A Data-Driven Machine-Learning Approach for Small Airway Disease Risk Prediction

Background and aims: Small airway disease (SAD) is a common aspect in chronic obstructivepulmonary disease (COPD). The nature of this disease is chronic and exacerbated over time withconsiderable morbidity and mortality. This study tries to use the modern science of artificial intelligenceand its application for early prediction of SAD by assessing the risk factors related tothis disease.Method: We used a large-scale dataset from the National Health and Nutritional ExaminationSurvey (NHANES) ۲۰۰۷-۲۰۱۲ on ۱۸۶۵۸ both males and females ۶-۷۹ years to develop a screeningtool for small airway disease based on ۲۸ features extracted using COPD assessment test(CAT) and other relevant risk factors for COPD. Our label was measured FEV۱/predicted FEV۱calculated from their raw curve spirometry. Multiple machine learning models (Naive Bayes,logistic regression, random forest, and gradient boosting) were compared on the level of classificationaccuracy and the area under the receiver operating characteristic curve (AUC-ROC). Toimprove prediction accuracy, the disparate models were combined to create a weighted ensemblemodel. We also used a trained Transformer, with state-of-the-art classification methods namedTabPFNClassifier.Results: Different machine learning models including Naïve Bayes, logistic regression, randomforest, and gradient boosting reached an accuracy of about ۶۱%, ۶۲%, ۶۱%, and ۶۵% respectively,and AUC-ROC score of ۵۳.۳%, ۷۶.۴%, ۷۰.۷% in the first three model respectively in comparisonof XGBoost which performed the best at ۷۸.۵%. TabPFNClassifier on a smaller part of datasetreached an accuracy of about ۶۹.۴% (AUC-ROC = ۶۹.۷%) and developed ensemble model usingstacked generalization achieved an accuracy of about ۶۴.۴%.Top five predictors for small airway disease in decreasing order were age, weight, sleeping hours,gender, and presence of Shortness of breath on stairs/inclines.Conclusion: We conclude machine learning models based on survey questionnaires can be valuablefor predicting the risk of small airway disease in at risk patients. We also identified some ofthe key contributors to the prediction, which can be used in developing a screening tool. A broaderrange of health data and advanced ML techniques are needed to conduct further research.

کلیدواژه ها:

artificial intelligence ، machine learning ، small airway disease ، COPD ، spirometry