LifeSeniorProﬁle: A Multisensor Dataset for Elderly Real-time Activity

Track

Maicon Diogo Much

, Julio Alexander Sieg

, Ayalon Angelo de Moraes Filho

Vanessa de Moura Bartoski

, Guilherme Schreiber

and C

esar Marcon

School of Technology, Pontiﬁcal Catholic University of Rio Grande do Sul, Porto Alegre, Brazil

Keywords:

Multisensor Dataset, Real-time Activity Tracking, Elderly Motion Data, Elderly Physiological Data.

Abstract:

Real-time tracking and detection of risky situations in the elderly, such as falls and sudden changes in vital

signs, requires reliable, continuous, and automated monitoring systems based on relevant information. Wire-

less biosensors provide a great opportunity to remotely detect and monitor hazardous situations, allowing for

a fast response in an emergency. Motion data is widely used to track daily activities. Physiological data can

also be used for this exact purpose. However, there is yet to be a database available in the ﬁeld of research in

which the patient’s physiological and movement information were collected simultaneously, considering daily

activities and simulation of falls. This work presents a multisensor dataset for developing real-time tracking

systems for the daily activities of older people. The data sensed refer to movement, using a triaxial accelerom-

eter, and physiology, considering blood volume pulse, electrodermal activity, heart rate, inter-beat interval,

and skin temperature. We collected these data from ten volunteers while performing 36 daily activities in a

simulated environment.

1 INTRODUCTION

Monitoring, detecting, and classifying activities of

older people through non-invasive wearable systems

is an open research area due to the similarity be-

tween movement data sensed in daily activities and

risky situations, such as falls. Recently, many ma-

chine learning algorithms, especially those that ex-

plore deep layers of neural networks (Mauldin et al.,

2018) (Li et al., 2019) (Santos et al., 2019), have

tried to search for hidden details of motion data of

accelerometers, gyroscopes, barometers, and magne-

tometers to identify features that classical algorithms

cannot identify. Despite the promising results, even

these high-performance algorithms hardly reveal a

real risk, achieving a false positive rate close to zero.

Accelerometers are the most common motion sen-

sors used to identify human activities, providing real-

https://orcid.org/0000-0002-1760-907X

https://orcid.org/0000-0003-3966-9855

https://orcid.org/0000-0001-7044-1504

https://orcid.org/0000-0001-6365-9273

https://orcid.org/0000-0002-1037-4853

https://orcid.org/0000-0002-7811-7896

time relative acceleration in a given direction, which

varies according to available freedom degrees. In the

rest state, the accelerometers share quasi-static val-

ues, and at excessive motion, they provide an accel-

eration peak. This behavior is adequate to develop a

low-precision algorithm based on peak detection sig-

nals to detect fall situations and identify elderly risk

situations. On the one hand, this algorithm is ineffec-

tive since many risk-free activities also have the same

data signature. On the other hand, the foundation is

the same for many fall detection systems developed in

various research. Over the years, many studies have

tried to improve these results by adding different mo-

tion sensors, but this problem is still an open research

area.

Our primary motivation for this work is based on

the direct correlation between accidental falls and vi-

tal signs (Naschitz and Rosner, 2007); this correla-

tion is not explored in most of the research data avail-

able to develop systems that explore risk situations

for older people. Integrating vital signs with motion

signals offers a considerable advantage for identify-

ing, detecting, and classifying daily activities and fall

risks. However, this integration is rarely addressed

in the available works (Oliver and Healy, 2009) (Vas-

Much, M., Sieg, J., Moraes Filho, A., Bartoski, V., Schreiber, G. and Marcon, C.

LifeSeniorProﬁle: A Multisensor Dataset for Elderly Real-time Activity Track.

DOI: 10.5220/0011730000003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 453-460

ISBN: 978-989-758-631-6; ISSN: 2184-4305

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

453

sallo et al., 2009).

This work provides a dataset of multisensory in-

formation, which includes, in addition to traditional

motion sensors, some physiological sensors collected

during activities of daily living and risky situations.

We designed this database to analyze changes in vital

signs caused by daily activities and falls. These data

enable us to explore the fusion of physiological and

motion data and classify all sensors based on their

importance in detecting daily activities and falls, al-

lowing us to give weight to each element sensed.

2 RELATED WORK

This section describes articles encompassing datasets

of wearable sensors employed to predict falls and

daily activities, and Table 1 emphasizes the main

points of the related articles concerning the work pre-

sented here. Table 1 lets us notice the main similari-

ties and differences between this work and the others

for comparison and elaboration of some conclusions.

The ﬁrst comparison in the table can be made by

evaluating the position of the sensors on the partici-

pants’ bodies. While some related works seek to use

several sensors in different body positions, like (Casi-

lari et al., 2017) and (Saleh et al., 2021a), the LifeSe-

nior dataset seeks to use a single device located on the

participants’ wrists, which is the same strategy used

by (Garcia-Ceja et al., 2021) and (Bruno et al., 2014)

datasets. This strategy is based on the idea that the

wrist is a well-accepted place to create products that

can be used daily.

Another interesting point is related to the activities

identiﬁed in each dataset. Related works that bring

daily activities simulate them starting with a lack of

movement and then the activity to be considered. Dif-

ferently from that, the LifeSenior dataset brings the

idea of executing a speciﬁc activity/risk situation to

perform some daily activity/fall. The closest to this

approach is found in the (Saleh et al., 2021a) dataset,

where the fall activities are related to a previous daily

activity, which leads to the fall.

The LifeSenior dataset contains data for 36 daily

activities, including fall situations, which is the higher

number of activity and fall data in a single dataset

compared to related work. The closest to this quantity

is the sheer number of fall types found in the FallAllD

(Saleh et al., 2021a) dataset, the diversity of daily ac-

tivities brought in HMP (Bruno et al., 2014), and even

the fusion of falls and activities brought in UMAFall

dataset (Casilari et al., 2017).

Regarding the sensors used, most of the related

works use the accelerometer as the only or one of the

motion sensors, which is the same approach as the

LifeSenior dataset. Despite this, while (Saleh et al.,

2021a) and (Casilari et al., 2017) rely on commonly

used sensors such as the gyroscope and magnetome-

ter and the smartphone that appears in some cases,

the dataset proposed in this work is based on a few

different sensors: PPG, EDA and Skin Temperature.

These sensors are located in the same body position,

the wrist, and can bring information that was previ-

ously little considered or unknown in other works.

In addition, the authors believe that using data fusion

from these sensors can bring new and unexpected re-

sults in future work.

3 SENSORS

We developed this dataset to record natural and pro-

voked human actions, such as a fall, using several sen-

sors that enable us to gather information from mul-

tiple points of view about each human action. We

planned to include a triaxial accelerometer to detect

human movement and insert sensors of temperature,

electrodermal activity (EDA), and photoplethysmog-

raphy (PPG) to extract vital signs.

3.1 Motion Sensors

All datasets evaluated presented data on daily activ-

ities and falls collected through motion sensors, pri-

marily accelerometers. The main reason to focus on

motion sensors is that most activities usually vary sig-

niﬁcantly in the sensing signal. For example, a fall is

registered by a sequence of peaks and valleys in an

accelerometer signal.

An accelerometer is an electromechanical device

that measures the change in velocity over time (Pic-

cinno et al., 2019). Acceleration measurements can

be static, such as the force of gravity, or dynamic,

caused by motion. In general, accelerometers trans-

late an external acceleration signal into a displace-

ment of their moving mass, called inertial mass. An

accelerometer reports a drop as an abrupt change

in values, represented by peaks and valleys (Bourke

et al., 2007); a graph generated by an accelerometer

during a fall shows the pre-fall, critical, and post-fall

phases. Within the critical phase, it is still possible to

identify the free fall, impact, and adjustment (Saleh

et al., 2021b).

On the one hand, the presence of a high peak of

acceleration followed by inactivity is a solid indicator

to detect falls; on the other hand, more information

is needed to avoid false alarms. For example, lying

on a bed satisfy this accelerator sequence, making a

HEALTHINF 2023 - 16th International Conference on Health Informatics

454

Table 1: Comparisons between the related articles and our work.

Dataset Participants Sensors Sensors body position Year Activities

FallAllD 15 1, 2, 3 Neck, Wrist, Waist 2021 a

HTAD 3 1, 5 Wrist 2020 b, c, d, e, f, g, h

ShimFall&ADL 35 1 Chest 2020 a, i, j, k, l, m

HMP 16 1 Wrist 2014 e, l, m, n, r, s, t, u, v, w, x

UMAFall 17 1, 2, 3, 4 Ankle, Waist, Wrist, Chest 2017 a, k, l, m, n, p, q, r

LifeSenior 10 1, 5, 6, 7 Wrist 2022 a, l, m, o, r, y, z

Sensors: 1 - Accelerometer, 2 - Gyroscope, 3 - Magnetometer, 4 - Smartphone, 5 - PPG, 6 - EDA, 7 - Skin temperature.

Activities: a - Fall, b - Mop ﬂoor, c - Sweep the ﬂoor, d - Type on a computer keyboard, e - Brush teeth, f - Wash hands,

g - Eat chips, h - Watch TV, i - Jumping, j - Lying Down, k - Bending/Picking up, l - Sitting/Standing to/from a chair, m

- Walking, n - Climbing Stairs Down/Up, o - Loss of Motor Balance, p - Hopping, q - Light Jogging, r - Lying down (and

getting up) on (from) a bed, s - Comb hair, t - Drink from a glass, u - Eat with fork and knife, v - Eat with a spoon, w - Pour

water in a glass, x - Use telephone, y - Standing, z - Crouching and coming back

fall detection system with exclusive use of this ap-

proach somewhat limited. Therefore, we structured

this dataset by collecting vital signs to analyze how

this information can compose with motion informa-

tion to decide about a daily activity or a fall.

3.2 Vital Signs Sensors

The temperature sensor enables us to identify body

changes that may indicate simple disorders, such as

a fever crisis, or more severe disorders that can com-

promise the functioning of vital organs. If the body

temperature exceeds 42ºC, the individual is at risk of

dying, as well as when the measured body tempera-

ture drops below 30ºC (Holtzclaw, 1993). The tem-

perature sensor, combined with a motion sensor, can

help to identify the cause of a fall or fainting.

Electrodermal activity (EDA) is the term used to

deﬁne autonomous changes in the electrical proper-

ties of the skin. For this reason, the EDA sensor works

by identifying electrical changes on the surface of the

skin, allowing the monitoring of episodes of stress,

anxiety, and the neurological state of an individual

since it is much more susceptible to human emotions

than just the analysis of heart rate (Blain et al., 2008).

The photoplethysmography (PPG) sensor uses an

optical technique to identify blood volume changes

in the microvascular tissue bed under the skin due to

the pulsatile nature of the circulatory system (Kamal

et al., 1989). The sensor system consists of a light

source and a detector with a red and infrared light-

emitting diode (LED). The PPG sensor monitors light

intensity changes through reﬂection or tissue trans-

mission (Tamura et al., 2014). Studies report that

through PPG, it is possible to non-invasively estimate

signals such as Electrocardiogram (ECG), heart pulse

rate, saturation (Kamal et al., 1989), respiratory rate

(Jarchi et al., 2018), and blood pressure (Kurylyak

et al., 2013).

4 DATASET

LifeSeniorProﬁle was developed to contribute to

world research in risk situation detection in older

people based on wearable device solutions. Some

datasets are available for research in this area, but

almost all are based only on motion sensors. We

propose a novel approach bringing up important vi-

tal signs correlated with motion data during simulated

daily living activities and risk situations. We use the

E4 wristband from Empatica (McCarthy et al., 2016),

a medical-grade and trustworthy device that provides

accurate real-time signals. Figure 1 displays the de-

vice with its sensors.

Figure 1: Front and back view of Empatica E4 device

biosensor (McCarthy et al., 2016).

In addition to the pioneering nature of the pro-

posed dataset in correlating vital signs with motion

sensors, the proposal to use a device that is already

widely clinically validated for use in research and is

also certiﬁed by several regulatory agencies in the

health area makes LifeSeniorProﬁle dataset unique.

4.1 Data Collection Process

Once the ethics committee of Group Conceic¸

ao Hos-

pital approved our experimental protocol (number

5,431,965), we started to recruit volunteers that ﬁt

LifeSeniorProﬁle: A Multisensor Dataset for Elderly Real-time Activity Track

455

the eligibility criteria. All the volunteers were re-

cruited around the research laboratory; the recruiter

explained all the study details to the volunteers before

accepting them to participate.

The volunteers placed a bracelet (collection de-

vice) on the right pulse to collect and identify the vi-

tal and motion signs. The movement characterization

and performance were accompanied by a physical

therapist who recorded the movement type for each

volunteer in each situation. These records, from the

physical therapist and the collection device, known as

the gold standard, are collected to be confronted later

and related to the clinical variables of each research

subject.

All movements were performed in an area covered

by 50cm x 50cm and 20mm thick EVA plates to ab-

sorb impacts and avoid discomfort. For the fall sim-

ulations, a mattress with a fall area of 3m x 2m and

30cm thick was additionally positioned.

The movement simulation step involves executing

a previous dynamic gait, successively followed by the

unbalance or fall movement. The following dynamic

gaits were simulated prior to each speciﬁc movement:

• Walking on a level surface at normal speed for 6

meters;

• Gait with vertical head movements; Up - Down -

Forward;

• Walk around the obstacle in the format of an eight;

The study volunteers performed dynamic gait

prior to all movement simulations to be collected and

evaluated, including the performance of activities of

daily living (ADL), simulations of different causes of

imbalance, and simulations of different falls (Figure

2). These movements, which are described below,

were used in previous studies of drop detector wear-

ables to collect fall, non-fall, and near-fall signals

(Bourke and Lyons, 2008) (Aziz and Robinovitch,

2011):

Activity Daily Living

• Stopped - without movement;

• Walking straight for 10 meters;

• Standing up and sitting in the chair - Perform

the entire movement of sitting on a seat approx-

imately 50cm high, wait 30 seconds and get up

from the seat and stand up straight again;

• Squatting and coming back - Starting from the up-

right posture, the volunteer performs the squatting

movement until he/she rests on the heels, waits 30

seconds, and returns to the standing position;

• Lying down and getting up - the volunteer per-

forms the complete lying down movement with

Figure 2: Simulations of different causes of imbalance and

simulations of different falls.

the face-up on a bed approximately 40cm high,

waits 30 seconds, and gets out of bed, getting erect

again.

Loss of Balance

• Sitting and Standing - the volunteer completes

the entire movement of sitting on a seat approx-

imately 50cm high, waits 30 seconds, and when

standing up, simulates the loss of balance forward

and ﬁnishes the erect movement;

• Standing and sitting - the volunteer performs

the movement of sitting on a seat approximately

50cm high, simulating loss of balance due to in-

correct weight transfer to the seat and ending the

movement erect;

• Changing direction - Being initially stopped, per-

form a 180° change of direction and, in the end,

simulate the loss of balance to the side and ﬁnish

the erect movement;

• Reaching Object - Being initially stationary, the

volunteer moves to look for an object on the

HEALTHINF 2023 - 16th International Conference on Health Informatics

456

ground and, in the end, simulates the loss of for-

ward balance and ﬁnishes the upright movement.

Fall

• Obstacle - Initially walking, the volunteer sim-

ulates limb collision with an obstacle (10 cm x

15cm bulkhead) and then simulates the fall move-

ment on the mattress;

• Cadence continuity - Initially walking, the volun-

teer simulates the forward slip and then simulates

the movement of falling on the mattress;

• Syncope - Initially stopped, relax lower limbs to

simulate a fall on a mattress.

These movements were captured by the device

instrumentation, generating reference signals associ-

ated with these movement patterns, which allow train-

ing and validate different models of algorithms capa-

ble of characterizing movements and detecting falls.

4.2 Eligibility Criteria (Ethics

Committee)

Volunteers were selected following ethical principles

and with comprehensive selection criteria. Volunteers

are adults, without gender restriction, over 18 years

of age, and in total health; that is, volunteers who

presented at least one of the following characteristics

were considered unﬁt for the study: (i) neurological

diseases; (ii) gait disorders and musculoskeletal dis-

eases; (iii) uncorrected visual impairment; (iv) inabil-

ity to maintain orthostatism; (v) need and assistance

to get around; (vi) functional dependency; and (vii)

recent orthopedic trauma.

4.3 Characteristics of the Participants

The dataset consists of data from 10 volunteers with

an average age of 36 years; all of them meet the cri-

teria established in the 4.2 section, 60% male with

average weight and height of 85kg and 1.74m, re-

spectively. Regarding female volunteers, the average

weight and height were 72kg and 1.65m. Further in-

formation about the volunteers was preserved for con-

ﬁdentiality reasons provided in the consent form for

participation in the research.

4.4 Experimental Protocol

We divided the experimental protocol into three

groups that simulate Daily Living Activities (DLA),

Loss of Motor Balance (LMB), and Effective Fall

(EF). In order to standardize the beginning of data col-

lection, we created three dynamic gaits: (i) on a ﬂat

surface at normal speed for 6 meters, (ii) with vertical

head movements, and (iii) walking around an obstacle

in the shape of an 8. After performing these marches

for 30s, the collection of movements in the database

was started, as detailed in Table 2.

Altogether, we obtained 360 records identiﬁed

with the number of each individual, followed by the

initial simulation applied in the test, the type of gait,

and the effective simulation, according to the last col-

umn of Table 2.

4.4.1 Effect of Data Merging for Some Activities

The data collected by the E4 wristband is transmit-

ted to the Empatica platform and exported in CSV

(Comma-separated values) ﬁles to be manipulated in

spreadsheet and graphics software.

The sensors record data at different rates:

• Accelerometer - 32 points per second;

• PPG sensor - 64 points per second for blood vol-

ume and 1 point per second for heart rate;

• EDA sensor - 4 points per second;

• Temperature sensor - 4 points per second.

To analyze the inter-sensor correlation, we upsam-

pled the sensors with lower frequencies, using the lin-

ear interpolation method to normalize all data to 64Hz

(the highest frequency found among the sensors).

We used the accelerometer as a base sensor to

identify a volunteer’s movement, daily activities, loss

of balance, and fall. We looked for variations that

occurred in the other sensors at the points where the

events were identiﬁed on the accelerometer graph.

We noticed that the Blood Volume Parameter (BVP)

showed a strong correlation with the events recorded

by the accelerometer, as can be seen in Figure 3,

which shows the daily activity record, with gait and

vertical movements with the head, and getting up and

sitting down from the chair.

In this same record, when analyzing the normal-

ized data of heart rate, temperature, and electroder-

mal activity, large variations are not highlighted, ei-

ther null or very small, as shown in Figure 4. How-

ever, it is not possible to say that these signals are not

inﬂuenced by the movement of the volunteer, as it is

necessary to consider that they have a lower frequency

of records. What can be seen in Figure 4 is that the

Heart Rate (HR) has its records interrupted before the

end of the collection at the 6000 points of the graph.

These same characteristics, peaks in the ac-

celerometer and the synchronized blood volume sig-

nals, could be observed in the records of loss of bal-

ance activities and falls.

LifeSeniorProﬁle: A Multisensor Dataset for Elderly Real-time Activity Track

457

Table 2: Details of the movements present in the database.

Simulation

Dynamic

gait (30s)

Movement

simulation

Time

File name

DLA

A, D e G

1 10s V1 DLA A 1, V1 DLA D 1, V1 DLA G 1

2 The necessary V1 DLA A 2, V1 DLA D 2, V1 DLA G 2

3 2 repetitions V1 DLA A 3, V1 DLA D 3, V1 DLA G 3

4 2 repetitions V1 DLA A 4, V1 DLA D 4, V1 DLA G 4

5 2 repetitions V1 DLA A 5, V1 DLA D 5, V1 DLA G 5

LMB

The necessary

V1 LMB A 6, V1 LMB D 6, V1 LMB G 6

7 V1 LMB A 7, V1 LMB D 7, V1 DLA G 7

8 V1 DLA A 8, V1 DLA D 8, V1 DLA G 8

9 V1 LMB A 9, V1 LMB D 9, V1 LMB G 9

10 V1 EF A 10, V1 EF D 10, V1 EF G 10

11 V1 EF A 11, V1 EF D 11, V1 EF G 11

12 V1 EF A 12, V1 EF D 12, V1 EF G 12

Simulation: DLA - Daily Living Activities, LMB - Loss of Motor Balance, EF - Effective Fall

Dynamic Walking: A - Walking on a ﬂat surface at normal speed for 6 meters, D - Walking with vertical head movements:

Up – Down – Forward, G – Walking around an obstacle in ﬁgure eight.

Motion simulation: 1 – Standing, 2 – Walking, 3 – Standing up and sitting in the chair, 4 – Crouching and coming back, 5

– Lying down and standing up, 6 – Sitting down and standing up, 7 – From standing and sitting, 8 – Changing direction, 9 –

Reaching object, 10 – Obstacle, 11 – Cadence continuity, 12 – Syncope.

Figure 3: Accelerometer and PPG sensor signals.

The data collection of the Empatica platform pro-

vides a graphical view of the data with native lin-

ear interpolation for data normalization. This facility

enables us to verify the inﬂuence of the volunteer’s

movement also on the EDA and temperature sensors,

as shown in Figure 5 - recording a gait simulation

around obstacles in eight, followed by a fall simulat-

ing a syncope.

Figure 5 shows the last 10 seconds of the collec-

tion, where the accelerometer data are uniﬁed, with

no signiﬁcant variation. However, the EDA sensor

and BVP variations are quite different. Nevertheless,

the temperature undergoes a small drop that slightly

Figure 4: Signals from heart rate (PPG) and EDA and tem-

perature sensors.

changes the graphic.

Machine learning algorithms see data raw without

the need for graphics. Thus, these slight variations in

the sensors and the accelerometer serve as points of

differentiation that can make a big difference in de-

ciding whether an event can be classiﬁed as a fall or

just a movement of an individual purposefully lower-

ing himself, for example.

HEALTHINF 2023 - 16th International Conference on Health Informatics

458

Figure 5: Signals of accelerometer, BVP, EDA, and temper-

ature sensors.

4.5 Dataset Characteristics

Each LifeSeniorProﬁle dataset ﬁle encompasses a dy-

namic gait simulation followed by a motion simu-

lation that can be a daily activity, loss of balance,

or a fall. Preceding the movement simulation with

a dynamic gait makes the collected data closer to a

real event, where some previous situation of normal-

ity will always precede a speciﬁc movement situation.

This procedure is another differential of the proposed

model concerning those currently available.

We separated data according to the sensor used

to perform the measurement; therefore, all participant

ﬁles were uniﬁed and separated by the sensors in the

wearable. To avoid losing the references present in

each simulation and to allow the individual analysis

of the curves, a “Label” ﬁle was created containing

the information of the last column of the previous ta-

ble. This process resulted in the following ﬁles:

• ACC-X.csv - Containing the accelerometer data

referring to the X axis;

• ACC-Y.csv - Containing the accelerometer data

referring to the Y axis;

• ACC-Z.csv - Containing the accelerometer data

referring to the Z axis;

• BVP.csv - Containing blood volume pulse data;

• EDA.csv - Containing electrodermal activity data;

• HR.csv - Containing heart rate values;

• TEMP.csv - Containing temperature data;

• LABEL.csv - Containing the characteristics of the

experimental protocol.

For convenience, the entire dataset was split into

the train and test sets. Seven subjects were included in

the train set and three in the test set. This procedure

helps develop algorithms that use the training set to

learn the characteristics, and the developed model is

not seen as data (test set).

4.6 Dataset Limitations

This dataset has some limitations that need to be taken

into consideration:

• Limited number of volunteers - as the number of

volunteers performing the simulated movements

is not higher than other datasets, the researcher

needs to consider speciﬁc algorithms; some in-

sights can be the product of a poor number of col-

lections and not from the algorithm performance.

• Age of volunteers - the average age of volun-

teers is lower than that of the elderly, generating a

loss of speciﬁc characteristics found only in older

adults. This choice prevents older adults from get-

ting injured, even with the falls being controlled

and assisted by a medical team.

5 CONCLUSION

Different methods of detection and monitoring of

daily activities performed by the human body have

already been developed and explored by numerous

scientiﬁc articles in the literature. Each method has

its peculiarities, but there is a consensus that motion

sensors are the ones that present the most acceptable

result within a feasible scenario in wearable devices.

Despite this good performance, these types of sensors

also present a high rate of false positives, like detect-

ing an activity that was not executed. This fact occurs

due to the motion signal characteristics that can be

easily confused with normal daily movements.

This article displays that the association of physi-

ological sensors to the usual movement provides rele-

vant information for reducing the number of false pos-

itive detections, given the evident correlation between

the different behavior of the vital signs in a real de-

tectable activity and a mistakenly detected one. This

dataset enables the researchers to explore in detail the

behavior associated with falls and daily living activi-

ties, thanks to a long time of data collection for each

simulated movement and the context available; there-

fore, providing a way to consider not only the exact

movement but also what occurred before and after

it. Algorithms that consider movement context, like

a fall, can ﬁnd important information about the vital

sign status before the fall, what the user was doing be-

fore the fall, and also how is the motion characteristics

after the fall, enabling us to detect if it was senseless

or moving.

Our proposed dataset is an essential new tool to

improve the results of activity detector algorithms

based on non-invasive wearable sensors. It is now

LifeSeniorProﬁle: A Multisensor Dataset for Elderly Real-time Activity Track

459

available for new researchers and can be downloaded

at https://github.com/lifeseniorproject/proﬁle.

For future work, we plan to collect more data from

users and include more older people, which was not

included in this article due to the difﬁculty of recruit-

ing them. Increasing this dataset enables us to use

deep learning algorithms.

ACKNOWLEDGMENT

This study was ﬁnanced in part by the Coordina-

tion for the Improvement of Higher Education Per-

sonnel - Brazil (CAPES) - Finance Code 001, Na-

tional Council for Scientiﬁc and Technological Devel-

opment (CNPq) and Financier of Studies and Projects

(FINEP).

REFERENCES

Aziz, O. and Robinovitch, S. N. (2011). An analy-

sis of the accuracy of wearable sensors for classify-

ing the causes of falls in humans. IEEE transac-

tions on neural systems and rehabilitation engineer-

ing, 19(6):670–676.

Blain, S., Mihailidis, A., and Chau, T. (2008). Assessing

the potential of electrodermal activity as an alterna-

tive access pathway. Medical Engineering & Physics,

30(4):498–505.

Bourke, A., O’Brien, J., and Lyons, G. (2007). Evaluation

of a threshold-based tri-axial accelerometer fall detec-

tion algorithm. Gait & Posture, 26(2):194–199.

Bourke, A. K. and Lyons, G. M. (2008). A threshold-based

fall-detection algorithm using a bi-axial gyroscope

sensor. Medical engineering & physics, 30(1):84–90.

Bruno, B., Mastrogiovanni, F., and Sgorbissa, A. (2014). A

public domain dataset for adl recognition using wrist-

placed accelerometers. In The 23rd IEEE Interna-

tional Symposium on Robot and Human Interactive

Communication, pages 738–743.

Casilari, E., Santoyo-Ram

on, J. A., and Cano-Garc

ıa, J. M.

(2017). Umafall: A multisensor dataset for the re-

search on automatic fall detection. Procedia Com-

puter Science, 110:32–39. 14th International Con-

ference on Mobile Systems and Pervasive Computing

(MobiSPC 2017) / 12th International Conference on

Future Networks and Communications (FNC 2017) /

Afﬁliated Workshops.

Garcia-Ceja, E., Thambawita, V., Hicks, S. A., Jha, D.,

Jakobsen, P., Hammer, H. L., Halvorsen, P., and

Riegler, M. A. (2021). Htad: A home-tasks activities

dataset with wrist-accelerometer and audio features.

In Loko

c, J., Skopal, T., Schoeffmann, K., Mezaris,

V., Li, X., Vrochidis, S., and Patras, I., editors, Multi-

Media Modeling, pages 196–205, Cham. Springer In-

ternational Publishing.

Holtzclaw, B. J. (1993). Monitoring body temperature. Ad-

vances in cardiac care limiting care, 4(1):44–55.

Jarchi, D., Salvi, D., Tarassenko, L., and Clifton, D. A.

(2018). Validation of instantaneous respiratory rate

using reﬂectance ppg from different body positions.

Sensors.

Kamal, A., Harness, J., Irving, G., and Mearns, A. (1989).

Skin photoplethysmography — a review. Computer

Methods and Programs in Biomedicine, 28(4):257–

269.

Kurylyak, Y., Lamonaca, F., and Grimaldi, D. (2013). A

neural network-based method for continuous blood

pressure estimation from a ppg signal. 2013 IEEE In-

ternational Instrumentation and Measurement Tech-

nology Conference (I2MTC), pages 280–283.

Li, H., Shrestha, A., Heidari, H., Le Kernec, J., and Fio-

ranelli, F. (2019). Bi-lstm network for multimodal

continuous human activity recognition and fall detec-

tion. IEEE Sensors Journal, 20(3):1191–1201.

Mauldin, T. R., Canby, M. E., Metsis, V., Ngu, A. H., and

Rivera, C. C. (2018). Smartfall: A smartwatch-based

fall detection system using deep learning. Sensors,

18(10):3363.

McCarthy, C., Pradhan, N., Redpath, C., and Adler, A.

(2016). Validation of the empatica e4 wristband. In

2016 IEEE EMBS international student conference

(ISC), pages 1–4. IEEE.

Naschitz, J. E. and Rosner, I. (2007). Orthostatic hypoten-

sion: framework of the syndrome. Postgraduate med-

ical journal, 83(983):568–574.

Oliver, D. and Healy, F. (2009). Falls risk prediction tools

for hospital inpatients: do they work? Nursing Times,

105(7):18–21.

Piccinno, A., Hemmatpour, M., Ferrero, R., Montrucchio,

B., and Rebaudengo, M. (2019). A review on fall pre-

diction and prevention system for personal devices:

Evaluation and experimental results. Advances in

Human-Computer Interaction.

Saleh, M., Abbas, M., and Le Jeann

es, R. B. (2021a). Fal-

lalld: An open dataset of human falls and activities

of daily living for classical and deep learning applica-

tions. IEEE Sensors Journal, 21(2):1849–1858.

Saleh, M., Abbas, M., Prudhomm, J., Somme, D., and

Le Bouquin Jeann

es, R. (2021b). A reliable fall de-

tection system based on analyzing the physical activ-

ities of older adults living in long-term care facilities.

IEEE Transactions on Neural Systems and Rehabili-

tation Engineering, PP:1–1.

Santos, G. L., Endo, P. T., Monteiro, K. H. d. C., Rocha, E.

d. S., Silva, I., and Lynn, T. (2019). Accelerometer-

based human fall detection using convolutional neural

networks. Sensors, 19(7):1644.

Tamura, T., Maeda, Y., Sekine, M., and Yoshida, M. (2014).

Wearable photoplethysmographic sensors—past and

present. Electronics, 3(2):282–302.

Vassallo, M., Mallela, S. K., Williams, A., Kwan, J., Allen,

S., and Sharma, J. C. (2009). Fall risk factors in el-

derly patients with cognitive impairment on rehabili-

tation wards. Geriatrics & gerontology international,

9(1):41–46.

HEALTHINF 2023 - 16th International Conference on Health Informatics

460