Automated Screening Methodology for Asthma Diagnosis that Ensembles Clinical and Spirometric Information

Abstract

Asthma, a chronic disease, involves various clinico-epidemiological factors. In view of this, its early detection followed by management is required for providing better healthcare support to patients. This paper proposes an efficient machine learning methodology for the computerized detection of asthma that ensembles clinico-epidemiological and spirometric information to help pulmonologists in diagnostic decision-making. A total of 42 features (30 clinico-epidemiological and 12 spirometric parameters) are considered for diagnosing asthma. Statistical evaluation is employed to identify potential features for characterizing asthma. A pattern classification scheme based on three neural network models (probabilistic, radial basis function, and multilayer perceptron neural networks) and the alternating decision (AD) tree technique are investigated. Finally, the performance of the designed diagnostic scheme is evaluated in terms of sensitivity, specificity, and overall accuracy. 17 features are found to be statistically significant for discriminating two groups (asthma vs. healthy). Of the trained neural network models, the multilayer perceptron model has the highest accuracy (96.55 % sensitivity, 97.18 % specificity, and 96.86 % overall classification accuracy) for the automated classification of asthma-based statistically significant features. Moreover, the AD tree has a higher asthma screening accuracy (97.73 % sensitivity, 99.53 % specificity, and 99 % overall classification accuracy) than that of the multilayer perceptron model. The AD tree converges to the highest accuracy based on only seven features, which were selected iteratively and automatically.