Classiﬁcation of Respiratory Diseases Using the NAO Robot

Rafael Andrade Rodriguez, Jireh Ferroa-Guzman and Willy Ugarte

Universidad Peruana de Ciencias Aplicadas, Lima, Peru

Keywords:

Classiﬁcation, NLP, NAO, Respiratory.

Abstract:

This work proposes an interface that connects the NAO robot with a development environment in Azure Ma-

chine Learning Classic for the prediction of respiratory diseases. The developed code uses Machine Learning

algorithms trained for the prediction of diseases and fatal symptoms in order to provide the user with a scope

of his health status and the possible conditions associated with his age, sex, symptoms and severity. During

this process, a brief discard of COVID-19 is made with the symptoms obtained, which indicates if they cor-

respond to those of this disease. Additionally, we offer a friendly interaction with the NAO robot to facilitate

the exchange of information and, at the end of the algorithm ﬂow, it is always suggested to use a professional

doctor to provide users with more details about their current status based on the overall results obtained. The

tests carried out on the work show that it is possible to speed up the time of care in medical care centers in Peru

through the Nao Robot. Additionally, it has been possible to predict respiratory diseases, which also helps the

doctor to have a notion of the patient prognosis.

1 INTRODUCTION

On March 11, 2020, the World Health Organisation

(WHO) declared SARS-Cov-2 a pandemic, due to

its far-reaching, affecting millions of people in sev-

eral countries around the world

. In Peru, COVID-

19 was ofﬁcially reported on March 6, 2020; Faced

with this, the Peruvian state declared on March 15 of

the same year, a state of Emergency, considering the

speed of progression of the disease and ruled manda-

tory quarantine at the national level (Miyahira, 2020).

On the 25th of the same month, the Peruvian Gov-

ernment established the measures that would lead cit-

izens towards a new social coexistence and the state

of emergency was extended due to the serious circum-

stances that affected the nation as a result of SARS-

Cov-2 (Barrutia-Barreto et al., 2021)

Despite the measures taken, the numbers of deaths

in Peru continued to grow. For the month of August

2020, Peru reached 613,378 infections, which made

it the sixth country with the most reported cases. At

that time, we reached 28 thousand deaths due to the

pandemic with a mortality rate of 85.8 per thousand

inhabitants

. During this period, there was evidence

https://orcid.org/0000-0002-7510-618X

WHO - https://covid19.who.int/

“Peru has the world’s highest COVID death rate.

Here’s why” - NPR - https://n.pr/3EBRukh

of a deﬁcient response by the public health system of

Peru, taking into account the number of deaths over

the number of infections (Gianella et al., 2021). How-

ever, this deﬁciency did not begin with the COVID-19

pandemic.

In the public sector of the Peruvian health system,

the government offers health services to uninsured

people in exchange for the payment of a fee through

the Integral Health System (SIS), with EsSalud be-

ing the entity that offers the services (Gianella et al.,

2021). During previous years, complaints and even

denunciations have been expressed by patients who

have health insurance within this center. Problems

such as speed of care, lack of medication and medical

malpractice are part of the large list of claims against

EsSalud. For example, in 2016, more than 111,000

claims for medical malpractice were ﬁled in EsSalud

for poor provision of services, which shows the dis-

satisfaction of users about the care provided

According to the National Institute of Statistics

and Informatics (INEI) of Peru, 25% of patients

treated at EsSalud have to wait between 15 to 30 days

to be able to schedule a simple medical consultation

while for surgical interventions, the time between the

date of programming and the date of intervention rises

to 2 months. On the other hand, for outpatient med-

“Complaints for lack of medical care and negligence

persist in EsSalud” (in spanish) - https://bit.ly/2XAjVby

940

Andrade Rodriguez, R., Ferroa-Guzman, J. and Ugarte, W.

Classiﬁcation of Respiratory Diseases Using the NAO Robot.

DOI: 10.5220/0011782700003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 940-947

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)

ical care, the waiting time in a Peruvian clinic is 42

minutes on average, a ﬁgure that for Essalud rises to

81 minutes. Therefore, many people are discouraged

from carrying out medical consultations in the public

health sector, also considering other problems such as

the level of distrust in medical personnel that still pre-

vails to date. Synthesizing the main problems during

consultations, it has been identiﬁed: insufﬁcient at-

tention time, high workload, patient anxiety or fear,

fear of physical and verbal abuse, unrealistic expecta-

tions of patients, fear of demands, patient resistance to

change and lack of training in this area. For all these

reasons, there are still barriers between doctors and

patients, which hinder the efﬁcient exchange of infor-

mation, which can have an impact on misdiagnoses,

which currently cover 11% of cases in the country

Within this context, it is important to consider

these shortcomings in the health system, maintain an

adequate doctor-patient relationship through patient

training and information, encourage health prevention

and promote adherence to treatments. For this, Peru-

vians need a source of information that serves as an

assistant and allows them to foresee the presence of

certain diseases based on probabilities. To meet these

needs, an interface capable of delivering results with

suggested illnesses and feedback on symptom sever-

ity to patients was developed through brief interaction

with the robotic assistant, NAO. The algorithm that is

handled in the interface was rigorously selected after

performing a comparative analysis with other classi-

ﬁcation algorithms prioritizing accuracy and avoiding

overtraining.

This paper is organized as follows. Therefor, in

Section 2, an analysis of the state of the art consid-

ered for this work will be made. Section 3, ﬁrst, in-

troduces the technologies used for the development of

the proposed solution.

Finally, Section 4 shows the experimental proto-

col, the results obtained, and the discussion. To con-

clude with Section 5

2 RELATED WORKS

The work of Fale (Fale and Gital, 2022) proposes

a hybrid of Mamdani type and Fuzzy Sugeno type

models by means of a fuzzy controller, follow a se-

quence of three steps: fuzziﬁcation; inference; and

defuzziﬁcation. Yuan’s work (Rozo et al., 2021) fo-

cuses on qualitatively detecting normal breathing and

Cheyne-Stokes breathing in patients with non-contact

“Medical error rate is around 11% in hospitals

and technology could change this ﬁgure” - https://bit.ly/

3V0PWHi

heart failure using orthogonal frequency division and

multiplexing technology (OFDM). On the other hand,

Mubashir’s work (Rehman et al., 2021) develops a

machine learning (ML) classiﬁcation model that is in-

telligent, secure, reliable and contributes to current

health systems by exploiting several Machine Learn-

ing algorithms to classify eight respiratory anomalies:

eupnea, bradypnea, tachypnea, Biot, sighs, Kussmaul,

Cheyne-Stokes and central sleep apnea (CSA). All

these works are oriented to respiratory diseases, just

like ours. However, we use Multiclass Decision Jun-

gle and Two Class Decision Forest as algorithms for

prediction, unlike the other methods mentioned.

In (Romero-Garc

ıa et al., 2021) evaluates the per-

formance of symptoms as a diagnostic tool for SARS-

CoV-2 infection using Mantel-Haenszel logistic re-

gression. In this area, in (Arslan, 2021), the authors

develop a prediction method based on the similarity of

the genome of human SARS-CoV-2 and a coronavirus

similar to bat SARS-CoV to predict this same disease.

Also, in (Brunese et al., 2020), the authors develop a

supervised machine learning model that discriminates

between COVID-19 and other lung diseases. These

three works are based on the detection or prediction

of COVID-19 and obtained the accuracies of 83.45%,

99.8% and 96.5% respectively according to the tests

performed. Unlike them, our work simply performs

a quick discard considering the symptoms mentioned

by the user, without additional analysis.

In (Yoon et al., 2019), a deep learning system is

developed using a recurrent neural network capable

of encoding and deciphering people’s postures in im-

ages and videos, and then being able to imitate them.

Similar to this, in (Filippini et al., 2021), the authors

design a CNN-based FER (Facial Expression Recog-

nition) model for facial expression recognition in real-

life situations. Both works handle neural networks

and employ computer vision, unlike us, who mainly

use the Audio service of the NAOqi library.

In (Burns et al., 2022), the authors attempt to

prove that the walking speed of the humanoid NAO

can be improved without modifying its physical con-

ﬁguration using decision trees and the ANN and

NAive Bayes models. On the other hand, in (Hoff-

mann et al., 2021), the authors develop a process

model with the components that are required to pass

the recognition test in front of a mirror. Regarding

our work, instead of working with decision trees, we

manage multiple DAG’s for disease prediction.

Classiﬁcation of Respiratory Diseases Using the NAO Robot

941

3 CLASSIFICATION TASK

3.1 Preliminary Concepts

3.1.1 Human-Robot Interaction

The study of human-robot interactions (also called

HRI) represents a multidisciplinary ﬁeld with contri-

butions from human-computer interaction, artiﬁcial

intelligence, robotics, natural language understand-

ing, design, and social sciences.

1. Robot NAO: Nao is a programmable and au-

tonomous humanoid robot developed by Alde-

baran Robotics. Nao is 4.3 kg in weight and has

a height of 58 centimeters. It is relatively light

and small, which makes it an ideal solution to

live with humans. Thanks to its prehensile ﬁnger

hands with tactile sensors, it is capable of lifting

objects of up to 600 grams. The different elements

of NAO, such as sensors, motors and software are

controlled by a powerful operating system called

NAOqi. All versions have an inertial measure-

ment unit with gyrometer, accelerometer and 4

ultrasound sensors, which provide the robot with

stability, while the leg versions include 8 force de-

tection resistors and 2 stops. The collaborative

robot includes 4 microphones, 2 speakers and 2

high-deﬁnition cameras. Also, it presents inter-

esting attributes and features such as a 25-degree

movement, 2 HD cameras, 2 speakers, Wi-ﬁ con-

nection and an Intel Atom 1.6 GHz processor

2. The robot has functionalities that the programmer

can use as resources to automate processes with

the NAO robot. In the interaction with the NAO

robot, it is necessary to have a copy connected

to a local IP of the home so that it can be con-

nected to a remote computer in which it is going

to be programmed. This process of connecting to

a network and synchronizing in the working envi-

ronment of the Python programming language is

understood as “AlProxy”.

3. Naoqi: NAOqi is an interpreter between the Nao

robot and Python programming that will allow us

to interact with the robot. It consists of a frame-

work that will allow to use the functionalities of

the robot and implement the Machine Learning al-

gorithm to the robot to process the received input

data.

• “ALProxy” command: This command is used

as a means of communication between the pro-

“Programming NAO robot with Python” - Softbank

Robotics Europe (2015) -https://www.youtube.com/watch?

v=iAeis7j5LmE

gramming interface and the NAO robot. Previ-

ously, the robot must be connected to the inter-

net so that it can be recognized by the program.

Its syntax consists of: an action to perform, the

IP address where the robot is connected and the

port where the NAO robot is connected.

• Action Name: NAOqi’s own SDK comes

with pre-programmed actions. All these pre-

programmed actions can help the program-

mer perform processes or automations with

the robot. From simple commands like say-

ing something by voice to taking pictures

and interpreting symbols. According to Soft-

Bank Robotics (2022), these are separated by

groups

(a) NAOQi core: Contains a list of functions that

allows you to interact with the NAO robot to

perform complex actions.

(b) NAOqi sensors and led: Contains the action

codes of the NAO robot with which it can in-

teract and program.

for managing video cameras, stereo cameras

and 2D cameras of the NAOqi robot.

(d) NAOqi Audio: Contains modules for record-

ing and playing audio, as well as for handling

the robot’s language.

(e) NAOqi people perception: Contains com-

mands are used to analyze human behavior

around the robot.

(f) Naoqi Motion: Contains the commands that

allow the movement of the NAO robot.

From these functionalities of the NAOqi library,

we take advantage of ALSpeechRecognition, a

NAOqi Audio command, which allows to inter-

pret the sounds or words that a human can make.

In this way, the NAO robot is able to capture the

input data necessary for symptom processing and

also, the user information that will be used to send

the results.

3.1.2 Basic Notions About Health and

Symptomatology

1. Symptomatology: Set of symptoms characteristic

of a given disease or grouping of symptoms that

occur in a patient.

2. Comorbidity: The presence of two or more asso-

ciated disorders or diseases in the same person,

occurring at the same time or one after the other

Co-morbidities - WHO - https://www.who.int/

southeastasia/activities/co-morbidities-tb

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

942

3. Interconsultation: Occurs when a doctor refers the

patient to another specialist doctor to handle com-

munication with different areas of expertise.

3.1.3 Classiﬁcation Models

1. Multiclass Decision Jungle: Represents an exten-

sion or modiﬁcation of the Decision Forest algo-

rithm. However, this consists of a set of acyclic

graphs that are driven by a decision (DAG). Mu-

ticlass Decision Jungle has the following advan-

tages:

(a) By allowing tree branches to merge, a deci-

sion DAG typically takes up less memory space

and has better generalization performance than

a decision tree, albeit at the cost of somewhat

longer training time.

(b) Decision jungles are nonparametric models that

can represent nonlinear decision boundaries.

built-in features and are resistant in the pres-

ence of noisy features.

The algorithm has many advantages in terms of

machine learning and has had considerable suc-

cess when testing. However, this also has a fun-

damental limitation that, with a lot of data, the

number of nodes in decision trees will grow ex-

ponentially with depth, limiting their use to only

certain platforms that can support this amount of

processing.

Figure 1: DAG visualization (Shotton et al., 2013).

2. Two Class Decision Forest: A decision forest de-

scribes a model made of multiple decision trees.

The prediction of a decision forest is the aggrega-

tion of the predictions of each decision tree.

Figure 2: Decision Forest ﬂow. (H

ansch, 2021).

3.2 Method

Our main contribution is the prediction of diseases

based on symptoms using a Multiclass Decision Jun-

gle. This algorithm is powered by a dataset published

on the Kaggle platform by American Health Info

For its adaptation in the Azure environment, we have

made some improvements such as data normaliza-

tion; That is, those symptoms that were written dif-

ferently, but corresponded to the same, were uniﬁed

into one. Also, because the dataset was unbalanced;

in other words, with unrepresented classes, we aggre-

gate records with missing severities (high, medium,

or low) for each symptom of each disease; so that the

code can predict diseases with any level of severity of

their symptoms. To exemplify this idea, if there are

no records of people with disease: asthma and symp-

tom: chest tightness, speciﬁcally with severity: low,

it is assumed that no user with these last two charac-

teristics can have asthma.

From the dataset handled, we use “Diseases”, as a

column to predict, and the columns of: “symptom”,

“sex”, “age” and “severity” as predictive variables.

We also use this dataset with an additional column of

Mortality, which indicates whether a symptom con-

sidering age and severity, can become fatal. To do

this, the code sends the data to another development

environment and tells the user if it detects several

deadly symptoms.

On the Azure Machine Learning Classic plat-

form, we ﬁrst load the normalized dataset and sec-

tion the columns relevant to our algorithm. Then,

we use 88% of the dataset for processing and 12%

for testing. Records for both proportions are random.

Then, we generate the training model, which connects

with the multiclass decision jungle algorithm. This

has at the beginning of 16 DAG’s (directed acyclic

graphs), with a maximum depth of 156 DAG’s and

https://www.kaggle.com/datasets/abbotpatcher/

respiratory-symptoms-and-treatment

Classiﬁcation of Respiratory Diseases Using the NAO Robot

943

Figure 3: Image of the updated Dataset.

width of 140 DAG’s. Subsequently, we create the

Score Model, which contains the percentages calcu-

lated with each test record selected for the testing pro-

cess and is part of the basis for knowing the accu-

racy of the algorithm in general. Finally, the Evaluate

model shows us the metrics calculated on the same

platform and we connect the input and output services

to make our architecture functional.

Figure 4: Architecture in Azure Machine Learning.

According to the ﬂow, in the ﬁrst instance, the user

indicates his personal data to the robot: Name, Age,

Sex, Symptoms and Severities; for the latter variables,

an iterative ﬂow is traversed that detects when there

are 3 or more symptoms of COVID-19. In case the

detection is positive, the user is consulted if a discard

test was done; If so, continue with the code and other-

wise, the ﬂow is cut off indicating that a test is carried

out as soon as possible. After completion of symp-

tom and mortality uptake; The NAO robot sends the

captured symptoms to the ﬁrst development environ-

ment in Azure, which will process one by one and

save the individual results, to then be averaged and

thus, calculate those diseases with more probability.

After this process, it connects with the second devel-

opmental environment and comments are issued re-

garding the mortality of the symptoms. Finally, the

user is queried for their WhatsApp number and, after

having captured it, the overall results are sent. During

the entire process of connection to the development

environments, it is necessary to execute a reconnec-

tion script with the NAO robot to connect to a net-

work that has access to the internet; which is required

to receive results from Azure. Subsequently, we link

again with the robot so that it continues its ﬂow.

Figure 5: Connection Flow with the environments.

This developed ﬂow has allowed us to speed up

and have a shorter attention time, thanks to the ﬂuid

conversation facilitated by the humanoid robot. Addi-

tionally, during the development and tests carried out,

we have been able to observe certain characteristics of

the robot that allow the correct reception of the input

data. These are as follows:

1. When talking to the NAO: We have to position

ourselves at a considerable height (higher than

that of the robot), because the microphones are

located on the upper front of the robot (Fig. 1)

2. Response time: When the robot makes a query,

you have to wait approximately 3 seconds to re-

spond to it.

3. Detection of the person: As seen in Fig. 1, the

sensors are located in the chest of the robot, which

means that it will detect the person in front of it,

therefore, it is recommended that, when talking to

the robot, the patient is in front of it, Not on the

sides.

Likewise, the symptoms have been classiﬁed in

order to simplify the process of receiving input data

so that the user can mention those symptoms recorded

that ﬁt their own during the consultation. The classi-

ﬁcations are as follows:

1. Cough Related

2. Related to breathing

3. Pains

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

944

4. Weight loss

5. Fatigue related

6. General discomfort

On the limitations and scope of our proposal, we

have the following:

1. Our model recognizes a total of 65 symptoms and

18 different respiratory diseases.

2. Our model is limited to respiratory diseases.

3. We must consider a waiting time to be able to give

an answer to each question asked by the robot.

This time is 5 seconds.

4. The ﬂow does not include adding new symptoms

or diseases (conditioned by the dataset).

The use of the NAO robot for our proposal, unlike

any other technological solution, is justiﬁed by rely-

ing on one of the causes of our main problem: the

level of distrust that exists towards medical person-

nel, which hinders an efﬁcient exchange of informa-

tion with patients. Faced with this, a study entitled

“Making eye contact with a robot: Psychophysiologi-

cal responses to eye contact with a human and with

a humanoid robot” concludes, after an experimen-

tal process, that in human-robot and human-human

condition, eye contact in front of the deviated gaze

caused a greater conductance of the skin, responses

associated with positive affect and deceleration of the

heart, that index attention allocation. In conclusion,

eye contact provokes affective and attentional reac-

tions when shared with a humanoid robot as well as

with another human.

4 EXPERIMENTATION

In this section, we present the experimental analysis

to demonstrate the feasibility of the proposal. We will

outline our experimental process and explain all the

considerations that were fundamental to obtain the ﬁ-

nal results.

4.1 Experimental Protocol

To carry out the experiments, various tools have been

used, both hardware and software. First, the NAO v.6

humanoid robot enabled interaction with users as well

as data capture in the experimental process. The pro-

gramming language of Python 2.7 and 3.10 has been

handled, along with the Naoqi framework. Regarding

our algorithms in Azure Machine Learning classic,

we have made use of the web service. On the other

hand, the computer that was responsible for execut-

ing the developed code has an AMD Ryzen 5 2500U

processor with 2.00 GHz and 24.0 GB of RAM.

Within our experimentation we have also tested

with different algorithms to be able to know which of

them would be better to be able to predict the diseases

that we are going to enter. Among these algorithms

are the Multiclass Decision Jungle, Multiclass Deci-

sion Forest, Multiclass Logitic Regression and Multi-

class Neural Network. Additionally, these algorithms

have internal parameters that we have been modify-

ing in order to ﬁnd the most appropriate model for the

prediction we want. In Table 1, we can ﬁnd the re-

sults of the algorithms with the default values given

by the Azure Machine Learning Classic; and in Ta-

ble 1b you can ﬁnd the results of the same algorithms

with the modiﬁed values. In Table 1c you can ﬁnd

the experimentation that has been carried out for the

part of the algorithm that predicts the mortality of the

disease.

Our code is currently available at https://github.

com/gareia/Dr Nao.git.

4.2 Results

In this section, we will detail the results of the tests

performed and show some videos in which the work-

ﬂow and each functionality mentioned below are vi-

sualized. The complete ﬂow of our proposal can be

visualized in our video

Reception Input Data. The reception of name is

conditioned by a dataset, which initially overloaded

the vocabulary allowed by Naoqi. Other input values

such as Age and Sex are received without major prob-

lems; However, in the symptoms section, at times the

mentioned symptom is not recognized. Here, several

factors such as ambient noise or similarity in the name

of different symptoms can inﬂuence.

Connection to Development Environment. The

API Key and URL generated by Azure Machine

Learning classic allow us to connect with the devel-

opment environment. This is functional at all times

and the disconnection/reconnection script with NAO,

as well as the network with deﬁned internet access,

works properly without errors.

Discard COVID-19: The ﬂow is properly met. When

symptoms related to COVID-19 are detected, the user

is asked if the COVID-19 test was performed to end

the ﬂow in case of responding negatively.

Reception and Sending of Results. It was possible

to send results through a WhatsApp account linked

to the selected browser. It is recommended that it is

https://youtu.be/sw7LpUie2TA

Classiﬁcation of Respiratory Diseases Using the NAO Robot

945

Table 1: Machine Learning Algorithms Metrics Charts.

(a) Default Values.

Multiclass

Decision

Jungle

Multiclass

Decision

Forest

Multiclass

Logitic

Regression

Multiclass

Neural

Network

Overall accuracy .820627 .805957 .914870 .907313

Average accuracy .980070 .978440 .990541 .989701

Micro-averaged precision .820627 .805957 .914870 .907313

Macro-averaged precision .861564 .838173 .934235 .918068

Micro-averaged recall .820627 .805957 .914870 .918068

Macro-averaged recall .760552 .768652 .896562 .903236

(b) Modiﬁed values.

Multiclass

Decision

Jungle

Multiclass

Decision

Forest

Multiclass

Logitic

Regression

Multiclass

Neural

Network

Overall accuracy .817070 .805957 .918649 .907535

Average accuracy .979674 .978440 .990961 .989726

Micro-averaged precision .817070 .805957 .918649 .907535

Macro-averaged precision .862696 .838173 .937841 .918490

Micro-averaged recall ,817070 .805957 .918649 .907535

Macro-averaged recall .755444 .768652 .899862 .901638

Two Class

Bayes

Point

Machine

Two Class

Averaged

Perceptron

Two Class

Boosted

Decision

Tree

Two Class

Decision

Forest

Accuracy .962 .273 .273 .988

Precision .951 .273 .273 .984

Recall precision .907 1.000 1.000 .972

F1 Score .929 .430 .430 .978

Threshold .500 .500 .500 .500

AUC .992 .164 .000 .999

already open and that it has synchronized correctly to

avoid unnecessary delays.

According to Tables 1 and 1b, and an analysis car-

ried out, for the detection of diseases, we have cho-

sen the Multiclass Decision Jungle as an algorithm

for the detection and diagnosis of diseases. On the

other hand, according to Table 1c, for the detection

of mortality we have selected the Two Class Decision

Forest.

4.3 Discussions

The reason we chose the Multiclass Decision Jungle

over the other classiﬁcation algorithms is justiﬁed in

the precision matrix. The matrix, as shown in Figure

7, with the lowest amount of empty blocks is beneﬁ-

cial for the proposal, since it rules out fewer diseases

during detection and favors obtaining more real re-

sults. In addition to that, it is the one that has the best

parameters and results at the time of experimentation,

not having an overtraining or having too low values.

The reason we chose the Two Class Decision For-

est is that, over the other classiﬁcation algorithms we

have tested, this is the one that has returned us bet-

ter results. This, because it has a better Accuracy

and Precision that will help to obtain better results

from our algorithm. Additionally, it is better than

the Two Class Averaged Perceptron and Two Class

Boosted Decision Tree algorithms, as these return

very low values to be selected. Regarding the Two

Class Decision Forest, it obtained metrics quite simi-

lar to the Two Class Bayes Point Machine, so any of

these would have been useful.

Initially, we want to recognize the user’s name by

assigning a Dataset of names; However, the number

Figure 6: Multiclass Decision Jungle Precision Matrix.

of records exceeded 40,000 names, which overloaded

the vocabulary of the robot and began to present errors

and slowness during testing. From there, we limit the

number of names to recognize and avoid overloading

the vocabulary with a lot of data. Among those that

occupy the most space are combinations of numbers

and names. Subsequently, we make improvements so

that, before receiving each input data, the vocabulary

is conﬁgured to be more limited and less overloaded.

5 CONCLUSIONS

First, the Nao has technical tools that have very been

useful, however, it has limitations in terms of Speech

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

946

Recognition. Additionally, it has sensitivity to noise,

that is, if in the place where we are there is too much

noise around, the Nao will not be able to detect the

voice of the person or may have difﬁculties. Testing

the code was difﬁcult at this stage because it’s not pos-

sible to test with Choreography if input data in audio

format is needed. Also, the person cannot be at a great

distance from the microphones of the Robot (which

are in his head), otherwise, the listening of this will

be low and may have problems to understand the mes-

sage. According to this, we noticed that the Nao robot

has a greater facility to capture numbers when listen-

ing to them, than large words. You are more likely to

ask to repeat the word than to ask to repeat the dic-

tated number. Another limitation is related to Internet

access, because if it were possible to access the In-

ternet connected to the Nao robot, our full ﬂow time

will be severely reduced. Finally, special care must

be taken when training with the classiﬁcation algo-

rithms, because some datasets can generate overtrain-

ing, which would generate irregular results (Burga-

Gutierrez et al., 2020).

For future improvements, more diseases can be

added to the dataset so that it can cover a larger ﬁeld

and can be run again with the same algorithm. Al-

though our premise is that the input data is said aloud,

our ﬂow time can decrease if the patient, instead of

dictating the symptoms one by one, can have a table

with the total symptom numbers and tell the numbers

to the robot. Similar to this, the reception of the tele-

phone number generates that the time ﬂow increases

considerably. We suggest that this data can be typed

and the results can be sent to the doctor if applica-

ble. Additionally, by skipping this step, we can avoid

scaring the patient, because we do not know how sen-

sitive he may be and may even misinterpret the robot’s

comments on the results. Furthermore, combining our

approach with other kinds of smart health allocation

systems (Ugarte, 2022).

REFERENCES

Arslan, H. (2021). COVID-19 prediction based on genome

similarity of human sars-cov-2 and bat sars-cov-like

coronavirus. Comput. Ind. Eng., 161.

Barrutia-Barreto, I., S

anchez-S

anchez, R. M., and Silva-

Marchan, H. A. (2021). Consecuencias econ

omicas

y sociales de la inamovilidad humana bajo covid – 19

caso de estudio per

u. Lecturas de Econom

ıa, 1(94).

Brunese, L., Martinelli, F., Mercaldo, F., and Santone, A.

(2020). Machine learning for coronavirus covid-19

detection from chest x-rays. Procedia Computer Sci-

ence, 176.

Burga-Gutierrez, E., Vasquez-Chauca, B., and Ugarte, W.

(2020). Comparative analysis of question answering

models for HRI tasks with NAO in spanish. In SIM-

Big.

Burns, R. B., Lee, H., Seiﬁ, H., Faulkner, R., and Kuchen-

becker, K. J. (2022). Endowing a NAO robot with

practical social-touch perception. Frontiers Robotics

AI, 9.

Fale, M. I. and Gital, A. Y. (2022). Dr. ﬂynxz - A ﬁrst aid

mamdani-sugeno-type fuzzy expert system for differ-

ential symptoms-based diagnosis. J. King Saud Univ.

Comput. Inf. Sci., 34(4).

Filippini, C., Perpetuini, D., Cardone, D., and Merla, A.

(2021). Improving human-robot interaction by en-

hancing NAO robot awareness of human facial expres-

sion. Sensors, 21(19).

Gianella, C., Gideon, J., and Romero, M. J. (2021). What

does covid-19 tell us about the peruvian health sys-

tem? Canadian Journal of Development Studies / Re-

vue canadienne d’

etudes du d

eveloppement, 42(1-2).

ansch, R. (2021). Handbook of Random Forests - Theory

and Applications for Remote Sensing. Series in Com-

puter Vision. WorldScientiﬁc.

Hoffmann, M., Wang, S., Outrata, V., Alzueta, E., and

Lanillos, P. (2021). Robot in the mirror: Toward

an embodied computational model of mirror self-

recognition. K

unstliche Intell., 35(1).

Miyahira, J. (2020). Lo que nos puede traer la pandemia.

Revista Medica Herediana, 31(2).

Rehman, M., Shah, R. A., Khan, M. B., Shah, S. A., Abuali,

N. A., Yang, X., Alomainy, A., Imran, M. A., and

Abbasi, Q. H. (2021). Improving machine learning

classiﬁcation accuracy for breathing abnormalities by

enhancing dataset. Sensors, 21(20).

Romero-Garc

ıa, R., Mart

ınez-Tom

as, R., Pozo, P., de la

Paz, F., and Sarri

a, E. (2021). Q-CHAT-NAO: A

robotic approach to autism screening in toddlers. J.

Biomed. Informatics, 118.

Rozo, A., Buil, J., Moeyersons, J., Morales, J. F., van der

Westen, R. G., Lijnen, L., Smeets, C., Jantzen, S.,

Monpellier, V., Ruttens, D., Hoof, C. V., Huffel, S. V.,

Groenendaal, W., and Varon, C. (2021). Controlled

breathing effect on respiration quality assessment us-

ing machine learning approaches. In IEEE CinC.

Shotton, J., Sharp, T., Kohli, P., Nowozin, S., Winn, J. M.,

and Criminisi, A. (2013). Decision jungles: Compact

and rich models for classiﬁcation. In NIPS.

Ugarte, W. (2022). Vaccination planning in peru using con-

straint programming. In ICAART.

Yoon, Y., Ko, W., Jang, M., Lee, J., Kim, J., and Lee, G.

(2019). Robots learn social skills: End-to-end learn-

ing of co-speech gesture generation for humanoid

robots. In IEEE ICRA.

Classiﬁcation of Respiratory Diseases Using the NAO Robot

947