Ensemble Learning for Cough-Based Subject-Independent

COVID-19 Detection

Vincenzo Conversano and Stavros Ntalampiras

Department of Computer Science, University of Milan, Via Celoria 18, Milan, Italy

Keywords:

Medical Acoustics, Audio Pattern Recognition, Machine Learning, COVID-19 Detection.

Abstract:

This paper belongs to the medical acoustics ﬁeld and presents a solution for COVID-19 detection based on the

cough sound events. Unfortunately, the use of RT-PCR Molecular Swab tests for the diagnosis of COVID-19

is associated with considerable cost, is based on availability of suitable equipment, requires a speciﬁc time

period to produce the result, let alone the potential errors in the execution of the tests. Interestingly, in addition

to Swab tests, cough sound events could facilitate the detection of COVID-19. Currently, there is a great

deal of research in this direction, which has led to the development of publicly available datasets which have

been processed, segmented, and labeled by medical experts. This work proposes an ensemble composed of

a variety of classiﬁers suitably adapted to the present problem. Such classiﬁers are based on a standardized

feature extraction front-end representing the involved audio signals limiting the necessity to design handcrafted

features. In addition, we elaborate on a prearranged publicly available dataset and introduce an experimental

protocol taking into account model bias originating from subject dependency. After thorough experiments,

the proposed model was able to outperform the state of the art both in patient-dependent and -independent

settings.

1 INTRODUCTION

The coronavirus disease, widely known as COVID-

19, is a severe acute respiratory syndrome (SARS-

CoV2) which ﬁrst appeared in Wuhan, China, and

quickly spread to the entire world

. COVID-19,

which has been declared a pandemic by the World

Health Organization, causes severe infections to the

respiratory human system and is associated with very

high mortality rates as it has led to approximately 5.7

million deaths. A fundamental step towards facing

and potentially containing this pandemic consists in

having available knowledge regarding contamination,

i.e. reliable testing tools (Lippi et al., 2020). This

posed a signiﬁcant challenge since testing tools might

be time-consuming and/or of limited quantities to sat-

isfy the ever-growing demand. Unfortunately, the

COVID-19 pandemic demonstrated the lack of suit-

able testing capacity across the globe

. The world is

facing an unprecedented loss of human lives, not to

mention the huge consequences across the entire eco-

https://orcid.org/0000-0003-3482-9215

https://www.who.int/emergencies/diseases/novel-

coronavirus-2019

https://www.aacc.org/science-and-research/covid-19-

resources/aacc-covid-19-testing-survey

nomic sector, where there are large inequalities be-

tween developed and developing countries (Nessiem

et al., 2021). Such inequalities are particularly evi-

dent in testing equipment and materials in develop-

ing countries resulting to poor assessment in the dif-

fusion of the virus in the community and ineffective

decision making regarding lockdown measures with

severe consequences on the society. Such a challenge

could be lightened by the availability of pre-screening

tools which are inexpensive and can be easily ac-

cessed by interested subjects. This work is based on

the premises that the analysis of cough sound events

could comprise such a pre-screening tool.

Motivated by the speciﬁc problem, during the last

couple of years there has been a great deal of research

in the ﬁeld of medical acoustics (Poir

e et al., 2022;

Cozzatti et al., 2022) focusing on the use of signal

processing and pattern recognition algorithms for the

diagnosis and prognosis of COVID-19 suspected pa-

tients. There are several works in the ﬁeld of image

processing which employ lung radiographs (X-rays)

or lung computed tomography (CT) scans. However,

these methods are certainly invasive and imply a con-

siderable cost (Casiraghi et al., 2020; Ning et al.,

2020). Indeed, the usage of audio signal processing

and pattern recognition tools has been shown to fa-

798

Conversano, V. and Ntalampiras, S.

Ensemble Learning for Cough-Based Subject-Independent COVID-19 Detection.

DOI: 10.5220/0011651700003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 798-805

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)

Figure 1: The block-diagram of the proposed method for diagnosing COVID-19 in cough sound events.

cilitate the diagnosis of respiratory diseases (Ntalam-

piras and Potamitis, 2019). Interestingly, during the

last couple of years, several researchers have followed

such a line of thought and approached the COVID-

19 detection problem based on the associated cough

sound events. The literature includes several audio

processing methodologies (both handcrafted features

and automatically discovered) combined with various

classiﬁcation mechanisms, where the efﬁcacy of both

traditional and deep learning methods has been in-

vestigated. Some representative papers are described

next. Tena et al. (Tena et al., 2022) designed an

automated front-end combined with Random Forest

classiﬁer while they employed a combination of pro-

prietary and publicly available datasets. Imran et al.

(Imran et al., 2020) developed a smartphone applica-

tion which is based on Mel-spectrograms and a Con-

volutional Neural Network using a proprietary dataset

recorded using a smartphone. Last but not least, Er-

dogan and Narin (Erdo

gan and Narin, 2021) report

a framework exploiting both traditional and deep-

learned features modeled by means of Support Vector

Machines using data captured by a mobile applica-

tion

A common challenge in existing works is the

availability of reliable data along with a standardized

experimental protocol. For example, most studies

report results based on cross validation experiments

without mentioning whether a subject independent di-

vision was followed which might introduce a bias in

the analysis. Moreover, when using multiple datasets

it is essential that data balancing and a consistent an-

notation protocol are followed.

Unfortunately, to the best of our knowledge such

aspects have not been considered so far in the re-

lated literature. The only work which scratches the

surface of these issues is reported in (Xia et al.,

2021) where a systematic comparison of available

datasets is included. It comes out that the only dataset

which has been processed and labeled by medical ex-

perts is COUGHVID (Orlandic et al., 2021). Im-

https://virufy.org/

Figure 2: Representative Mel-scaled spectrograms of a

COVID-19 and a healthy sample along with the PCA-based

visualization of the feature space.

portantly, it encompasses more that 2,800 record-

ings which have been annotated by four experienced

physicians to diagnose medical abnormalities present

in the coughs. Cough sound events labeled as symp-

tomatic and COVID-19 originate from countries with

high infection rates. As such, it is reliable, e.g. not

based solely on crowd-sourced data; nonetheless it is

of limited quantity and not adequate to train very deep

models. On top of that, it is publicly available, i.e.

every interested researcher can access it, thus provid-

ing the opportunity of adopting a standardized experi-

mental protocol facilitating the comparison of diverse

frameworks.

Motivated by the above, this work a) considers

a subject-independent experimental protocol, b) is

based exclusively on reliable and publicly available

data ensuring reproducibility, c) optimizes a series

of classiﬁers to the speciﬁc problem, the learning

of which does not require enormous data quantities,

d) proposes a systematic cooperative framework com-

bining the beneﬁts of heterogeneous classiﬁers, and

e) addresses the need of prediction interpretability

since the operations carried out by the considered

classiﬁers can be easily backtracked to the input fea-

tures which are free from domain expertise.

Ensemble Learning for Cough-Based Subject-Independent COVID-19 Detection

799

2 THE PROPOSED COVID-19

DETECTION FRAMEWORK

This section describes the two main modules of the

proposed framework, i.e. feature extraction and clas-

siﬁcation.

2.1 Feature Extraction and PCA-Based

Visualization

Aiming at a standardized front-end, i.e. not necessi-

tating domain expertise, we employed the following

audio features, which are considered informative for

generalized sound recognition:

• Mel-Frequency Cepstral Coefﬁcients (40 fea-

tures): They comprise a summarization of the

spectrum, appropriately converted in the Mel

scale and spaced using the log operator (Ntalam-

piras, 2020b).

• MFCCs derivative (40 features): they are useful

to understand the direction of the evolving power

spectrum over time.

• Spectral Contrast (7 features): it monitors the dif-

ference between peak and valley energy across the

frequency content.

Fig. 2 demonstrates Mel-scaled spectrograms, i.e.

before the application of DCT, extracted out of repre-

sentative healthy and COVID-19 cough sound events.

In order to compare the subsets of the features and

identify the optimal combination, we computed the

PCA plots and F1-score for every combination. As

we see in Table 1, the best-performing subset con-

sists in MFCCs and Contrast. A PCA-based visual-

ization of the obtained feature space is illustrated in

Fig. 2, where we see a great overlap between healthy

and COVID-19 recordings.

2.2 Classiﬁcation Models

This section describes brieﬂy the considered classi-

ﬁcation models including information regarding the

hyper-parameters optimization process. Classiﬁers of

heterogeneous characteristics were included aiming

at an ensemble able to beneﬁt from the advantages

of each individual classiﬁer. Each one was cross-

validated and tuned to optimize the F1-score in each

fold, while data division was kept constant during ev-

ery experimental phase. For each fold, we stored pre-

dictions of the test samples with respect to each opti-

mized classiﬁer. These predictions allowed us to com-

pute global performance metrics and, importantly, a

similarity score between classiﬁers which was con-

sidered when populating the ensemble.

2.3 k-Nearest Neighbors

k-NN classiﬁer is a lazy learner that does not cre-

ate explicit models but rather exploits the k-nearest

neighbors’ classes to categorize new data points. The

parameters we tuned for this classiﬁer are:

• N: is represents the number of neighbors con-

sidered when making a prediction (search space

{1,3,5,7,9,11,13,15}).

• Weighting scheme: it describes the way neighbor-

hood points are weighted (uniform or distance-

based)

• Metric: it is the distance metric used to evalu-

ate neighbors; it was optimized among Euclidean,

and Manhattan.

• Algorithm: it deﬁnes the type of algorithm used

for searching the nearest neighbors; it was opti-

mized among auto, ball-tree, kd-tree, and brute.

• Leaf size: it is used in the case

of ball-tree or kd-tree (search space

{10,20,30,40,50,60,70,80,90}).

2.4 Random Forest

The speciﬁc classiﬁer creates multiple weak-learners

(that demonstrate high variance but low bias) and

combine them in order to robustify the achieved pre-

dictions (Ho, 1998). The Random Forest algorithm

ﬁts a plethora of Decision Trees (DT), typically on

different bootstrap samples, while each tree elabo-

rates on a random subset of features. Here, the fol-

lowing parameters were tuned:

• N-estimators: the number of DT (search space:

from 10 to 200 with step 10).

• Max depth: the maximum depth of each DT. It

should be noted that growing very deep trees can

lead to overﬁtting (search space {2,8,16,32,50}).

• Min samples split: it represents the minimum

number of samples required to split a node (search

space {2,4,6}).

• Min samples leaf: it represents the minimum

number of samples required for a node to be con-

sidered a leaf (search space {1,2}).

• Max features: the maximum number of features to

consider when searching for the best split (search

space {

√

# f eatures,log

(# f eatures)}.

• Bootstrap: considering the entire dataset or boot-

strap samples for growing each tree.

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

800

Table 1: Performance comparison between subsets of features (the highest F1 score is emboldened).

Features subset F1 score

MFCCs, MFCCs-delta, Contrast 0.7153

MFCCs, MFCCs-delta 0.7144

MFCCs, Contrast 0.7385

MFCCs-delta, Contrast 0.6983

MFCCs-delta 0.7056

Contrast 0.6352

MFCCs 0.7303

2.5 Extra Trees

Even though the Extra Trees classiﬁer follows a simi-

lar line of thought as RF, there is a relevant difference

when choosing the thresholds to use for each split.

In fact, these are drawn randomly for each candidate

feature and the optimal point is chosen as the splitting

rule. Such randomness may provide a diverse type of

information to the ensemble, while reducing the com-

putational complexity (Geurts et al., 2006). Finally,

the parameters which need tuning are the same with

the RF classiﬁer.

2.6 Support Vector Machine

Support Vector Machines map the initial feature space

X to a higher dimensional one φ(X) and aim at dis-

covering a hyperplane that separates the training in-

stances into two classes. As such, they are based

on the assumption that in higher dimensional spaces,

there is a hyperplane dividing the data representing

different classes (Cortes and Vapnik, 1995).

Here, the parameters that need tuning are:

• C: a regularization parameter that in the context of

soft-margin SVMs regulates the trade-off between

maximizing the margin and minimizing the train-

ing error. The larger C the more emphasis will be

placed on minimizing the training error. Usually,

it is preferable to achieve a reasonable training er-

ror at a larger margin to avoid model overﬁtting

(search space {0.1,1,10,100,1000}).

• Kernel function: computations in the high dimen-

sional space involve the dot product φ(x

)·φ(x

which has been deﬁned as the Kernel function; in-

terestingly, it is expressed in the original feature

space and it avoids the necessity of deﬁning the

mapping function. The following functions have

been explored: Radial Basis Function, Polyno-

mial, and Linear.

• Degree: it is used in case of Polynomial kernel

function (search space: {1, 2, 3, 4, 5, 6, 7, 8, 9}).

• Gamma: it represents the allowable curva-

ture of the decision boundary (search space:

{1/# f eatures ×var(X ), 1/# f eatures}).

2.7 Light Gradient Boosting Machine

This classiﬁer sequentially ﬁts a group of DTs, where

each training iteration focuses on previously misclas-

siﬁed samples as revealed by Residual Errors. We em-

ployed the gradient boosting framework LightGBM

(Ke et al., 2017) which can operate efﬁcienlty in dis-

tributed settings.

LightGBM is formed by the following parameters:

• N-estimators: the number of trees to grow (search

space: from 10 to 200 with step 10).

• Max depth: the maximum depth of each tree

(search space: {2,8,16,32,50}).

• Min child samples: it denotes the minimum num-

ber of samples required for a node to be consid-

ered as a leaf (search space {1,2}).

• Learning rate: it is the boosting learning rate

(search space {0.01,0.1,0.2,0.3}).

2.8 Multilayer Perceptron

A Multilayer Perceptron network employs multiple

perceptrons distributed in a multi-layered scheme

where the techniques of Gradient Descent and Back-

propagation are applied to update the included

weights and learn the characteristics of the training

set (Rumelhart et al., 1986).

The parameters we tuned for this classiﬁer are:

• Number of layers (search space {1,3,5}).

• Number of units: it is the number of perceptrons

for a layer (search space: from 32 to 128).

• Dropout rate: it discards information in order to

avoid overﬁtting (search space: from 0.1 to 0.9

with step 0.1).

• Learning rate: it represents the rate with which

each weight is updated (search space: from

0.00001 to 0.1).

Ensemble Learning for Cough-Based Subject-Independent COVID-19 Detection

801

Table 2: Similarity-based ranking of the considered classiﬁers (in ascending order).

Classiﬁer MLP k-NN SVM LGBM ET RF

0.7 0.7037 0.7304 0.7423 0.7497 0.7552

• Decay: it modiﬁes the learning rate over time

(search space: from 0.000001 to 0.01).

• Batch size: it represents the quantity of data points

considered in each training iteration (search

space: from 8 to 512 with step 8).

It should be noted that the number of units along with

the dropout rate are not ﬁxed and may change over the

layers.

3 SOFT MAJORITY VOTING

SCHEME

Combining predictions obtained from various classi-

ﬁers of heterogeneous properties may lead to perfor-

mance improvements (Ntalampiras, 2020a). To this

end, we create a collaborative meta-classiﬁer that uses

a Majority Voting scheme to perform the ﬁnal clas-

siﬁcation based on the predictions of the individual

classiﬁers.

More speciﬁcally, we employed a soft voting

scheme where every classiﬁer can predict the prob-

ability for a sample to belong to a class and the ﬁnal

prediction is the class maximizing the sum of proba-

bilities, i.e.

prediction = argmax

y∈Y

|C|

∑

i=1

P(C

(x) = y),

where x is the extracted feature vector, C the set of

classiﬁers of the ensemble, P(C

(x) = y) the probabil-

ity of a correct prediction given classiﬁer C

, and |•|

denotes the cardinality operator.

In order to maximize the efﬁciency the ensemble,

it is preferable to encompass diverse methodologies;

as such, we performed a pair-wise comparison be-

tween the available classiﬁers (Dietterich, 2000). To

this end, for each pair of classiﬁers (i, j), we com-

puted their similarity λ

i j

as the number of equally

classiﬁed samples. The result is demonstrated in Fig.

3. Subsequently, we used these similarity scores to

compute for each classiﬁer a global similarity score

6×3917

∑

j∈C

i j

, where 6 is the number of

considered classiﬁers and 3917 the number of testing

samples. The obtained ranking is shown in Table 2.

After early experimentations, we decided to heuristi-

cally populate the ensemble with the top four classi-

ﬁers, i.e. the most dissimilar four.

4 EXPERIMENTAL SETUP AND

RESULTS

This section describes a) the data preprocessing, seg-

mentation and ﬁltering processes, b) the parameter-

ization of the included features and classiﬁers, and

c) the experimental results including a comparison

with the state of the art.

4.1 Data Preprocessing and Cough

Segmentation

Keeping in mind the reproducibility of the experi-

ments, this work is based on the publicly-available

COUGHVID dataset

, which the largest COVID-19

audio dataset. As shown in Table 3, the dataset

is highly unbalanced across the considered classes,

i.e. healthy, unknown, symptomatic, and conﬁrmed

COVID-19 cases. As such, we analyzed the distribu-

tion of the cough detected attribute provided in the

metadata, reporting the probability of cough sound

events in a given audio sample (Orlandic et al., 2021).

This attribute was computed using a XGB classiﬁer,

trained and validated on 68 audio features (MFCCs,

EEPD, ZCR, etc.) of 215 randomly selected audio

samples. Since a threshold of 0.8 resulted in an aver-

age precision of 95.4%, it was used during the ﬁlter-

ing phase. Thus, we discarded audio samples a) with

cough detection score less than 0.8, b) labeled as un-

known or symptomatic, c) of poor quality based on

Figure 3: Similarity heatmap representing scores λ

i j

’s for

every pair of considered classiﬁers.

https://zenodo.org/record/4498364#.YcbxwVnSJPY

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

802

Table 3: Composition of the COUGHVID dataset (Orlandic

et al., 2021).

Class #audio samples Avg. duration (s)

Healthy 12479 7.58

Unknown 11326 7.27

Symptomatic 2590 7.57

COVID-19 1155 7.65

the associated label, and d) that were not analyzed

by medical experts. It should be noted that symp-

tomatic label denotes audio samples that come from

people who present COVID-19 symptoms but have

not received a diagnosis, i.e. they are not conﬁrmed

cases. Interestingly, such a cough segmentation pro-

cess resulted in 509 healthy audio samples encom-

passing 2088 cough sound events, and 454 conﬁrmed

COVID-19 audio samples with 1829 cough sound

events.

It should be mentioned that all audio signals were

sampled at 22050Hz, while cough segmentation is

based on the hysteresis comparator of the signal’s en-

ergy (Orlandic et al., 2021). As such, the updated

dataset containing only cough segments exhibits a sat-

isfactory balance across the healthy and COVID-19

classes.

4.2 Framework Parameterization

This section describes the parameterization process of

the feature extraction and classiﬁcation phases. As

regards to the feature extraction process:

• MFCCs: 40 DCT coefﬁcients were used along

with hamming-windowed frames of size 25 ms

overlapped by 50%. The FFT size is 512.

• Spectral Contrast: 7 dimensions were employed

to characterize it.

It should be mentioned that z-score normalization was

used (x

norm

x−¯x

) while the features were averaged

across each segmented cough.

For each of the considered classiﬁers, we carried

out ten fold cross-validation, while during each iter-

ation, hyperparameters were tuned to maximize F1-

score on a subsplit, i.e. 30%, of the training set. The

best-performing parameters with respect to each clas-

siﬁcation model, i.e. the ones that achieved the high-

est F1-score on are tabulated in Table 4.

4.3 Experimental Results

In this section, we report the experimental results and

compare the performance of the proposed COVID-19

detection framework to existing works addressing the

same task. Aiming at a reliable comparison, we relied

on ﬁgures of merit which are well-established in the

related literature, i.e. F1-score, Sensitivity, and Speci-

ﬁcity along with the ten-fold cross validation data di-

vision scheme. The achieved results are tabulated in

Table 5.

First, we observe that there is a signiﬁcant dif-

ference when subject dependency is considered dur-

ing the data division process. The experiments con-

ﬁrm that cough sound events are not statistically inde-

pendent since training and testing on samples coming

from the same subject heavily inﬂuences the models’

performance. More precisely, we see that all ﬁgures

of merit improve with the F1 score increasing from

0.6 to 0.82, sensitivity from 0.6 to 0.8 and speciﬁcity

from 0.64 to 0.84 when the proposed framework is

used. As such, the bias introduced by including sam-

ples of the same subject in both train and test sets is

conﬁrmed and this point should be considered when

conducting future research. Moreover, it is not rea-

sonable to assume a-priori availability of COVID-19

samples of a subject existing in the testing set, mean-

ing that it would have to be a least the second time

that the speciﬁc subject is infected.

Second, we observe that the performance reached

by the proposed framework is encouraging, especially

when considering that samples of different classes are

acoustically similar (see also Fig. 2) to the extent that

a non-expert human listener can assess. This high-

lights the great necessity of reliable data annotated by

medical experts. In fact, the obtained ﬁgures of merit

are well above chance, while the F1-score is 0.6. In-

terestingly, true positive and negative rates are 0.6 and

0.64 which demonstrates that the framework is almost

equally effective in identifying healthy and infected

cough sound events. This feature is particularly im-

portant given that our aim at a pre-screening tool, so

that not only potential COVID-19 cases are correctly

identiﬁed but, at the same time, unnecessary use of

medical services is limited. The overall performance

of the ensemble is severely boosted when a subject

dependent protocol is followed as explained earlier.

Third, we carried out a comparison with ap-

proaches using the same experimental protocol and

dataset. In Table 5, we see that the ensemble signif-

icantly outperforms (Rao et al., 2021) and (Agbley

et al., 2020) which achieve F1 score of 0.4 and 0.62

in a subject dependent setting. At the same time, it

offers a balanced sensitivity and speciﬁcity ﬁgures of

merit.

The results conﬁrm that a summarized Mel-scaled

spectrum modeled by a set of diverse classiﬁcation al-

gorithms may provide efﬁcient COVID-19 detection

in cough sound events. An important characteristic of

the proposed solution is the ability of the ensemble

Ensemble Learning for Cough-Based Subject-Independent COVID-19 Detection

803

Table 4: Classiﬁers’ parameters after optimization.

Classiﬁer Parameters

k-Nearest Neighbors N-neighbors: 1, Weights: uniform

Metric: Euclidean, Leaf size: 10

Random Forest N-estimators: 190, max depth: 16

Min samples split: 6, Min samples leaf: 2

Max features: log2, Bootstrap: false

Extra Trees N-estimators: 190, max depth: 50

Min samples split: 4 , Min samples leaf: 1

Max features: auto, Bootstrap: false

Support Vector Machine C: 10, kernel function: poly

Degree: 5, Gamma: auto

Light Gradient N-estimators: 160, Max depth: 50

Boosting Machine Min child samples: 2, Learning rate: 0.3

Multilayer Perceptron Number of layers: 5

Number of units: 128, 32, 32, 128, 32

Dropout rate: 0, 0, 0, 0.9, 0

Learning rate: 0.1, Decay: 0.01, Batch: 512

Table 5: The ﬁgures of merit achieved by the proposed and contrasted approaches (SI: subject independent, SD: subject

dependent). The highest one w.r.t each protocol is emboldened.

Work (protocol) F1-score Sensitivity Speciﬁcity

Ensemble (SI) 0.59±0.04 0.59±0.05 0.64±0.05

Ensemble (SD) 0.81±0.01 0.82±0.03 0.83±0.01

k-NN (SD) 0.69 ±0.03 0.64 ±0.02 0.74 ±0.03

RF (SD) 0.66 ±0.03 0.71 ±0.04 0.70 ±0.02

ET (SD) 0.63 ±0.04 0.72 ±0.04 0.68 ±0.02

SVM (SD) 0.71 ±0.03 0.71 ±0.03 0.74 ±0.03

LGBM (SD) 0.66 ±0.03 0.67 ±0.02 0.70 ±0.03

MLP (SD) 0.59 ±0.18 0.70 ±0.05 0.68 ±0.06

VGG-13 (SD) (Rao et al., 2021) 0.40 0.26 0.96

Wavelets (SD) (Agbley et al., 2020) 0.62 0.54 0.74

model to combine the advantages of each individual

model. In fact, as shown in Table 5, the ensemble

was able to provide: 1. improved performance with

respect to all individual models in subject dependent

experimental protocols; 2. satisfying performance in

subject independent experimental protocol.

Lastly, in case multimodal approaches are consid-

ered, where several clinical variables are fed to the

model, the work by Ahmed et al. (Fakhry et al., 2021)

comprises the state of art with F1 score 0.91, sensi-

tivity 0.85 and speciﬁcity 0.99, while considering a

subject independent protocol.

5 CONCLUSIONS

This article described a COVID-19 detection frame-

work based on an ensemble model combining the ad-

vantages of individual classiﬁers. After extensive ex-

periments considering the statistical dependency of

the available cough sound events, it was shown that

such a synergistic framework surpasses the state of

the art, and closes the gap existing between audio-

based and multimodal detection approaches. Impor-

tantly, the present experimental set-up is based on a

publicly available dataset, while the results are fully

reproducible and available at https://github.com/vin

cenzoconv99/.

We believe that the dataset-related considerations

expressed in this work are important to continue con-

ducting research on a problem of such critical im-

portance. In the future, as more data become avail-

able, we are going to experiment with deep learn-

ing based methods including automated feature learn-

ing and embedded prediction interpretability, which is

rapidly becoming a standard requirement in modern

AI-based tools and methodologies (European Com-

mission, 2020).

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

804

REFERENCES

Agbley, B. L. Y., Li, J., Haq, A. U., Cobbinah, B.,

Kulevome, D. K. B., Agbefu, P. A., and Eleeza, B.

(2020). Wavelet-based cough signal decomposition

for multimodal classiﬁcation. 2020 ICCWAMTIP,

pages 5–9.

Casiraghi, E., Malchiodi, D., Trucco, G., Frasca, M., Cap-

pelletti, L., Fontana, T., Esposito, A. A., Avola, E.,

Jachetti, A., Reese, J., Rizzi, A., Robinson, P. N.,

and Valentini, G. (2020). Explainable machine learn-

ing for early assessment of covid-19 risk prediction

in emergency departments. IEEE Access, 8:196299–

196325.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine Learning, 20(3):273–297.

Cozzatti, M., Simonetta, F., and Ntalampiras, S. (2022).

Variational autoencoders for anomaly detection in res-

piratory sounds. In Lecture Notes in Computer Sci-

ence, pages 333–345. Springer Nature Switzerland.

Dietterich, T. G. (2000). Ensemble methods in machine

learning. In Multiple Classiﬁer Systems, LBCS-1857,

pages 1–15. Springer.

Erdo

gan, Y. E. and Narin, A. (2021). COVID-19 de-

tection with traditional and deep features on cough

acoustic signals. Computers in Biology and Medicine,

136:104765.

European Commission (2020). White paper on artiﬁcial

intelligence: a european approach to excellence and

trust. Technical report, Brussels.

Fakhry, A., Jiang, X., Xiao, J., Chaudhari, G., Han, A., and

Khanzada, A. (2021). Virufy: A multi-branch deep

learning network for automated detection of COVID-

19. CoRR, abs/2103.01806.

Geurts, P., Ernst, D., and Wehenkel, L. (2006). Extremely

randomized trees. Machine Learning, 63(1):3–42.

Ho, T. K. (1998). The random subspace method for con-

structing decision forests. IEEE Transactions on

PAMI, 20(8):832–844.

Imran, A., Posokhova, I., Qureshi, H. N., Masood, U., Riaz,

M. S., Ali, K., John, C. N., Hussain, M. I., and Nabeel,

M. (2020). AI4covid-19: AI enabled preliminary di-

agnosis for COVID-19 from cough samples via an

app. Informatics in Medicine Unlocked, 20:100378.

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W.,

Ye, Q., and Liu, T.-Y. (2017). Lightgbm: A highly ef-

ﬁcient gradient boosting decision tree. NIPS’17, page

3149–3157, Red Hook, NY, USA. Curran Associates

Inc.

Lippi, G., Mattiuzzi, C., Bovo, C., and Plebani, M. (2020).

Current laboratory diagnostics of coronavirus disease

2019 (covid-19). Acta Bio Medica Atenei Parmensis,

91(2):137–145.

Nessiem, M. A., Mohamed, M. M., Coppock, H., Gaskell,

A., and Schuller, B. W. (2021). Detecting covid-19

from breathing and coughing sounds using deep neu-

ral networks. In 2021 IEEE CBMS, pages 183–188.

Ning, W., Lei, S., Yang, J., Cao, Y., Jiang, P., Yang, Q.,

Zhang, J., Wang, X., Chen, F., Geng, Z., Xiong, L.,

Zhou, H., Guo, Y., Zeng, Y., Shi, H., Wang, L., Xue,

Y., and Wang, Z. (2020). Open resource of clinical

data from patients with pneumonia for the prediction

of COVID-19 outcomes via deep learning. Nature

Biomedical Engineering, 4(12):1197–1207.

Ntalampiras, S. (2020a). Collaborative framework for auto-

matic classiﬁcation of respiratory sounds. IET Signal

Processing, 14(4):223–228.

Ntalampiras, S. (2020b). Identiﬁcation of anomalous

phonocardiograms based on universal probabilistic

modeling. IEEE Letters of the Computer Society,

3(2):50–53.

Ntalampiras, S. and Potamitis, I. (2019). Classiﬁcation of

sounds indicative of respiratory diseases. In Engineer-

ing Applications of Neural Networks, pages 93–103.

Springer International Publishing.

Orlandic, L., Teijeiro, T., and Atienza, D. (2021). The

COUGHVID crowdsourcing dataset, a corpus for the

study of large-scale cough analysis algorithms. Scien-

tiﬁc Data, 8(1):3–2.

Poir

e, A. M., Simonetta, F., and Ntalampiras, S. (2022).

Deep feature learning for medical acoustics. In

Lecture Notes in Computer Science, pages 39–50.

Springer Nature Switzerland.

Rao, S., Narayanaswamy, V., Esposito, M., Thiagarajan, J.,

and Spanias, A. (2021). Deep learning with hyper-

parameter tuning for covid-19 cough detection. In

2021 IISA, pages 1–5.

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986).

Learning Representations by Back-propagating Er-

rors. Nature, 323(6088):533–536.

Tena, A., Clari

a, F., and Solsona, F. (2022). Automated de-

tection of COVID-19 cough. Biomedical Signal Pro-

cessing and Control, 71:103175.

Xia, T., Spathis, D., Brown, C., Ch, J., Grammenos, A.,

Han, J., Hasthanasombat, A., Bondareva, E., Dang, T.,

Floto, A., Cicuta, P., and Mascolo, C. (2021). COVID-

19 sounds: A large-scale audio dataset for digital

COVID-19 detection. In Thirty-ﬁfth Conference on

Neural Information Processing Systems Datasets and

Benchmarks Track (Round 2).

Ensemble Learning for Cough-Based Subject-Independent COVID-19 Detection

805