Professional Documents
Culture Documents
Amaral Et Al 2020
Amaral Et Al 2020
Amaral Et Al 2020
https://doi.org/10.1007/s11517-020-02240-7
ORIGINAL ARTICLE
Abstract
To design machine learning classifiers to facilitate the clinical use and increase the accuracy of the forced oscillation technique (FOT) in
the differential diagnosis of patients with asthma and restrictive respiratory diseases. FOT and spirometric exams were performed in 97
individuals, including controls (n = 20), asthmatic patients (n = 38), and restrictive (n = 39) patients. The first experiment of this study
showed that the best FOT parameter was the resonance frequency, providing moderate accuracy (AUC = 0.87). In the second exper-
iment, a neuro-fuzzy classifier and different supervised machine learning techniques were investigated, including k-nearest neighbors,
random forests, AdaBoost with decision trees, and support vector machines with a radial basis kernel. All classifiers achieved high
accuracy (AUC ≥ 0.9) in the differentiation between patient groups. In the third and fourth experiments, the use of different feature
selection techniques allowed us to achieve high accuracy with only three FOT parameters. In addition, the neuro-fuzzy classifier also
provided rules to explain the classification. Neuro-fuzzy and machine learning classifiers can aid in the differential diagnosis of patients
with asthma and restrictive respiratory diseases. They can assist clinicians as a support system providing accurate diagnostic options.
Keywords Clinical decision support system . Forced oscillation technique . Diagnostic of respiratory diseases . Respiratory
oscillometry . Differential diagnosis
COPD Chronic obstructive pulmonary disease, a lung dis- majority of the class’s output by the individual de-
ease characterized by chronic obstruction of lung cision trees
airflow that interferes with normal breathing Rm Mean resistance in the 4–16 Hz range, reflecting
FOT Forced oscillation technique, a method to evaluate mid-frequency spectra, that is related to the resis-
respiratory mechanics using sinusoidal system iden- tance in the central airways, expressed as cmH2O/
tification techniques L/s
frA Fuzzy set (Gaussian membership function) related ROC Receiver operating characteristic curve
to resonant frequency fr for Asthma class Rrs Respiratory resistance, including airways, lung, and
FEV1 The forced expiratory volume in the first second, thoracic wall resistance, expressed as cmH2O/L/s
obtained from a maximal expiratory effort maneu- S Angular coefficient of resistance, the resistance
ver, expressed in L change with frequency in the 4–16 Hz range, which
FVC Forced vital capacity, the total amount of air ex- is associated with respiratory nonhomogeneities,
haled during the espirometric exams, expressed in L expressed as cmH2O/L/s2
fr Resonant frequency, the frequency at which Xrs Se Sensitivity, proportion of actual positives that are
becomes zero, associated with respiratory inhomo- correctly identified as such
geneity and expressed as Hz Sm Width matrix of the Gaussian membership
frR Fuzzy set (Gaussian membership function) related functions
to resonant frequency fr for Restrictive class Sp Specificity, proportion of actual negatives that are
K Number of nearest neighbor correctly identified as such
k Number of folds in k-fold validation procedure SSCG Speeding up Scaled Conjugate Gradient
KNN K-nearest neighbor. It is a ML algorithm. When it is SVM Support vector machines. It is a ML algorithm that
used for classification, the class of the query is de- uses support vectors to determine a decision bound-
termined by the majority vote among the class K- ary that is a hyperplane with optimal geometric mar-
nearest neighbors found in the training set gin from the classes, which, in turn, presents the
KS Kyphoscoliosis, an abnormal curvature of the spine highest generalization capacity
in both a coronal and sagittal plane. It is a combina- SVMR support vector machines with radial basis function
tion of kyphosis and scoliosis kernel. It is a SVM that employs the “kernel trick”
m Number of rules to allow SVM to be employed in nonlinear separa-
ML Machine learning, field of artificial intelligence ble problems. The “kernel trick” transforms the data
whose scope is the investigation of algorithms that in a new high-dimensional space where it is easier to
can recognize patterns and learn different relation- separate the classes
ships present in a set of data U Center matrix of the Gaussian membership
n Number of features (variables) functions
NFC Neuro-fuzzy classifier. It is a fuzzy rule-based sys- W The weight matrix among the rules and the classes
tem, which is encoded as a neural network. Hence, X4 Respiratory reactance at 4 Hz, expressed as cmH2O/
it is possible to apply neural network learning algo- L/s
rithms to determine the parameters of the fuzzy sys- Xm Mean reactance evaluated considering the 4 to
tems, such as the fuzzy rules and fuzzy membership 32 Hz frequency range, associated with respiratory
functions inhomogeneity and expressed as cmH2O/L/s
r Standard deviation of the radial basis function Xrs Respiratory reactance, including airways, lung, and
R0 Intercept resistance, obtained using linear regres- thoracic wall, expressed as cmH2O/L/s
sion in the 4–16 Hz range, a representative of the xj jth input variable in the fuzzy classifier
resistance in the low-frequency spectra, expressed Z4 Impedance module in 4 Hz, The total mechanical
as cmH2O/L/s load including resistance and elastic effects,
R0A Fuzzy set (Gaussian membership function) related expressed as cmH2O/L/s
to intercept resistance R0 for Asthma class
R0R Fuzzy set (Gaussian membership function) related
to intercept resistance R0 for Restrictive class 1 Introduction
R4 Respiratory resistance in 4 Hz, expressed as
cmH2O/L/s Reduction of the maximum air flow of the lungs is understood
RF Random forests. It is a ML algorithm that employs as an obstructive disease [37]. Asthma is a particular case of
an ensemble of decision trees. In classification prob- obstructive disease, a global health problem affecting 1–18%
lems, the final output class is obtained by the of the population in different countries [95]. Restrictive
Med Biol Eng Comput
disease is defined by the restriction of lung expansion due to FOT is rapidly becoming a key instrument in pulmonary
parenchymal, pleural, chest wall, or neuromuscular apparatus function analysis. However, despite the advantages of FOT in
changes. These disorders are characterized by a reduced vital terms of its noninvasiveness and lack of dependence on pa-
capacity and lung volume at rest, but with normal resistance. tient cooperation, the clinical use of this method is limited
Although asthma and restrictive diseases present similar clin- because, in the context of a diagnostic framework, the inter-
ical implications, including shortness of breath, severe cough, pretation of resistance and reactance parameters demands
and chest pain, they have different treatments for each condi- training and experience, and it is not a simple task for the
tion, requiring a special differential analysis [85]. untrained pulmonary specialist. To contribute to minimizing
Airway obstruction and lung expansion properties are usu- this limitation, prior studies from our laboratory have provided
ally indirectly evaluated by measuring the maximum expired clear evidence that machine learning (ML) methods may sim-
airflow and volume using spirometry. Whole body plethys- plify the routine assessment of lung function by FOT. Another
mography is another useful technique to measure airway ob- important characteristic that emerged from these studies was a
struction and lung restriction. Despite long-term clinical suc- significant improvement in the diagnostic accuracy of FOT in
cess, these methods have a number of problems in use. They several important clinical respiratory conditions [6–8, 68, 69].
require a high degree of collaboration and maximal effort, and In line with these observations, other methods of lung function
thus, these measurements may be unreliable and variable if analysis were also improved by the use of ML techniques [9].
suboptimal maneuvers are performed. Furthermore, forced Although previous studies have shown that these methods
maneuvers may alter the bronchial tone and modify the airway could help the clinical use of the FOT, simplifying test inter-
patency, rendering the obtained indices hardly physiologic. pretation and increasing accuracy, there are no previous stud-
Accordingly, the literature has emphasized that investigation ies using FOT combined with ML algorithms to improve the
on new methods to improve the noninvasive tests of pulmo- differential diagnoses of asthma and restrictive diseases.
nary function should be considered a great preeminence [13, Therefore, we have explored the hypothesis that the appli-
22, 30]. cation of ML algorithms combined with FOT analysis would
The forced oscillation technique (FOT) is a noninva- improve the differential diagnosis of asthma and restrictive
sive method that was introduced in the 1950s by Dubois respiratory diseases. In this context, the specific goals of this
et al. [31] to evaluate lung function. The method evaluates study were (1) to evaluate the ability of FOT parameters alone
respiratory impedance and its associated elements: respi- to correctly diagnose differences between patient groups
ratory resistance and reactance [55]. Owing to a number (asthma and restrictive); (2) to evaluate several ML methods
of recent technological improvements, this method nowa- to help in the differential diagnosis of these respiratory dis-
days represents the state of the art in lung function eval- eases; and (3) to explore whether a fuzzy classifier could aid in
uation [11]. Several authors have argued that it has the the differential diagnosis and provide a useful explanation
potential to improve diagnosis and monitor the treatment concerning how the classification is performed.
of respiratory diseases and that further studies are needed
in this area [16, 63]. Previous studies conducted by our
group suggest that this technique can contribute to the 2 Methods
detection of obstructive respiratory changes in patients
with asthma [18], and restrictive abnormalities in sarcoid- 2.1 Ethical issues, studied subjects, and inclusion and
osis [34], silicosis [25], rheumatoid arthritis [33], and sys- exclusion criteria
temic sclerosis [71].
Information obtained from FOT provided a detailed evalu- This research was approved by the research ethics board of the
ation of the respiratory abnormalities in asthma [18] and re- State University of Rio de Janeiro, and the post-informed con-
strictive disease [34]. Therefore, this method may contribute sent of all volunteers was obtained before inclusion in the
to the differential diagnosis of the respiratory abnormalities in study. The study was conducted in accordance with the
asthma and restrictive diseases. It was pointed out previously, Declaration of Helsinki. Anthropometric information was ob-
however, that identification of specific characteristics of re- tained from the volunteers before the beginning of the
strictive diseases by FOT analysis has not yet been fully pre- procedures.
sented [58], and only two recent studies have suggested that Ninety-seven volunteers were selected for the study, in-
FOT may be useful in the differential diagnosis between re- cluding 38 diagnosed with asthma and 39 with restrictive re-
strictive and obstructive respiratory abnormalities [48, 90]. spiratory disease. Among asthmatics, eighteen had severe ob-
Initial results obtained in our group provided additional evi- struction and twenty moderate obstruction. The group with
dence in favor of this hypothesis [86]. However, FOT was restrictive disease was composed of patients with sarcoidosis,
able to provide adequate clinical diagnosis only between scleroderma, idiopathic pulmonary fibrosis, asbestosis, and
groups presenting diseases in latter stages. silicosis, of whom twenty patients presented severe restriction
Med Biol Eng Comput
and eighteen presented moderate restriction. The inclusion respiratory nonhomogeneities [15]. The mean resistance (Rm)
criteria were as follows: age over 18 years; the clinical diag- was calculated in the 4–16 Hz range and reflects mid-
nosis of the respective diseases for patients with asthma and frequency spectra, which is related to the resistance in the
restrictive abnormalities; exclusion of history of smoking and central airways [64].
other cardiovascular or respiratory diseases. Volunteers with Three parameters were used to describe the results related
asthma during a crisis period were not included in the study. to reactance: resonant frequency (fr), mean reactance (Xm),
Although the objective of this study was to evaluate the and dynamic compliance (Cdyn). Resonant frequency is the
potential of FOT in the differential diagnosis of asthma and frequency, at which Xrs becomes zero [18], obtained by the
restrictive respiratory diseases, a control group (CG) was also interpolation of respiratory reactance (Xrs) values adjacent to
included to allow a deeper comprehension of the biomechan- the ones at which this variable changes from negative to pos-
ical abnormalities in asthma and restrictive diseases, as well as itive values. Mean reactance is associated with respiratory
the changes with the progression of these diseases. This group inhomogeneity and was evaluated considering the 4 to
was composed of 20 healthy subjects with FOT values within 32 Hz frequency range. Dynamic compliance, which was ob-
normal limits, with no history of tobacco use, as well as car- tained using the reactance at 4 Hz (Cdyn = 1/2πfX4), reflects
diac or pulmonary disease. the total compliance of the respiratory system, also being re-
lated to the respiratory homogeneity [64]. The total mechani-
2.2 Pulmonary function cal load, including resistance and elastic effects, was also in-
vestigated using the 4 Hz impedance module (Z4) [74].
FOT exams were conducted using an instrument developed in Spirometric exams were performed after FOT, following
our laboratory [24] that follows international standards [56]. local [54] and international [70] recommendations. The clas-
The measurements were conducted during spontaneous sifications of airway obstruction and pulmonary restriction
breathing using small peak-to-peak pressure oscillations were also performed following these guidelines, using spiro-
(2 cmH2O) generated by a speaker, which were applied at metric exams as a gold standard. The analyzed indexes were
the entrance of the individual’s airway through the oral cavity. forced expiratory volume in the first second (FEV1), forced
The nostrils of the volunteers were occluded with a nose clip. vital capacity (FVC), and FEV1/FVC ratio, which were
To minimize the shunt effect of the upper airways [8], the expressed as absolute values and as percentages of predicted
volunteer firmly held their cheeks and chin with their hands. values [79]. The GINA guidelines [95] were used to define
Three tests of 16 s were performed, and the result adopted was asthma.
the mean score. This measurement duration is an appropriate
compromise, achieving clinically acceptable statistical vari-
ability and a comfortable examination time for the patient. 2.3 Presentation of results and statistical analysis
The test was considerable acceptable if the volunteers present-
ed stable tidal volumes and rate and free of pauses. Common The results are presented as the means ± standard deviations.
artifacts such as swallows, coughs, and leaks were identified Statistical analysis was performed using the ORIGIN 8.0 pro-
by the evaluation of flow and pressure signals, and the acqui- gram (Microcal Software Inc., Northampton, MA, USA). The
sition was repeated until three stable and free of artifact mea- characteristics of the sample’s distribution were initially eval-
surements were obtained. Frequencies between 4 and 32 Hz uated (Shapiro-Wilk test). Then, a parametric test (indepen-
were used in these exams. To reduce the influence of the dent t test) was used when the data exhibited a normal distri-
spontaneous breathing, only exams with coherence function bution, and a nonparametric test (Mann-Whitney) was used
≥ 0.9 in the whole frequency range studied were accepted. To when the data do not presented a normal distribution. A p
exclude outlying values, the coefficient of variation of respi- value of less than 0.05 was considered significant for the re-
ratory resistance (Rrs) at the lowest oscillation frequency sults of all statistical analysis.
(4 Hz) for the 3 measurements was ≤ 10%. The analyses were Receiver operating characteristic (ROC) curves were plot-
performed using an instrument developed at our laboratory ted to analyze the clinical use and cutoff points of various FOT
and described previously [24]. parameters in the discrimination between asthma and restric-
The resistance results were interpreted using linear regres- tive respiratory abnormalities. Diagnostic accuracy was
sion in the 4–16 Hz range, which allowed us to obtain the assessed calculating the area under the curve (AUC).
intercept resistance (R0). This is a representative of the resis- Previous studies suggested that ROC curves with an AUC
tance in the low-frequency spectra, describing Newtonian re- between 0.80 and 0.90 are adequate for clinical use, while a
sistance of the respiratory system, as well as the effect of high diagnostic accuracy is observed in conditions of AUCs ≥
pendelluft (gas redistribution) [60]. The regression also pro- 0.90 [41, 91]. This accuracy allows the clinicians to easily
vided the angular coefficient of resistance (S), which depicts balance the values of sensitivity and specificity to the specific
the resistance change with frequency, which is associated with local conditions of use and needs. Diagnostic accuracy
Med Biol Eng Comput
analyses were performed using MedCalc 12 (MedCalc theory of fuzzy systems with computational intelligence tech-
Software, Mariakerke, Belgium). niques such as neural networks and evolutionary algorithms.
In the neuro-fuzzy methods [2, 52, 75], one possible strategy
2.4 Datasets is to code the fuzzy system as a neural network and to apply
established methods of training, such as backpropagation [46].
In the present work, experiments were performed in a dataset, When the strategy employs evolutionary algorithms, then the
which included 7 input features (FOT indexes) from 241 genetic algorithms are the most used. They provide a way of
exams. It included 114 measurements taken from volunteers codifying and evolving the following fuzzy blocks: member-
with asthma: 60 with moderate asthma and 54 with severe ship functions, aggregation operators, different rules compo-
asthma. It also has 117 measurements taken from patients with sitions, and defuzzification operators.
restrictive diseases: 60 with severe and 57 with moderate re- To further explore other ML methods, we also selected the
strictive diseases. fuzzy classifier because we also wanted to address the inter-
pretability, which is the ability to express the behavior of the
2.5 Machine learning algorithms real system in a compressive manner. It is an individual prop-
erty and is usually associated with several factors related to the
Machine learning is the field of artificial intelligence whose structure of the model, such as the number of input variables,
scope is the investigation of algorithms that can recognize number of rules, number of linguistic terms, and others [39].
patterns and learn different relationships present in a set of Therefore, we would like to answer the following inquiry: Is it
data [94]. In our previous studies [5, 7, 8], we have appraised possible to develop a classifier that presents interpretability
a wide variety of models including logistic linear classifier, and, at the same time, also have a satisfactory accuracy?
decision trees, neural networks, k-nearest neighbors, support As a result, in the present work, the chosen classification
vector machines (SVMs), and ensemble strategies such as methods were analyzed:
random forests. It was noted that the AdaBoost classifier, k-
nearest neighbor, random forest classifier, and SVM with ra- & K-nearest neighbor [45, 57];
dial basis kernel had presented outstanding performance. Our & Ensemble strategies (AdaBoost classifier with decision
incipient exploratory experiments confirm this fact; based on trees [38, 87] and random forest (RF) [14, 88]);
these observations, we concluded that we conserve only k- & SVM with radial basis functions [1, 44];
nearest neighbor, AdaBoost, SVM with radial basis kernel, & Neuro-fuzzy classifier [20, 47, 52].
and the random forest classifier. However, these models only
address accuracy, which is the capability of the model to rep- A description of the ML techniques previously used in
licate the real system’s results. It should be more significant as pulmonary function exams was presented elsewhere [9]. The
there is higher conformity between the responses of the real studied methods will be succinctly presented here, and a thor-
system and the model [39]. However, when one wishes to ough description is available in the References.
understand how the induced model can distinguish between KNN is a lazy learner because, in the training stage, it does
different classes or represent relations existing in the data in a not learn the relationships between the data in the training set,
comprehensible way, more symbolic approaches, such as it merely stores the training set (a set of labeled instances).
rule-based systems, become more attractive. In addition to When a new query must be classified, the class of the query is
the ability to express knowledge in a comprehensible way, obtained by using majority vote among the class of the K
they enable the introduction of the specialist’s knowledge. objects. Random forest (RF) is an ensemble strategy that as-
The fuzzy set theory [96] is one of the most important para- sembles and compounds several base decision trees [14]. It
digms of computational intelligence that explores aspects of employs the bootstrap aggregation (bagging) which helps to
inference and knowledge representation. The greatest motiva- improve accuracy and control the overfitting [45]. AdaBoost
tion for using the fuzzy set theory was to establish an interface employs a distinct ensemble strategy, called boosting, where
between quantitative patterns and qualitative knowledge the user can join several “weak classifiers” (base estimators)
structures that represent vague and imprecise information for- together to form a single “strong classifier” [87]. Each base
mulated in natural language. This feature allows the represen- estimator (decision tree) is designed to classify the instances
tation of the knowledge extracted from the database in linguis- misclassified by previous base estimators correctly. Once the
tic form, generating greater interpretability [47]. training process is terminated, the algorithm associated with
The most recurrent application of fuzzy set theory is the each base estimator, a weight related to its accuracy. The final
synthesis or inference of fuzzy rule-based systems (FRBS), output is a weighted combination of all base estimators (deci-
where several strategies have been developed to induce sion trees).
rules-based fuzzy models [21]. It is remarkably vital in the The basic principles from which support vector machines
field of learning fuzzy rules, the hybrid methods that join the (SVMs) were conceived were established by statistical
Med Biol Eng Comput
learning theory [93]. SVMs were employed in a myriad of and second-order methods as Levenberg–Marquardt [51]. In this
problems with the state-of-the-art performance [62, 65, 66]. work, we chose the approach described in [20], which employs a
Considering a binary, linearly separable classification prob- Speeding up Scaled Conjugate Gradient (SSCG) to optimize the
lem, SVM provides a decision boundary that is hyperplane parameters of the neuro-fuzzy classifier. The SSCG is a variation
with optimal geometric margin from the classes, which in turn of the Scaled Conjugate Gradient, which shortens the optimiza-
presents the highest generalization capacity. This conception tion time without compromising the convergence rate.
can be extended to a nonlinear separable problem by applying
an artifice called a “kernel trick.” This scheme transforms the 2.6 Performance analysis
data into a new high-dimensional space, where one expects
the classes to be effortlessly separable [45]. Generalization is a very important characteristic of a classifier
Fuzzy classification is the procedure of segmenting feature [72]. It is the faculty of the classifier to encounter a suitable
spaces into fuzzy classes. It assumes the frontier between two class estimate for new and unseen data; that is, data was not
neighbor classes as a continuous, super-imposed area within used to the training procedure. One key issue to carry out the
an object has partial membership in each class. A fuzzy clas- assessment of the generalization ability is the election of prop-
sifier can be characterized as a set of fuzzy classification rules er evaluation criteria [3, 94]. Since our research deals with the
Ri which depicts the relation between the input feature space medical diagnosis, we opt to embrace sensitivity (Se), speci-
and the classes, which can be expressed as follows: ficity (Sp), and the area under the receiver operating charac-
teristic (ROC) curve (AUC) [43]. Additionally, this choice
Ri : If x1 is Ai1 and…and xj is Aij and…and xn is Ain
grants permission to contrast and correlate our findings with
Then class is C k other previous studies carried out by our group [5, 7, 8, 69].
After the selection of the performance assessment criteria, it
where xj stands for the jth feature or input variable; Ck
was needed to design a proper evaluation structure to measure
represents the kth label of class; n represents the number of the performance of the trained model based on unseen exam-
features; and Aij indicates the fuzzy set of the jth feature in the
ples to infer its generalization ability. Since the dataset has a
ith rule and is depicted by the pertinent membership function modest size, k-fold validation procedure [53] is a more suit-
[51]. The fuzzy classifier employed in this work is like Jang’s able choice because it allows the valuation of the generaliza-
classifier and it is displayed in Fig. 1.
tion capability in the whole dataset. We subdivided the dataset
The feature space in Fig. 1 has two features {x1, x2}, and into k equal (or approximately equal) data subsets or folders,
the classifier discriminates them into two classes {C1, C2}. ensuring the same class proportional to each folder. Each sub-
The network structure of the fuzzy classifier is similar to the
set is utilized for testing, and the training of the classifier
one presented in the study by Jan and his collaborators [51]. model employs the remaining k − 1 subsets. The performance
However, every feature is expressed with two fuzzy sets. As a of each algorithm on each folder can be calculated. Upon the
result, there are four fuzzy rules.
end of the k-fold validation, k outcomes of the performance
The parameters of the fuzzy classifier θ = {Umxn Smxn Wmxn} metric are available, and they can be aggregated to produce a
can be fit by supervised training algorithms, where U and S are final estimation of the generalization capability of the model.
the center and the width matrices of the Gaussian membership
This aggregation in the k-fold validation is an essential feature
functions, respectively; W represents the weight matrix among because it circumvents the reporting of an idealist result
the rules and the classes; and m, n, and c are the number of rules, achieved from a single, particular division of the dataset in
features, and classes, respectively. The adaptive neuro-fuzzy net-
the training and test sets, as it could occur in holdout proce-
works have also been trained using different optimization dures. Furthermore, it is also possible to use these outcomes to
methods, such as the Kalman filter [50], gradient descent [52], contrast various ML methods using statistical hypothesis
testing, which is a vital element to compare two or more ML known to be unsusceptible to overfitting [87]. Schapire argues
algorithms. McNemar’s test and Wilcoxon’s signed-rank test that the increase in confidence of the predictions with additional
are endorsed by Dietrich [29], Demsar [28], and Japkowicz cycles of boosting is responsible for achieving improved gener-
and Shah [53]. Another methodology commonly employed is alization. Besides, the max depth of the base estimators was
the comparison of AUCs, which was performed as described controlled to avoid the overfitting of the base estimators. In the
in Delong et al. [27]. k-nearest neighbor classifier, the number of neighbors was al-
ways chosen to be higher than 1; as a result, the classification
2.7 Experimental analysis is based on a set of nearest neighbors, which turns the classifier
more resistant to outliers. The RF is an ensemble strategy that
Four experiments were conducted. Firstly, we figured out the aggregates the outcomes of several trees, which tends to mitigate
ability of each one of the FOT indexes alone to accurately set overfitting errors; also, similarly to AdaBoost, the max depth of
apart patients with asthma and with restrictive respiratory dis- the base estimator was also controlled, diminishing the possibil-
eases (experiment 1). ity of overfitting. The SVMR is built to obtain the optimal deci-
Secondly, ML algorithms were exploited to determine sion frontier between classes to ensure a higher generalization
whether a boost in performance could be accomplished. We ability. The fuzzy partition applied in the neuro-fuzzy classifier
did not carry out any feature selection; hence, all the original turns it resistant to small fluctuations in training, which could
FOT indexes were used. K-nearest neighbors, AdaBoost, ran- lead to variance errors. The training procedures can also adopt
dom forests, and SVMs classifiers were implemented with a the same strategies employed to avoid overfitting in neural net-
Python machine learning library Scikit-learn [78]. This library works, such as early stopping [10].
allows the user to tune several classifiers’ hyperparameters. Also, the use of nested cross-validation procedure [19] was
Here, AdaBoost employed the decision tree as its base estima- enforced to find the best classifier hyperparameters and to
tor (“weak classifier”), and the adjusted parameters were the provide extra protection against overfitting.
number of base estimators (50, 100, 150, 300) and the max The third experiment followed the same methodology of
depth of the base estimator (4, 5, 6, 7, 8). In the KNN, K (3, 5, experiment 2, but in the processing pipeline, a feature selec-
7) was the tuned hyperparameter. In the RF method, the tion procedure was included. The utility of the input feature
hyperparameters were the number of base estimators (50, selection is to encounter the smallest number of pertinent and
100) and the max depth of the base estimator (4, 5, 6, 7). informative features that can result in adequate performance.
The SVM with an RBF kernel had two hyperparameters, the Our primary motivation to perform feature selection is to al-
regularization parameter C (1, 10, 100, 1000) and the standard low data visualization (2D or 3D) [42]. The chosen feature
deviation of the radial basis function r (0.001, 0.0001). The selection method was the recursive feature elimination, which
investigation for the best hyperparameters was executed in the is a backward search [42], which recursively removes attri-
training routine for each classifier. Because of the use of 10- butes and builds a model on those features that remain. The
fold cross-validation, the training was reiterated ten times; chosen model was a linear support vector machine. It is a low
each training routine selects one fold as a test set and the complexity model (linear) with the ability to produce the hy-
remaining ones as a training set. A second 10-fold cross-val- perplane to maximize the margin and, as a result, to provide
idation (also called inner, or nested), which employed only the better generalization. Also, the low complexity model pro-
training set, was carried to completion to spot suitable vides protection against overfitting in the choice of the feature,
hyperparameters for each classifier. The performance metric which may happen if a high complexity model was applied.
was average AUC. The fourth experiment also employed feature selection, but
We also implemented in MATLAB a neuro-fuzzy classifi- it used a method stability selection [76]. It is a relatively new
er based on the source code provided by Cetişli and Barkana method for feature selection, where the basic idea is to apply a
[20]. It has, as its main hyperparameter, the number of clusters feature selection algorithm on distinct subsets of data and with
per class, which is used by K-means [49, 51] to obtain the diverse subsets of features. Once the process is rerun numer-
initial parameters (centers and widths of the Gaussian mem- ous times, the selection results can be accounted for, for ex-
bership functions) and to formulate the fuzzy rules. Later, ample, by checking how many times a feature was elected as
those parameters are optimized by a modified version of the necessary when it was in a scrutinized feature subset.
scaled conjugate gradient algorithm [73]. Also, we add an Therefore, strong features would have scores close to 100%
early stopping procedure, which is known to reduce (since they would always be selected when possible), weaker,
overfitting and improve generalization [81]. but somewhat essential features would have non-zero scores,
Overfitting can be responsible for poor predictive perfor- and irrelevant features would have a score close to 0.
mance in unseen data when a model has not acquired the ability For all experiments in the study, the ROC curve, obtained
to generalize. Therefore, to circumvent overfitting, we thorough- using only the best individual FOT parameter (BFP) as an input
ly manage the classifier complexity and model. AdaBoost is feature, was used to confront the performance of the classifiers.
Med Biol Eng Comput
Resistance (hPa.s.l )
-1
4.0
FEV1 [70]. For these experiments, the comparative analysis be- 3.5
tween the classifiers was executed using MedCalc 8.2 (Medicalc 3.0
Software, Mariakerke, Belgium) [27]. 2.5
2.0
1.5
1.0
0.5
3 Results 0.0
4 8 12 16 20 24 28 32
Regarding the biometric characteristics of the studied individ- Frequency (Hz)
uals (Table 1), no significant differences were observed in
b
terms of weight and height among groups (p > 0.05). The 2
control group produced higher FEV1 and FVC values (within
normal limits). As expected, the spirometric parameters were
0
reduced in the patient groups and FVC was reduced in patients
Reactance (hPa.s.l )
-1
with restrictive diseases in comparison with the asthma group,
while FEV1/FVC were higher in patients with restrictive dis- -2
a b
18 18
p<0.001
15 15
p<0.001
12 12
Rm (hPa.s.l )
R0 (hPa.s.l )
-1
-1
9 * 9
6 6 *
*
*
3 3
0 0
Control Restrictive Asthma Control Restrictive Asthma
c d
200 2
* * 0 * *
0
Xm (cmH2O/L/s)
S (hPa.s .l )
2 -1
-200 -2
-400 -4
-600 -6
p<0.001
p<0.001
-800 -8
Control Restrictive Asthma Control Restrictive Asthma
e 60
f
0.18
50 0.15
p<0.001
40 ns
0.12
*
Cdyn (l/hPa)
fr (Hz)
30 0.09
*
20 0.06
*
10 0.03
*
0 0.00
Control Restrictive Obstructive Control Restrictive Asthma
g
24
p<0.01
20
16
Z4 (hPa.s.l )
-1
12 *
8
*
0
Control Restrictive Asthma
Med Biol Eng Comput
The resonance frequency (fr) was the best FOT parameter Table 2 Results of experiment 1. The area under the ROC curve
(AUC), the standard error (SE), and the 95% confidence interval (95%
and it provided moderate accuracy diagnoses (0.7 ≤ AUC <
CI) of each FOT parameter
0.9).
AUC SE 95% CI
Fig. 5 ROC curves of the best FOT parameter (BFP), and the FOT
Fig. 4 ROC curves for the first experiment, describing the diagnostic parameters associated with the ML algorithms and the neuro-fuzzy clas-
performance of each FOT parameter sifier (NFC)
Med Biol Eng Comput
Table 3 Results of
experiment 2. Best Se Sp AUC
studied classifiers with (%) (%)
the original FOT
parameters as inputs. The BFP 89.7 77.2 0.87
95% confidence interval ADAB 87.2 96.5 0.97
is shown in parenthesis
below each performance KNN 92.3 90.4 0.95
metric. The AUC RF 92.3 89.5 0.97
standard error is also SVMR 99.1 80.7 0.94
shown in parenthesis
NFC 88.9 78.1 0.90
BFP 0.103 ± 0.027** 0.083 ± 0.028** 0.095 ± 0.027** 0.067 ± 0.029* 0.025 ± 0.026
ADAB - 0.020 ± 0.009* 0.008 ± 0.008 0.035 ± 0.016* 0.078 ± 0.022**
KNN - - 0.012 ± 0.010 0.016 ± 0.016 0.058 ± 0.025*
RF - - - 0.028 ± 0.013* 0.070 ± 0.023**
SVMR - - - - 0.042 ± 0.028
BFP, best FOT parameter (obtained without the use of classifiers); ADAB, AdaBoost with decision tree classifiers; KNN, K-nearest neighbor (K = 1); RF,
random forests; SVMR, support vector machines with radial basis kernel; NFC, neuro-fuzzy classifier
*p < 0.05 (Delong et al. [25])
**p < 0.01 (Delong et al. [25])
Med Biol Eng Comput
Table 5 Results of
experiment 3. Best Se Sp AUC
studied classifiers with (%) (%)
the original FOT
parameters as inputs. The BFP 89.7 77.2 0.87
95% confidence interval ADAB 92.3 91.2 0.96
is shown in parenthesis
below each performance KNN 93.2 82.5 0.94
metric. The AUC RF 95.7 84.2 0.96
standard error is also SVMR 99.1 82.5 0.93
shown in parenthesis
NFC 89.7 79.8 0.89
mean FOT curves in the asthmatic and restrictive groups were that integrates resistive and elastic effects, presented increased
clearly distinct (Fig. 2). The resistance values in the group of changes in patients with asthma compared with patients with
asthmatic patients were higher than those observed in the restrictive diseases. These results further support two hypoth-
groups of restrictive patients, which, in turn, were higher than esis: (1) that asthma introduces increased effects in the resis-
those observed in controls (Fig. 2a). Perhaps the most impor- tance values in comparison with restrictive diseases and (2)
tant finding was that the reactance values were more influ- that the effect of the small airway obstruction on dynamic
enced by asthma than by restrictive abnormalities (Fig. 2b). compliance [26] may be stronger than the effects of elastic
This unanticipated finding may be explained, at least in part, changes observed in restrictive diseases.
considering that the effect of the small airway obstruction on Among the historiography of FOT, several studies have
dynamic compliance [26] and inertance [59], which are pres- described that this method may provide a sensitive analysis
ent in obstructive diseases, in the respiratory reactance may be for detecting abnormalities in respiratory mechanics [35, 36,
stronger than the effects of elastic changes, typical in restric- 71, 77, 80, 82–84]. The results of this study provide additional
tive diseases. evidence of this hypothesis (Table 2) and extend these finding
Accordingly, FOT parameters were distinct in these groups to the differential diagnosis of patients with asthma and re-
(Fig. 3). The current study found that resistive parameters strictive diseases.
were higher in asthmatic patients (Fig. 3 a, b, and c), which In the first experiment, we investigated the ability of each
is consistent with the characteristics of the studied groups. The of the FOT indexes alone to adequately discern patients with
changes observed in Xm (Fig. 3d), fr (Fig. 3e) and Cdyn (Fig. asthma and restrictive (Table 2). The best FOT index was fr,
3f) reflect the changes in the mean reactance curves (Fig. 2b) which had an AUC equal to 0.87. None of the FOT parameters
and can be explained by the same observations described pre- alone allowed a highly accurate discrimination among patients
viously. Another important finding was that Z4, a parameter with asthma and restrictive abnormalities.
BFP 0.090 ± 0.028** 0.073 ± 0.028* 0.087 ± 0.029** 0.059 ± 0.031 0.023 ± 0.026
ADAB - 0.017 ± 0.009 0.003 ± 0.008 0.031 ± 0.017 0.067 ± 0.026**
KNN - - 0.013 ± 0.010 0.015 ± 0.015 0.050 ± 0.027
RF - - - 0.0280 ± 0.016 0.064 ± 0.027*
SVMR - - - - 0.036 ± 0.027
BFP, best FOT parameter (obtained without the use of classifiers); ADAB, AdaBoost with decision tree classifiers; KNN, K-nearest neighbor (K = 1); RF,
random forests; SVMR, support vector machines with radial basis kernel; NFC, neuro-fuzzy classifier
*p < 0.05 (Delong et al. [25])
**p < 0.01 (Delong et al. [25])
Med Biol Eng Comput
BFP 0.072 ± 0.030* 0.089 ± 0.027** 0.084 ± 0.027** 0.057 ± 0.031 0.032 ± 0.028
ADAB - 0.016 ± 0.017 0.012 ± 0.016 0.015 ± 0.022 0.040 ± 0.028
KNN - - 0.004 ± 0.011 0.031 ± 0.018 0.056 ± 0.024*
RF - - - 0.027 ± 0.016 0.052 ± 0.026*
SVMR - - - - 0.0250 ± 0.030
BFP, best FOT parameter (obtained without the use of classifiers); ADAB, AdaBoost with decision tree classifiers; KNN, K-nearest neighbor (K = 1); RF,
random forests; SVMR, support vector machines with radial basis kernel; NFC, neuro-fuzzy classifier
*p < 0.05 (Delong et al. [25])
**p < 0.01 (Delong et al. [25])
Med Biol Eng Comput
services provided to these patients, making the use of the FOT be useful for simplifying the use of the FOT in the everyday
simpler and improving the discrimination of the cited evaluation of respiratory function. Particularly, the neuro-
diseases. fuzzy classifier provides simple rules to explain the achieved
This study confirms that the use of ML algorithms is asso- classification and a graphical interface, which is very easy to
ciated with a significant improvement in the diagnostic accu- use. These classifiers hold the promise of improving the med-
racy. This is consistent with our earlier observations, which ical services provided to patients with asthma and restrictive
showed that these algorithms improved the early diagnosis of diseases.
the smoking effects [8], COPD identification [4, 6], automatic
analysis of disease severity [5], and respiratory abnormalities Funding information This study was supported by the Brazilian Council
for Scientific and Technological Development (CNPq) and the Rio de
in sickle cell anemia [68, 69]. These results further support
Janeiro State Research Supporting Foundation (FAPERJ) and in part by
previous works of other researchers describing improvements the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior -
in the description of biological systems [32], as well as in the Brasil (CAPES) - Finance Code 001.
diagnostic accuracy of respiratory exams based on pulmonary
sounds [89], magnetic resonance imaging [61], and spirome- Compliance with ethical standards
try [67, 92]. The interested reader may find in the supplement
(Table S1) a detailed description of the increase in the diag- This research was approved by the research ethics board of the State
University of Rio de Janeiro, and the post-informed consent of all volun-
nostic accuracy observed in the present study and in previous teers was obtained before inclusion in the study. The study was conducted
studies due to the use of ML methods. in accordance with the Declaration of Helsinki.
A clear description of the potential limitations of this paper
is necessary. The principal limitation of this study is that we Conflict of interest The authors declare that they have no conflict of
only studied 77 subjects with asthma and restrictive diseases interest.
and that the precise sensitivity and specificity continues un-
known. More research using a higher number of volunteers is
necessary. However, this preliminary result significantly con- References
tributes to elucidate important debates concerning the differ-
ential diagnosis of obstructive and restrictive diseases [48, 58, 1. Abe S (2009) Support vector machines for pattern classification,
90], and the use of ML algorithms in the clinical diagnosis of advances in computer vision and pattern recognition, 2nd edn.
Springer, New York
respiratory abnormalities [23].
2. Abraham A (2005) Adaptation of fuzzy inference system using
Provided that the work was limited to Brazilian population neural learning. In: Nedjah N, Macedo Mourelle Ld (eds) Fuzzy
at a single practice site, it was not possible to guarantee its systems engineering, vol 181. Springer Berlin Heidelberg Berlin,
generalizability to a different population. Future works should pp. 53–83
include multicenter data to expand the generalizability of the 3. Abu-Mostafa YS, Magdon-Ismail M, Lin H-T (2012) Learning
from data: a short course. AMLbook.com, S.l.
findings. It is noteworthy that readers can easily assess wheth-
4. Amaral JL, Faria AC, Lopes AJ, Jansen JM, Melo PL (2010)
er they are likely to obtain similar results in their own patient Automatic identification of chronic obstructive pulmonary disease
population by examining the inclusion and exclusion criteria based on forced oscillation measurements and artificial neural net-
adopted and the biometric features. It is relevant to consider works. Conference proceedings : Annual International Conference
of the IEEE Engineering in Medicine and Biology Society IEEE
also that the experimental design of the present work increases
Engineering in Medicine and Biology Society Conference 2010:
its generalizability. Globally recognized inclusion and exclu- 1394–1397. https://doi.org/10.1109/IEMBS.2010.5626727
sion criteria were used, and the work was conducted under 5. Amaral JL, Lopes AJ, Faria AC, Melo PL (2015) Machine learning
usual clinical procedures in a typical setting. algorithms and forced oscillation measurements to categorise the
airway obstruction severity in chronic obstructive pulmonary dis-
ease. Comput Methods Prog Biomed 118:186–197. https://doi.org/
10.1016/j.cmpb.2014.11.002
5 Conclusions 6. Amaral JL, Lopes AJ, Jansen JM, Faria AC, Melo PL (2012)
Machine learning algorithms and forced oscillation measurements
We developed and analyzed the performance of various clas- applied to the automatic identification of chronic obstructive pul-
monary disease. Comput Methods Prog Biomed 105:183–193.
sifier algorithms to elaborate a clinical decision support sys-
https://doi.org/10.1016/j.cmpb.2011.09.009
tem to help in the differential diagnosis of asthma and restric- 7. Amaral JL, Lopes AJ, Jansen JM, Faria AC, Melo PL (2013) An
tive respiratory diseases. FOT indexes are able only to reach improved method of early diagnosis of smoking-induced respirato-
moderate diagnostic accuracy (0.70–0.90). The introduction ry changes using machine learning algorithms. Comput Methods
of the neuro-fuzzy and ML classifiers resulted in a significant Prog Biomed 112:441–454. https://doi.org/10.1016/j.cmpb.2013.
08.004
improvement in the accuracy, attaining high accuracy in the 8. Amaral JL, Lopes AJ, Veiga J, Faria AC, Melo PL (2017) High-
differential diagnosis of patients with asthma and restrictive accuracy detection of airway obstruction in asthma using machine
respiratory diseases. Additionally, the developed system may learning algorithms and forced oscillation measurements Computer
Med Biol Eng Comput
methods and programs in biomedicine. 144:113–125. https://doi. 26. Dellaca RL, Duffy N, Pompilio PP, Aliverti A, Koulouris NG,
org/10.1016/j.cmpb.2017.03.023 Pedotti A, Calverley PM (2007) Expiratory flow limitation detected
9. Amaral JLM, Melo PL (2020) Clinical decision support systems to by forced oscillation and negative expiratory pressure. Eur Respir J
improve the diagnosis and management of respiratory diseases. In: 29:363–374. https://doi.org/10.1183/09031936.00038006
Barh D (ed) Artificial intelligence in precision health. Elsevier, 27. DeLong ER, DeLong DM, Clarke-Pearson DL (1988) Comparing
USA the areas under two or more correlated receiver operating charac-
10. Azar AT, Hassanien AE (2015) Dimensionality reduction of med- teristic curves: a nonparametric approach. Biometrics:837–845
ical big data using neural-fuzzy classifier. Soft Comput 19:1115– 28. Demšar J (2006) Statistical comparisons of classifiers over multiple
1127. https://doi.org/10.1007/s00500-014-1327-4 data sets. J Mach Learn Res 7:1–30
11. Bates JHT, Irvin CG, Farré R, Hantos Z (2011) Oscillation mechan- 29. Dietterich TG (1998) Approximate statistical tests for comparing
ics of the respiratory system. In: Terjung R (ed) Comprehensive supervised classification learning algorithms. Neural Comput 10:
physiology. John Wiley & Sons, Inc., Hoboken 1895–1923
12. Bit A, Chattyopadhay H, Nag D (2009) Study of airflow in the 30. Drummond MB, Buist AS, Crapo JD, Wise RA, Rennard SI (2014)
trachea of a bronchopulmonary patient using CT data. Indian Chronic obstructive pulmonary disease: NHLBI workshop on the
Journal of Biomechanics:31–36 primary prevention of chronic lung diseases. Annals of the
13. Bousquet J, Tanasescu CC, Camuzat T, Anto JM, Blasi F, Neou A, American Thoracic Society 11(Suppl 3):S154–S160. https://doi.
Palkonen S, Papadopoulos NG, Antunes JP, Samolinski B, org/10.1513/AnnalsATS.201312-432LD
Yiallouros P, Zuberbier T (2013) Impact of early diagnosis and 31. Dubois AB, Brody AW, Lewis DH, Burgess BF Jr (1956)
control of chronic respiratory diseases on active and healthy ageing. Oscillation mechanics of lungs and chest in man. J Appl Physiol
A debate at the European Union Parliament. Allergy 68:555–561. 8:587–594
doi:https://doi.org/10.1111/all.12115 32. Eswari JS, Majdoubi J, Naik S, Gupta S, Bit A, Rahimi-Gorji M,
14. Breiman L (2001) Random forests. Mach Learn 45:5–32 Saleem A (2020) Prediction of stenosis behaviour in artery by neu-
15. Brochard L, Pelle G, de Palmas J, Brochard P, Carre A, Lorino H, ral network and multiple linear regressions. Biomech Model
Harf A (1987) Density and frequency dependence of resistance in Mechanobiol. https://doi.org/10.1007/s10237-020-01300-z
early airway obstruction. Am Rev Respir Dis 135:579–584. https:// 33. Faria AC, Barbosa WR, Lopes AJ, Pinheiro Gda R, Melo PL
doi.org/10.1164/arrd.1987.135.3.579 (2012) Contrasting diagnosis performance of forced oscillation
16. Brusasco V, Barisione G, Crimi E (2015) Pulmonary physiology: and spirometry in patients with rheumatoid arthritis and respiratory
future directions for lung function testing in COPD. Respirology symptoms. Clinics 67:987–994
20:209–218. https://doi.org/10.1111/resp.12388 34. Faria AC, Lopes AJ, Jansen JM, Melo PL (2009) Assessment of
17. Busse WW, Erzurum SC, Blaisdell CJ, Noel P (2014) Executive respiratory mechanics in patients with sarcoidosis using forced os-
summary: NHLBI workshop on the primary prevention of chronic cillation: correlations with spirometric and volumetric measure-
lung diseases. Annals of the American Thoracic Society 11(Suppl ments and diagnostic accuracy. Respiration; international review
3):S123–S124. https://doi.org/10.1513/AnnalsATS.201312- of thoracic diseases 78:93–104. https://doi.org/10.1159/000213756
421LD 35. Faria AC, Lopes AJ, Jansen JM, Melo PL (2009) Evaluating the
18. Cavalcanti JV, Lopes AJ, Jansen JM, Melo PL (2006) Detection of forced oscillation technique in the detection of early smoking-
changes in respiratory mechanics due to increasing degrees of air- induced respiratory changes. Biomed Eng Online 8:22. https://doi.
way obstruction in asthma by the forced oscillation technique. org/10.1186/1475-925X-8-22
Respir Med 100:2207–2219. https://doi.org/10.1016/j.rmed.2006. 36. Faria ACD, Lopes AJ, Jansen JM, PLd M (2009) Assessment of
03.009 respiratory mechanics in patients with sarcoidosis using forced os-
19. Cawley GC, Talbot NLC (2010) On over-fitting in model selection cillations. Respiration 78:93–104
and subsequent selection bias in performance evaluation. J Mach 37. Ferguson GT, Enright PL, Buist AS, MW H (2000) Office spirom-
Learn Res 11:2079–2107 etry for lung health assessment in adults: a consensus statement
20. Cetişli B, Barkana A (2010) Speeding up the scaled conjugate gra- from the National Lung Health Education Program. Chest 117:
dient algorithm and its application in neuro-fuzzy classifier training. 1146–1161
Soft Comput 14:365–378. https://doi.org/10.1007/s00500-009- 38. Freund Y, Schapire R, Abe N (1999) A short introduction to
0410-8 boosting. Journal-Japanese Society For Artificial Intelligence 14:
21. Cordón O (2011) A historical review of evolutionary learning 1612
methods for Mamdani-type fuzzy rule-based systems: designing 39. Gacto MJ, Alcalá R, Herrera F (2011) Interpretability of linguistic
interpretable genetic fuzzy systems. Int J Approx Reason 52:894– fuzzy rule-based systems: an overview of interpretability measures.
913. https://doi.org/10.1016/j.ijar.2011.03.004 Inf Sci 181:4340–4360. https://doi.org/10.1016/j.ins.2011.02.021
22. Croxton TL, Weinmann GG, Senior RM, Hoidal JR (2002) Future 40. GOLD (2013) Global Initiative For Chronic Obstructive Lung
research directions in chronic obstructive pulmonary disease. Am J Disease – UPDATE (2013). In: Global strategy for the diagnosis,
Respir Crit Care Med 165:838–844. https://doi.org/10.1164/ management, and prevention of chronic obstructive pulmonary dis-
ajrccm.165.6.2108036 ease. NHLBI/WHO
23. Das N, Topalovic M, Janssens W (2018) Artificial intelligence in 41. Golpe R, Jimenez A, Carpizo R, Cifrian JM (1999) Utility of home
diagnosis of obstructive lung disease: current status and future po- oximetry as a screening test for patients with moderate to severe
tential. Curr Opin Pulm Med 24:117–123. https://doi.org/10.1097/ symptoms of obstructive sleep apnea. Sleep 22:932–937
MCP.0000000000000459 42. Guyon I, Lisseff A (2003) An introduction to variable and feature
24. de Melo PL, Werneck MM, Giannella-Neto A (2000) New imped- selection. J Mach Learn Res 3:1157–1182
ance spectrometer for scientific and clinical studies of the respira- 43. Hajian-Tilaki K (2013) Receiver operating characteristic (ROC)
tory system. Rev Sci Instrum 71:2867–2872 curve analysis for medical diagnostic test evaluation. Caspian J
25. de Sá PM, Lopes AJ, Jansen JM, de Melo PL (2013) Oscillation Intern Med 4:627–635
mechanics of the respiratory system in never-smoking patients with 44. Hastie T, Tibshirani R, Friedman J (2009) The elements of statisti-
silicosis: pathophysiological study and evaluation of diagnostic ac- cal learning, 2nd edn. Springer-Verlag
curacy. In: Clinics (Sao Paulo), 68. 5. pp 644-651. doi:https://doi. 45. Hastie T, Tibshirani R, Friedman J (2009) The elements of statisti-
org/10.6061/clinics/2013(05)11 cal learning. Springer Series in Statistics, New York
Med Biol Eng Comput
46. Haykin SS (2009) Neural networks and learning machines. 3rd ed and support vector machine. BioMedical Engineering OnLine 14.
edn. Prentice Hall, New York doi:https://doi.org/10.1186/s12938-015-0003-y
47. Hüllermeier E (2005) Fuzzy methods in machine learning and data 66. Majid A, Ali S, Iqbal M, Kausar N (2014) Prediction of human
mining: status and prospects. Fuzzy Sets Syst 156:387–406. https:// breast and colon cancers from imbalanced data using nearest neigh-
doi.org/10.1016/j.fss.2005.05.036 bor and support vector machines. Comput Methods Prog Biomed
48. Ionescu CM, Machado JT, De Keyser R (2011) Is multidimensional 113:792–808. https://doi.org/10.1016/j.cmpb.2014.01.001
scaling suitable for mapping the input respiratory impedance in 67. Manoharan SC, Veezhinathan M, Ramakrishnan S (2008)
subjects and patients. Comput Methods Prog Biomed 2011:189– Comparison of two ANN methods for classification of spirometer
200 data. MEASUREMENT SCIENCE REVIEW 8:53–57
49. Jain AK (2010) Data clustering: 50 years beyond K-means. Pattern 68. Marinho CL, Maioli MCP, Amaral JLM, LA J, PL M (2018)
Recogn Lett 31:651–666. https://doi.org/10.1016/j.patrec.2009.09. Respiratory resistance and reactance in adults with sickle cell ane-
011 mia: part 2 - fractional-order modeling and a clinical decision sup-
50. Jang J-SR, others Fuzzy modeling using generalized neural net- port system for the diagnostic of respiratory disorders. PLoS One
works and Kalman filter algorithm. In, 1991 1991. pp 762–767 14:e0213257. https://doi.org/10.1371/journal.pone.0213257
51. Jang J-SR, Sun C-T, Mizutani E (1997) Neuro-fuzzy and soft com- 69. Marinho CL, MCP M, do JLM A, AJ L, PL M (2017) Respiratory
puting; a computational approach to learning and machine resistance and reactance in adults with sickle cell anemia: correla-
intelligence tion with functional exercise capacity and diagnostic use. PLoS One
52. Jang JSR (1993) ANFIS: adaptive-network-based fuzzy inference 12:e0187833. https://doi.org/10.1371/journal.pone.0187833
system. IEEE Transactions on Systems, Man, and Cybernetics 23: 70. Miller MR, Hankinson J, Brusasco V, Burgos F, Casaburi R,
665–685. https://doi.org/10.1109/21.256541 Coates A, Crapo R, Enright P, CPMvd G, Gustafsson P, Jensen
53. Japkowicz N, Shah M (2011) Evaluating learning algorithms: a R, DC J, MacIntyre N, McKay R, Navajas D, Pedersen OF,
classification perspective. Cambridge University Press, Pellegrino R, Viegi G, Wanger J (2005) Standardisation of spirom-
Cambridge, New York etry. https://doi.org/10.1183/09031936.05.00034805
54. Jornal Brasileiro de Pneumologia - Diretrizes para Testes de Função 71. Miranda IA, Dias Faria AC, Lopes AJ, Jansen JM, Lopes de Melo P
Pulmonar. (2002). http://www.jornaldepneumologia.com.br/ (2013) On the respiratory mechanics measured by forced oscillation
detalhe_suplemento.asp?id=45 technique in patients with systemic sclerosis. PLoS One 8:e61657.
55. King GG, Bates J, Berger KI, Calverley P, de Melo PL, Dellaca RL, https://doi.org/10.1371/journal.pone.0061657
Farre R, Hall GL, Ioan I, Irvin CG, Kaczka DW, Kaminsky DA, 72. Mohri M, Rostamizadeh A, Talwalkar A (2012) Foundations of
Kurosawa H, Lombardi E, Maksym GN, Marchal F, Oppenheimer machine learning. Adaptive computation and machine learning se-
BW, Simpson SJ, Thamrin C, van den Berge M, Oostveen E (2019) ries. MIT Press, Cambridge, MA
Technical standards for respiratory oscillometry. Eur Respir J 55: 73. Møller MF (1993) A scaled conjugate gradient algorithm for fast
1900753. https://doi.org/10.1183/13993003.00753-2019 supervised learning. Neural Netw 6:525–533. https://doi.org/10.
56. King GG, Bates J, Berger KI, Calverley P, de Melo PL, Dellaca RL, 1016/S0893-6080(05)80056-5
Farre R, Hall GL, Ioan I, Irvin CG, Kaczka DW, Kaminsky DA, 74. Nagels J, Landser FJ, van der Linden L, Clement J, Van de
Kurosawa H, Lombardi E, Maksym GN, Marchal F, Oppenheimer Woestijne KP (1980) Mechanical properties of lungs and chest wall
BW, Simpson SJ, Thamrin C, van den Berge M, Oostveen E (2020) during spontaneous breathing. J Appl Physiol Respir Environ
Technical standards for respiratory oscillometry. Eur Respir J 55: Exerc Physiol 49:408–416
1900753. https://doi.org/10.1183/13993003.00753-2019 75. Nauck D, Kruse R, Klawonn F (1997) Foundations of neuro-fuzzy
57. Kuncheva LI (2004) Combining pattern classifiers: methods and systems. John Wiley, Chichester ; New York
algorithms. John Wiley & Sons 76. Nicolai M, Peter B (2010) Stability selection. Journal of the Royal
58. Lappas AS, Tzortzi AS, Behrakis BK (2014) Forced oscillations in Statistical Society: Series B (Statistical Methodology) 72:417–473
applied respiratory physiology: clinical applications. Clin Res 77. Nilsson AM, Theander E, Hesselstrand R, Piitulainen E, Wollmer
Pulmonol 2:1016–1033 P, Mandl T (2014) The forced oscillation technique is a sensitive
59. Lima AN, Faria AC, Lopes AJ, Jansen JM, Melo PL (2015) Forced method for detecting obstructive airway disease in patients with
oscillations and respiratory system modeling in adults with cystic primary Sjogren’s syndrome. Scand J Rheumatol 43:324–328.
fibrosis. Biomed Eng Online 14:11. https://doi.org/10.1186/ https://doi.org/10.3109/03009742.2013.856466
s12938-015-0007-7 78. Pedregosa F, Varoquaux G, Gramfort A, Bertrand Thirion VM,
60. Lorino AM, Zerah F, Mariette C, Harf A, Lorino H (1997) Grisel O, Blondel M, Müller A, Nothman J, Louppe G,
Respiratory resistive impedance in obstructive patients: linear re- Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A,
gression analysis vs viscoelastic modelling. Eur Respir J 10:150– Cournapeau D, Brucher M, Perrot M, Duchesnay E (2011) Scikit-
155 learn: machine learning in Python. J Mach Learn Res 12:2825–
61. Lungu A, Swift AJ, Capener D, Kiely D, Hose R, Wild JM (2016) 2830
Diagnosis of pulmonary hypertension from magnetic resonance 79. Pereira CAdC, Barreto SdP, Simöes JG, Pereira FWL, Gerstler JG,
imaging-based computational models and decision tree analysis. Nakatani J (1992) Reference values for spirometry in Brazilian
Pulmonary circulation 6:181–190. https://doi.org/10.1086/686020 adults. doi:lil-123525
62. Ma Y, Guo G (2014) Support vector machines applications. 80. Peters U, Hernandez P, Dechman G, Ellsmere J, Maksym G (2016)
Springer Early detection of changes in lung mechanics with oscillometry
63. MacIntyre NR (2012) The future of pulmonary function testing. following bariatric surgery in severe obesity. Applied physiology,
Respir Care 57:154–161; discussion 161-154. doi:https://doi.org/ nutrition, and metabolism = Physiologie appliquee, nutrition et
10.4187/respcare.01422 metabolisme 41:538-547. doi:https://doi.org/10.1139/apnm-2015-
64. MacLeod D, Birch M (2001) Respiratory input impedance mea- 0473
surement: forced oscillation methods. Medical & biological engi- 81. Raskutti G, Wainwright MJ, Yu B (2014) Early stopping and non-
neering & computing 39:505–516 parametric regression: an optimal data-dependent stopping rule.
65. Madero Orozco H, Vergara Villegas OO, Cruz Sánchez VG, Ochoa The Journal of Machine Learning Research 15:335–366
Domínguez HdJ, Nandayapa Alfaro MdJ (2015) Automated system 82. Reisch S, Schneider M, Timmer J, Geiger K, Guttmann J (1998)
for lung nodules classification based on wavelet feature descriptor Evaluation of forced oscillation technique for early detection of
Med Biol Eng Comput
airway obstruction in sleep apnea: a model study. Technology and 92. Topalovic M, Das N, Burgel PR, Daenen M, Derom E,
health care : official journal of the European Society for Haenebalcke C, Janssen R, Kerstjens HAM, Liistro G, Louis R,
Engineering and Medicine 6:245–257 Ninane V, Pison C, Schlesser M, Vercauter P, Vogelmeier CF,
83. Reisch S, Steltner H, Timmer J, Renotte C, Guttmann J (1999) Wouters E, Wynants J, Janssens W, Pulmonary Function Study I,
Early detection of upper airway obstructions by analysis of acous- Pulmonary Function Study I (2019) Artificial intelligence outper-
tical respiratory input impedance. Biol Cybern 81:25-37. doi:DOI forms pulmonologists in the interpretation of pulmonary function
https://doi.org/10.1007/s004220050542 tests. Eur Respir J 53:1801660. https://doi.org/10.1183/13993003.
84. PMd S, AJ L, JM J, PLd M (2013) Oscillation mechanics of the 01660-2018
respiratory system in never-smoking patients with silicosis: patho- 93. Vapnik VN (2000) The nature of statistical learning theory.
physiological study and evaluation of diagnostic accuracy. Clinics Springer New York, New York, NY
(Sao Paulo) 68:644–651. https://doi.org/10.6061/clinics/2013(05) 94. Witten IH, Frank E, Hall MA, Pal CJ (2016) Data mining: practical
11 machine learning tools and techniques. Morgan Kaufmann
85. Sahin D, Ubeyli ED, Ilbay G, Sahin M, Yasar AB (2010) Diagnosis 95. World Health Organization WHO (2019) GINA – Global Initiative
of airway obstruction or restrictive spirometric patterns by for Asthma
multiclass support vector machines. J Med Syst 34:967–973. 96. Zadeh LA (1965) Fuzzy sets. Inf Control 8:338–353. https://doi.
https://doi.org/10.1007/s10916-009-9312-7 org/10.1016/S0019-9958(65)90241-X
86. Sancho AG, Faria ACD, Amaral JLM, Lopes AJ, Melo PL
Evaluation of the forced oscillation technique in the differential Publisher’s note Springer Nature remains neutral with regard to jurisdic-
diagnosis of obstructive and restrictive respiratory diseases. In: tional claims in published maps and institutional affiliations.
IFMBE Proceedings of the XXVI Brazilian Congress on
Biomedical Engineering, Búzios, Rio de Janeiro, 2018. Springer,
The International Federation for Medical and Biological
Engineering (IFMBE) Proceedings book series., p 45 to 50. doi: JorgeAmaral received his D.Sc.
https://doi.org/10.1007/978-981-13-2119-1_7 degree from the Catholic
87. Schapire RE (2013) Explaining adaboost. In: Empirical inference. University of Rio de Janeiroin
Springer, pp. 37–52 2006. He is an Associate
88. Scornet E, Biau G, Vert J-P (2015) Consistency of random forests. Professor at the State University
Ann Stat 43:1716–1741. https://doi.org/10.1214/15-AOS1321 of Rio de Janeiro.His main re-
search area is Machine Learning
89. Sen I, Saraclar M, Kahya YP (2015) A comparison of SVM and
applied to instrumentation.
GMM-based classifier configurations for diagnostic classification
of pulmonary sounds. IEEE Trans Biomed Eng 62:1768–1776.
https://doi.org/10.1109/TBME.2015.2403616
90. Sugiyama A, Hattori N, Haruta Y, Nakamura I, Nakagawa M,
Miyamoto S, Onari Y, Iwamoto H, Ishikawa N, Fujitaka K,
Murai H, Kohno N (2013) Characteristics of inspiratory and expi-
ratory reactance in interstitial lung disease. Respiratory medicine
107:875-882. doi:DOI https://doi.org/10.1016/j.rmed.2013.03.005
91. Swets JA (1988) Measuring the accuracy of diagnostic systems.
Science 240:1285–1293