possible to

prevent treatment delay and avoid potential morbidity

or mortality (Group, 2008).

In neonates, watchful waiting until more obvious

signs are revealed may already be too late to achieve

a complete recovery. On the other hand, over treat-

ment may be detrimental as it is the case of excessive

use of antibiotics that increases the likelihood of se-

lecting resistant bacteria hence rendering future treat-

ments more difﬁcult (Edmond and Zaidi, 2010). Cur-

rent monitoring tools used in the NICU require direct

contact with the patient which might cause harm at

times or, at least, might interfere with patient care or

parent-infant bonding. This is particularly true for ex-

tremely premature infants who have thin gelatinous

skin that is prone to sloughing in case of strong bond

or erroneous signal due to poor adherence.

For the above reasons, minimally invasive moni-

toring tools using “non-contact” electrodes have been

the subject of extensive analytical and data oriented

research (Wu et al., 2012; Poh et al., 2010; Alghoul

et al., 2017; Christinaki et al., 2014) that requires

Sharafeddin, H., Charafeddine, L., Khalaf, J., Kanj, I. and Zaraket, F.

Neonatal Video Database and Annotations for Vital Sign Extraction and Monitoring.

DOI: 10.5220/0011637800003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 767-774

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

767

collecting data sets for training, learning and evalu-

ation. We review these techniques and others in the

Related Work section 3. In this work, we present a

video database annotated with vital signs for neonate

infants. The work makes the database available to

the research community aims to advance research and

development efforts towards non-invasive monitoring

techniques for neonates.

2 CONSTRUCTION AND

CONTENT

The following describes our work to collect video

recordings of neonates while in care and curate them

with heart rate, respiratory rate and other health care

relevant annotations. The data collection process pro-

ceeded as follows.

• We designed our video capture environment with

the help of the NICU staff and director. Our target

was to allow studying the possibility of establish-

ing norms for each neonate population and to es-

tablish research data resources for public access.

• Simultaneously, we designed a method to collect

synchronized data from bedside monitors con-

nected to the neonates through electrodes. We

used the bedside monitors to obtain automatic

heart rate and respiratory rate annotations. We

bridged and synchronized the existing bedside

monitoring sensory devices with the same ma-

chine capturing the video recordings. We were in-

terested in capturing the change in heart rate and

heart rate variability (Task Force of the European

Society of Cardiology the North American Soci-

ety of Pacing Electrophysiology, 1996; Sztajzel,

2004) as we would like to study their correlation

with skin color variations.

• We constructed our annotation terms and notes

based on discussions with the NICU healthcare

team. The terms concern (i) medical state of

the neonate, and (2) technical conditions such as

reasons for obstructions to recordings or lighting

conditions.

• We drafted a consent form to collect consent from

the parents. The form describe to parents the aims

of the study and the possible risks involved in-

cluding how we planned to preserve privacy and

anonymity.

• We drafted our design as proposal to the Institu-

tional Review Board (IRB) of the American Uni-

versity of Beirut and obtained their approval to be-

gin the study. Institutional Review Board (IRB)

Table 1: Sample demographics.

Variables Value

Number of infants enrolled 27

Birth weight (g) Mean (±sd) 1302(576)

Gestational age (wk) Mean (±sd) 32.33(4.98)

Total hours recorded 55:34:28

Number of videos 127

Infants with bradycardia events 11

Infants with apnea events 2

who also approved the sharing of the anonymised

videos as an open access repository for further re-

search in this area.

• We trained our research assistants on how to ap-

proach parents for consent, and how to setup the

video recording environment to be compliant with

the objective of the studies.

• The research assistants approached the parents for

consent and once they received written informed

consent from the parents, they started the video

recordings.

• To date, we recorded, annotated and analyzed

over 55 hours of videos from 27 term and preterm

neonates.

We evaluated the utility of our collected data set

by running existing techniques (Wu et al., 2012; Al-

ghoul et al., 2017) on the data. Knowing that one

of the ﬁrst signs of physiologic disturbances in new-

born is the change in heart rate and heart rate variabil-

ity (Task Force of the European Society of Cardiology

the North American Society of Pacing Electrophysi-

ology, 1996; Sztajzel, 2004), we focused on correlat-

ing heart rate with skin color changes. The results we

obtained are promising and we discuss them further

in Section 4.

2.1 Video Database Content

Table 1 shows that the data set contains 56 hours in

127 recordings of 27 enrolled infants. Eleven infants

had bradycardia events and two infants had Apnea

events.

Table 3 shows the detailed schematic of the an-

notations. We implemented the annotations in XML

format and associated one XML ﬁle with each video.

The published video ﬁles capture the forehead region

of interest (ROI) that is instrumental in determining

the vital signs. Table 2 illustrated statistics of the

video recordings.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

768

Table 2: Recording Statistics.

Recording Time

(minutes)

Total all videos 3334

Overall Average(SD) 26.2(14.1)

Average per patient (SD) 120.5(131)

Age at ﬁrst recording (days) Min - Max 1 - 185

Recording Place

( % time)

Open incubator 61%

Closed incubator 39%

Infant status

(average % time)

Awake 31%

Deep Sleep 8%

Light Sleep 61%

Infant motion

(average % time)

At rest 8%

Minimal 63%

Agitated 25%

Agitated & Crying 5%

Detection

(average % time)

Detected 81%

Not detected 19%

Figure 1: RECORDING SETUP: Tripod with camera

mounted near incubator.

2.2 Video Acquisition

Video acquisition started November 2016 and contin-

ued until April 2018. The videos feature infants ad-

mitted to the NICU at the AUBMC after parental writ-

ten and informed consent. A total of 160 videos were

captured using the LogitecC920 high deﬁnition cam-

era with 1920 × 1080 pixel resolution at 30 frames

per second. The camera was mounted on a tripod 40

to 60cm away from the baby incubator as depicted in

Figure 1 showing the camera mounted facing the baby

incubator in the NICU. We excluded 33 out of the 160

videos due to technical issues or to failure of forehead

detection. The forehead regions extracted from the

remaining 127 videos are all admitted to the database

with their annotations.

2.3 Data De-Identiﬁcation

We performed data anonymization to clear patient

names, dates of birth and identifying facial features.

We replaced patient names and dates of birth with

a unique patient ID number and a day index while

preserving order, respectively. We only publish the

detected forehead region of the videos that is instru-

mental for vital sign extraction. Consequently, this

removes identifying facial features such as the eyes,

nose and lips. We reviewed the automatically detected

ROI videos and made sure no identifying facial fea-

Figure 2: CHALLENGES IN FACE DETECTION AND

TRACKING. FRAMES A-C: (a) Patient ID 08, sleeping at

rest. Medical equipment occluding face and challenging

face detection. (b) Baby moving and turning; challenging

continuous ROI localization. (c) face detection and fore-

head (ROI) localization (note green superimposed box). .

Figure 3: TWO PATIENT SAMPLES EACH REQUIRING

ITS SPECIALIZED INITIAL TEMPLATE FOR MATCHING.

FRAMES A-D: (a) Patient P09 in video1. (b) Patient P08

in video2. (c) Template for video1. (d) template for video2.

tures exist in them.

2.4 Forehead Detection

Typically, forehead detection proceeds by face de-

tection followed by facial feature tracking, and then

segmenting the forehead portion. However, inter-

and intra- video variability in patient appearance and

state complicated this task. For example, the pres-

ence of medical tape and nasal tubes affected the qual-

ity of features necessary for automatic face detection.

Hence, forehead detection robustness suffered and we

needed semi-automatic detection. Figure 2 shows an

example of such challenges for one baby. The baby

face is continuously partially occluded with medical

dressing. Occasionally more occlusion occurs due to

baby face or hand motion. This sometimes leads to

lost detection.

We created several image templates that cover dif-

ferent scenarios, and for each video, we manually

identiﬁed the speciﬁc template that best matches the

nose and surrounding area. Once there is a match,

the forehead region to be extracted is the rectangular

region above the template. Figure 3 shows two differ-

Neonatal Video Database and Annotations for Vital Sign Extraction and Monitoring

769

ent patient presentations together with their identiﬁed

templates and extracted regions.

The forehead tracking proceeded frame-by-frame.

Once we matched the template in an initial frame,

we automatically track the region across subsequent

frames measuring similarity mostly with no need

for human intervention. When a signiﬁcant scene

change occurs such as nurse intervention or heavy

baby movement, the tracking fails. We remedy this by

restarting the forehead detection and skipping frames

where detection fails. We manually inspected the

videos to make sure that forehead tracking failure did

not result in identifying facial features leaking into the

published videos.

2.5 Video Organization and Annotation

We organized the videos of each patient into a sepa-

rate directory. The videos of patient 01 and the as-

sociated annotation ﬁles go into directory P01. Each

patient directory contains a number of videos with an

associated annotation ﬁle in XML format.

File RecodingList includes metadata records

about all patient videos. Each record has the name

of the video ﬁle, the number of frames per second,

the day of the recording in preserved order, the length

of the video, the age of the baby at recording since

inception, and the status of the incubator.

It might also have temporal annotations indicating

when the baby was in states such as deep sleep, light

sleep, agitated, bradycardia, apnea and rest. Other

temporal annotations indicate when the region of in-

terest was detected, re-detected or lost.

Each video in a patient directory comes with an

associated xml video annotation ﬁle. The video anno-

tation ﬁle includes two sections. The ﬁrst section pro-

vides metadata about the ﬁle including its name, dura-

tion, number of frames per second, width and height

of the captured region of interest, the status of the in-

cubator, feeding method, whether an event happened

during the recording and whether a nutritive feed hap-

pened during the recording.

The second section provides a time stamp from

the start of recording, the heart and respiratory rates.

For this purpose, we conﬁgured a dual MIB/RS232

serial cord to extract realtime vital signs from the

MP40-70 Philips Intellivue monitor. The second sec-

tion also has the status of the baby, whether an action

happened, and the location of the captured ROI with

respect to the video frame. It might also have some

notes taken by the healthcare providers. Table 3 de-

scribes the entire annotation scheme with examples

for each ﬁeld.

3 RELATED WORK

Recently, several methods for extracting HR data

from video recordings of adult subjects have been

published. Eulerian Video Magniﬁcation (Wu et al.,

2012) (EVM) utilizes minute skin color variations and

low amplitude motion that are magniﬁed to reveal sig-

nals of interest which reﬂect physiologic changes at

the cellular level such as changes in skin perfusion,

temperature and heart rate. Poh et al. (Poh et al.,

2010) used Blind Source Separation (BSS) based on

independent component analysis (ICA) of the red,

green, and blue (RGB) intensity channels in facial

videos for HR measurement. Alghoul et al. (Alghoul

et al., 2017) compared between EVM and ICA and

found approaches based on ICA to deal better with

lighting-related noise; however, approaches based on

EVM performed better with motion-related noise. A

comparison of three BSS-ICA based methods is found

in (Christinaki et al., 2014) where different statistical

transformations aimed at enhancing component sepa-

rability for extracting the HR.

EVM and BSS only require video recordings from

a close distance without any close contact with the pa-

tient hence they are non-invasive. In (Wu et al., 2012),

colour changes are tracked over time, thus permit-

ting analysis of physiological state changes such as

heart rate and subsequently perfusion. Those changes

would then be correlated with particular condition or

disease states for the purpose of automatic diagnosis

and alarm issuance.

4 UTILITY AND DISCUSSION

We performed video analysis to extract vital signs us-

ing two methods reported in the literature to validate

the sanity of our data set and establish its utility. The

video analysis proceeds in two steps: face detection

and heart rate extraction.

Face Detection: Baby faces maybe often partially

covered with medical equipment and dressing while

in the incubator. That led to failure of off-the-shelf

face detection algorithms. Thus for each video, and

in the ﬁrst frame, we selected an initial region of in-

terest (ROI) containing the face. We set this as a face

template and used ROI tracking to capture it in the

rest of the video. This resulted in successful detection

throughout the recordings.

HR Extraction: We tested two methods to recover

the PPG signal from the detected faces. One is based

on analyzing the frequency content of the green chan-

nel(Wu et al., 2012), and the other is based on in-

dependent component analysis (ICA)(Alghoul et al.,

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

770

Table 3: Description of annotation scheme.

Label Description Examples

Patient ID The code given to the patient in chronological order P15

VideoName combination of patient ID, date index and video sequence num-

ber

P25-Day0505-V25-1.avi

Video porperties recording parameters

Frame rate how many image frames are captured per second (fps) of

recording

30 fps

Duration the video length in h:m:s 1:18:49

Width frame width in pixels 90

Height frame height in pixels 50

Recording Info environment-related information

Incubator incubator status being open or closed Closed

Position the baby’s position being supine or prone Supine

Non-nutritive feed whether non-nutritive feeding (paciﬁer) is being used at record-

ing time

yes

Recording Annotations Instantaneous heart rate, Status, Action, Detection, Coordi-

nates, Notes

HR Instantaneous heart rate extracted from the Intellivue Philips

Monitor every second

167

Status baby’s status in three states :

• deep sleep (absolutely no motion except for breathing re-

lated)

• light sleep (possible slow movement of hand, face, etc..)

• awake (open eyes, possible movements or interaction with

environment)

deep sleep

Action We identiﬁed four states :

• rest (usually coinciding with deep sleep)

• minimal (light or slow movement of face, limbs, etc.)

• agitated (heavy / fast movements)

• agitated with crying

agitated

Detection whether detection has occured at this second which means fore-

head availability

Detected

X,Y coordinates of the forehead in the initial frame

• x is the top left corner

• y is the bottom right corner

759,250

Notes Any activity observed during recording such as nurse care,

changes in ambient light, etc.

nurse care

2017). Both methods aim to extract rate of the car-

diovascular pulse wave which circulates throughout

the body when the heart beats.

The green channel method proceeds as follows.

a. R=Select ROI

b. Obtain raw signal s from each frame of selection

c. Extract hgi the green channel from s

d. Compute discrete signal hd

i = zero-mean, win-

dowed hgi.

e. Select the best component d

best

MAXPEAK

240

FFT(d

) by applying Fast Fourier

Transform (FFT) on on hd

i and choose the

resulting FFT component with the highest peak

within the range of 60 and 240 beats per minute

(bpm).

The ICA method follows the (a,b) steps from

above and proceeds as follows.

c. Extract hr, g, bi the red, green and blue (RGB)

channels from s

d. Compute zero-mean, unit-variance normalized

discrete signals hd

, d

i from hr, g, bi.

e. Apply ICA to compute three independent source

signals hi

, i

i = JADE(s, r, g, b). The signals

break up the raw elements hr, g, bi into three in-

dependent source signals (i

, i

) using the joint

approximate diagonalization of Eigen matrices

(JADE) algorithm.

f. Apply FFT on i

, i

, and i

g. Compute the best FFT component with the high-

est peak within the 40 and 240 bpm range.

Neonatal Video Database and Annotations for Vital Sign Extraction and Monitoring

771

Figure 4: HEART RATE EXTRACTION EXAMPLES: Four

examples of HR extraction by the Green Channel method.

The title of each subplot has the source video name and

the MAPE. The HR monitor output (considered the ground

truth) is the blue line. The extracted HR is shown in green

when the baby is at ”rest” and in red plus when the baby is

”agitated” - as deﬁned in Table 3.

Results: Both methods were tested on a sample of 10

recordings and gave acceptable heart rate estimates.

We compared the estimated heart rate with that of

the baby monitor, considered as ground truth, on a

second-by-second basis. The sample had a mean ab-

solute percentage error (MAPE) range of 3% to 8%

and standard deviation of 6% to 8% with increased

error during agitated (medium) baby motion. When-

ever our algorithm could not detect the baby face, we

considered that the signal has been lost. This usually

happens when the baby turns more than 45 degrees

away from the camera or when there is total occlu-

sion.

Figure 4 shows four sample results from the green

channel method with MAPE between 1.87% and

3.18%. The red crosses are the estimated HR at in-

stances when the baby is mildly agitated. The blue

and the green lines show the heart rates from the HR

monitor (the ground truth) and from the green channel

method respectively.

Figure 5 shows a sample comparison between the

green channel and the ICA methods. The mean ab-

solute percentage error for the green channel method

(MAPEG) is 2.1%, slightly less that that of the ICA

method (MAPEICA), 3.33%. We note that the chal-

lenge in the green channel method is its sensitivity to

the bandwidth selection to which the green signal

5 CONCLUSION

Automated early detection of physiological distur-

bances in newborns is essential for initiating therapy

as soon as possible. A contactless and non-invasive

system would be particularly helpful in vulnerable

populations such as sick neonates. This is a cross

sectional, observational study where all infants ad-

Figure 5: GREEN CHANNEL VS ICA HR EXTRACTION:

An example juxtaposing the ground truth HR (blue line)

from the attached monitor with both HR extraction meth-

ods, the Green Channel (green line) and the ICA method

(black line). The mean absolute percentage error for the

green channel method (MAPEG) is slightly less that that of

the ICA method (MAPEICA). In this recording of abour 3

minutes the baby is at ”rest” (see Table 3).

mitted to the NICU are eligible. After parent con-

sent, 127 video recordings of infants (total of 56hours

of recording) were obtained and analyzed. To estab-

lish the utility of the data set, the study performed

blind source separation (BSS) for detecting HR and

HR variability of infants admitted to the intensive

care unit. The estimated HR was compared to the

“ground truth” values of the regular monitors used in

the NICU. The testing results showed feasible heart

rate measurements on premature babies. There was

an average absolute error range of 3% to 8% and

standard deviation of 6% to 8% with increased error

during baby motion. As a conclusion, the proposed

method proves the utility of the dataset for vital sign

extraction. This technique proved beneﬁcial for pa-

tient monitoring and we hope further work presents

techniques for early detection of diseases, leading to

earlier treatment and possibly improved outcome.

6 LIST OF ABBREVIATIONS

• NICU: Neonatal intensive care unit

• HR: Heart rate

• BSS: Blind source separation

• AUBMC American University of Beirut Medical

Center

• ROI: Region of interest

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

772

7 DECLARATIONS

Ethics Approval and Consent to

Participate

All research procedures have been reviewed and ap-

proved by the institutional review board (IRB) at

the American University of Beirut Medical Center

(AUBMC). We obtained written consent statements

from the investigators and the participants. The in-

vestigator and participant consent statements follow.

Investigator’s Statement

I have reviewed and explained in detail, the informed

consent document for this research study with [name

of participant, legal representative or guardian/ parent

if the participant is a minor or is unable to sign] the

nature and purpose of the study and its risks and ben-

eﬁts. I have answered to all his/her questions clearly

to the best of my ability. I will inform the participant

in case of any changes of this study or its negative

impacts or beneﬁts in the event of their occurrence.

Consent to Participate

I have read this consent form and understood its con-

tent. All my questions have been answered. Accord-

ingly, I agree, with my own free will, to be part of

this research study, and I know that I can contact Dr.

Lama Charafeddine at 01-350000 Ext: 5874 or any

of her designee/assistant involved in the study in case

of any questions. If I feel that my questions have not

been answered or need further clariﬁcation I can con-

tact one of the members of the Institutional Review

Board for human rights or its chair Dr. Fouad Zyadeh

at 01-350000 Ext: 5445.

I understand that I am free to withdraw my con-

sent from this study and discontinue my participation

at any time, even after signing this form without prej-

udice to the medical care provided to me. I know that

I will receive a copy of this signed informed consent.

• I agree to participate in this study and authorize

the investigator and her designee the access to my

child’s AUBMC medical records

• I authorize the investigator and her designee to

contact me at a later stage for future follow-up if

needed.

Consent for Publication

We have obtained consent to publish de-identiﬁed and

anonymized versions of the data. We are only sharing

publicly areas of foreheads which do not include any

identiﬁcation features.

Availability of Data and Material

The dataset supporting the conclusions of this arti-

cle is available in the American University of Beirut

research repository via the link: http://research-fadi.

aub.edu.lb/neonates/home.html.

Funding

Authors (LC, HS, and FZ) used Lebanese National

Council for Scientiﬁc Research Awards and Lebanese

Univeristy Research Funding Awards to hire research

assistants to complete the work. The funders did not

intervene in the process of the research work.

Authors’ Contributions

Authors JK and IK helped deﬁne and implement the

curation process. Authors HS and FZ helped deﬁne

and implement the curation process and the signal

processing process. Author LC deﬁned and imple-

mented the clinical aspects and processes. All authors

contributed equally to the ideation and organization of

the manuscript, revised and approved the ﬁnal version

of the manuscript.

ACKNOWLEDGEMENTS

We would like to thank the baby parents for their con-

sents in recording their newborns and recording them

on video, the engineers at the AUBMC for their co-

operation and help in setting up the medical monitor

equipment for data extraction and the nurses for their

patience and cooperation in the data collection pro-

cess.

REFERENCES

Alghoul, K., Alharthi, S., Al Osman, H., and El Sad-

dik, A. (2017). Heart rate variability extraction from

videos signals: Ica vs. evm comparison. IEEE Access,

5:4711–4719.

Christinaki, E., Giannakakis, G., Chiarugi, F., Pediaditis,

M., Iatraki, G., Manousos, D., Marias, K., and Tsik-

nakis, M. (2014). Comparison of blind source sepa-

ration algorithms for optical heart rate monitoring. In

2014 4th International Conference on Wireless Mo-

bile Communication and Healthcare - Transforming

Healthcare Through Innovations in Mobile and Wire-

less Technologies (MOBIHEALTH), pages 339–342.

Neonatal Video Database and Annotations for Vital Sign Extraction and Monitoring

773

Edmond, K. and Zaidi, A. (2010). New approaches to

preventing, diagnosing, and treating neonatal sepsis.

PLOS Medicine, 7(3):1–8.

Group, Y. I. C. S. S. (2008). Clinical signs that predict se-

vere illness in children under age 2 months: a multi-

centre study. The Lancet, 371(9607):135–142.

Liu, L., Johnson, H. L., et al. (2012). Global, regional, and

national causes of child mortality: an updated system-

atic analysis for 2010 with time trends since 2000. The

Lancet, 379(9832):2151–2161.

Lozano, R., Naghavi, M., et al. (2012). Global and re-

gional mortality from 235 causes of death for 20 age

groups in 1990 and 2010: a systematic analysis for

the global burden of disease study 2010. The Lancet,

380(9859):2095–2128.

Poh, M.-Z., McDuff, D. J., and Picard, R. W. (2010). Non-

contact, automated cardiac pulse measurements using

video imaging and blind source separation. Opt. Ex-

press, 18(10):10762–10774.

Sztajzel, J. (2004). Heart rate variability: A noninva-

sive electrocardiographic method to measure the auto-

nomic nervous system. Swiss Medical Weekly, pages

514–522.

Task Force of the European Society of Cardiology the

North American Society of Pacing Electrophysiology

(1996). Heart rate variability. task force of the euro-

pean society of cardiology the north american society

of pacing electrophysiology. Circulation, 93(5):1043–

1065.

Wu, H.-Y., Rubinstein, M., Shih, E., Guttag, J., Durand, F.,

and Freeman, W. (2012). Eulerian video magniﬁca-

tion for revealing subtle changes in the world. ACM

Trans. Graph., 31(4):65:1–65:8.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

774