A Prognostic Model to Improve Asthma Prediction Outcomes Using Machine Learning

2.1. Data Description and Acquisition

A dataset containing a mix of clinical findings from the pulmonary function test data, along with the clinical symptoms, was used to validate the proposed prognostic model. The dataset is mixed in that it contains information on both symptomatic representations and medical history in the form of categorical data along with lung function parameters estimated using a spirometer (with the data basically being quantitative (numerical) in nature). Feature engineering involves the adoption of domain knowledge to generate features so that it makes the prediction process more effective for well-tailored machine learning models. In addition to the available lung function parameters, we generate new features from a host of anthropometric indices included in the dataset by employing ARTP respiratory equations [10, 11].

The validation of the feature-driven prediction model is performed by deploying MSFET for the extraction of significant features that signify asthma severity indicators. MSFET involves the selection of severity indicators from the available set of features using feature-scoring techniques. This is followed by a comparative analysis of the various machine learning classifiers. Logistic regression, Support vector machine, and Naive Bayes classifiers were used to validate the performance of the model, and the best-performing classifier was adopted for the approach used. Further deep-learning techniques were also explored for the same [12, 13]. The performance of the model was evaluated on the complete and optimized feature sets obtained via MSFET, and the inferences drawn from an empirical analysis are presented.

We deployed the data documented during a study on the operation of the lungs and its diseases by the University of Innsbruck in the district of Brixlegg in Austria. The dataset included a variety of attributes perceived as covariates with respect to lung disease. The data contained the responses recorded for 1549 children. The missing values for the attributes were recorded as -1. The variables with respect to which the disease was expressed included the degree of pollution in the environment at the place of residence characterized through three categorical levels, extremely polluted, moderately polluted, and highly polluted with ozone, along with the other attributes including parental character-istics, such as details with respect to paternal/ maternal smoking, besides parental level of education and existence of comorbidities including cold, cough and existing allergies. The gender attribute was encoded as 0 for males and 1 for females. A variable on the presence/absence of bronchial tube disease, one of the most important parameters that help in ascertaining asthma as a disease rather than a symptomatic presentation, was also included in the data collected. PEF, the maximum speed of expiration (airflow at exhalation), as measured with a peak flow meter, was also included. Furthermore, clinical findings in the form of spirometer readings obtained by performing spirometry were recorded for the predominant and most common pulmonary function parameters.

2.2. Model Building using Domain Knowledge and Feature Engineering

Pulmonary function tests involving results revealing lung functionality play a major role in the process of making a prognosis of the disease and the assessment of treatment effects. However, it could lead to mis- management of the related disease and the patients affected in the case of encountering differences in the way the lung function is expressed and interpreted [14-16]. Of the several respiratory lung equations that thrive on predicting the expected levels of PFT parameters for a given height, weight, and gender, we try to adopt the ARTP reference equations by introducing a few more lung function parameters for effectively predicting asthma predisposition [17-19].

Initially, a few attributes, including identification number, month of birth and examination, day of month, and examination, were eliminated as we found that they do not significantly contribute to the prediction process. However, the “age” attribute was deduced from the year of examination and the year of birth, which were included in the raw data. Subjects containing missing values for any of the attributes were eliminated. Additional features indicative of pulmonary function parameters fevp and fvcp, representing predicted values for FEV1 and FVC, were added to the dataset.

The fevp and fvcp were computed and deduced using the existing attributes, “height” and “age” for the males and females separately using the formulae published by the Association for Respiratory Technology and Physio- logy. The figures in the formulae are based on a regression model from a cohort study where “height” is in meters and “age” is in years and is expressed as follows for male and female genders separately:

(i) Male:

(ii) Female:

Using fevp and fvcp, we further evaluated the ratio between both, characterizing the Tiffeneau-Pinelli index, a common measure used by clinicians for the monitoring of lung functionality, which was added to the existing set of features. The ratio between the actual fvc recorded by the test and the predicted fvc computed as shown above was also included. Thus, the resulting dataset now included a host of features incorporating clinical findings and symptomatic and medical history. The feature “bmi” was added to the set of existing attributes as it was found to be one of the indicators for diseases in the children of the age group under study, as per the literature [20-22].

A summary of the different variables involved in the dataset is presented in Table 1.

2.3. Balanced Data Creation using OCSVM

2.3.1. Deploying One Class SVM

One class SVM offers a different approach to classification when compared to standard algorithms by modeling the distribution of one class only, and is one of the most preferred solutions in the case of data exhibiting class imbalance. The one-class SVM is used for training only on one class, which in our case represents the non-asthmatics outnumbering the asthmatics, thereby adopting a strategy to eliminate samples that would be regarded as outliers. The OCSVM learning algorithm attempts to input the data into a high-dimensional feature space while iteratively searching the margin that maximizes the hyperplane that best separates the training data from the origin. The OCSVM may be viewed as a regular two-class SVM where all the training data lies in the first class, and the origin is taken as the only member of the second class [23, 24].

One-Class Support Vector Machines (OC-SVMs) are a natural extension of SVMs. In order to identify distrustful observations, an OCSVM approximates a distribution that encompasses the best of the observations and then labels them as “suspicious” for those that lie far from them with respect to a suitable metric. An OCSVM solution is constructed by estimating a probability distribution function that treats most of the observed data more prospective than the rest and a decision rule that separates this observation by the largest possible margin. Involving a quadratic programming problem, the computational complexity of the learning phase is exhaustive, but once the decision function is decided, it can be used to effortlessly predict the class label of unseen data. Further, the method is seen to be effective at handling data containing both continuous and categorical values, and, itwas adopted for the data under study.

Solving the OCSVM optimization problem is equivalent to solving the dual quadratic programming problem (Eqs 5-7).

(5)

With the constraint

(6)

and

(7)

where α is a lagrange multiplier (or “weight” on instance “I” such that vectors associated with non-zero weights are called “support vectors” and exclusively determine the optimal hyperplane), ν is a parameter that controls the trade-off between maximizing the distance of the hyperplane from the origin and the number of data points contained by the hyperplane, “n” is the number of points in the training dataset, and K (xi,xj) is the kernel function. Using the kernel function to project input vectors into a feature space, we allow for nonlinear decision boundaries. Given a feature map (Eq 8):

(8)

wherein, Φ maps training vectors from input space to a high dimensional feature space, we can define the kernel function as (Eq 9):

(9)

With this, the feature vectors do not need to be computed explicitly, and in fact, this capability greatly improves computational efficiency in directly computing kernel values.

The adoption of OCSVM as an alternative method to sampling data has resulted in the selection of an ideal representation of the data rather than the regular sampling approach that is used for undersampling, as there are all chances of leaving out significant instances that might be contributing to the prediction capability as the choice of samples happens randomly. With sampling, there is no underlying computational restriction when the samples are drawn from the original population, and, it might not be a preferred approach to represent the overall population effectively [25, 26].

3.1. Characterizing Asthma Populations with Box Plots

Figs. (2a-c) adopt box plots to depict the data concentrations of the discrete features, namely cold, paternal smoking, and lung disease, chosen by the MSFET technique. The labels “0” and “1” along the bars indicate non-asthmatics and asthmatics, respectively. While “blue” represents absence, “red” represents presence. It can be observed that the features “cold” and “Paternal smoking” are dormant in the category representing non-asthmatics. The feature representing “lung disease” is, however, present in both groups, though the proportion of it is higher in asthmatics, as per the data density observed in the two groups.

Figs. (3a-c) on similar lines depict the box plots for the numerical attributes PEF, FEF50, and BMI, respectively. It is observed that a PEF and FEF50 of 5.244+1.459 and 3.077+ 0.942, respectively, characterize an asthmatic population with a BMI of 17.402+2.69.

It can be observed from Table 3 that of all the classifiers, the Naïve Bayes classifier performed comparatively well with the optimized feature set. The model was able to predict the disease outcome with a precision of 86.1% and recall of 84.7%, accounting for an F1 measure of 84.5%. The AUC and CA were evaluated to be 92.2% and 84.7%, respectively. Furthermore, when the raw dataset excluded the features generated by feature engineering but included all the spirometer readings along with the symptomatic and history data were included, the model performed low, with the naïve Bayes classifier performing comparatively better with a reduced F1-measure of 82.8% accounting for a precision and recall of 83.3% and 82.8% respectively. The AUC and CA obtained with the complete feature set also showed a decline compared to the results with the optimized feature set. The sensitivity, which is most desirable from any of the clinical decisions, is quite low with the raw feature set (82.8%) as compared to the optimized feature set (84.7%). It is evident from this fact that it is of utmost importance to use the optimized feature set, which can always improve the prediction outcomes and enhance the performance efficiency of the models. With the optimized feature set containing the additional features generated by feature engineering, the classification accuracy was considerably increased. Feature engineering further places a significant impact in scenarios involving large data where there are exceedingly a greater number of features and where some of the useful, previously unknown features can be regenerated by using existing features backed by knowledge of the relevant domain.