Classification of pulmonary diseases using machine learning and deep learning models on GLI-2012 standardized spirometry features.

backgroundArtificial intelligence applications play an increasingly important role in medical diagnosis processes, nowadays. Spirometry, which is widely used among lung function tests, is an important method in the early diagnosis of pulmonary diseases and monitoring. In this study, the performances of different machine learning and deep learning algorithms in classifying pulmonary diseases were examined using a set of Global Lung Initiative 2012 standardized spirometry features.

methodsThe study analyzed demographic variables and different spirometry parameters to make classification for multi class classification. This was defined for both a standard four class and an extended five class problem. In this context, logistic regression, support vector machine, k-nearest neighbor, random forest, and XGBoost methods, as well as deep neural networks, convolutional neural networks models were applied in the study. Model accuracy was increased by applying preprocessing and standardization techniques to the data set. Model performances were analyzed with evaluation criteria such as accuracy, sensitivity, specificity, F1-score, F1-macro and the area under the receiver operating characteristic curve metrics.

resultsUnder the subject-disjoint test set, tree-based models achieved the highest generalization performance. The XGBoost model achieved the highest success rate with 99.24 ± 0.59% accuracy, followed by the random forest model at 97.23 ± 2.04%. The deep learning models, deep neural networks and convolutional neural networks achieved 81.61 ± 5.83% and 77.57 ± 5.55% accuracy, respectively, on the same conventional normal feature set.

conclusionsMachine learning based approaches have offered strong potential in early diagnosis of pulmonary diseases. In addition, the need for additional optimization to improve the performance of these models has been emphasized. This study provides important findings for the integration of artificial intelligence supported spirometry analysis into clinical use. However, validation of the proposed methods on larger patient groups is a critical step for the integration of these technologies into the healthcare system.