Machine learning for prediction of eosinophilic asthma using clinical indicators as an alternative to induced sputum.

BACKGROUND

Induced sputum differential cell counts can help classify asthma into eosinophilic and non-eosinophilic types. The collection of induced sputum is time-consuming. The aim of this study was to develop a machine learning model to identify eosinophilic asthma (EA) with clinical indicators.

METHODS

The data from 103 asthma patients were collected and divided into training cohort (83 patients) and validation cohort (20 patients). Five AI algorithms, namely logistic regression, support vector machine (SVM), k-Nearest Neighbor, neural network and Naive Bayes, were employed to establish predictive models using clinical indicators and were evaluated using the validation cohort. The threshold parameter of SVM was selected by balancing accuracy, true positive, and false positive rates. And two hyperparameters of SVM, c and γ, were optimized by k-fold cross-validation and grid search methods.

RESULTS

The SVM model presented good discrimination of EA, and the Area Under the Curve(AUC) and accuracy were 0.93 and 81.58 %, respectively. The top six clinically important predictors were Fraction of exhaled Nitric Oxide, blood eosinophils, immunoglobulin E, allergy, Asthma Control Questionnaire -7 and Asthma Control Test. Considering the model performance and generalization ability, the threshold of SVM was selected as 0.5, and the c and γ hyperparameters were set as 1 and 1, respectively. The evaluation results of the validation cohort showed that the SVM model achieved an AUC of 0.77, indicating a favorable ability to identify EA.

CONCLUSION

Machine learning for prediction of eosinophilic asthma using clinical indicators can be an alternative to induced sputum.