Comparison of the performances of machine learning methods in the diagnosis of pulmonary diseases with a voice transmission speed

It has been possible to create artificial intelligence by creating models on existing data and defining complex relationships with the methods of machine learning. Machine learning includes two main steps as the teaching-training and trial-test phases. A model is created in the teaching-training phase and the validity of the model is checked during the trial-test phase. There are 2 types of machine learning methods in terms of teaching-training. These are counseling and non-counseling methods of machine learning. The counseling machine learning methods include linear regression, logistic regression, discriminant analysis, Support Vector Machine, Decision tree, K-nearest neighbor algorithm, Simple (Native) Bayes Classification, and artificial neural network algorithms. The non-counseling machine learning methods include k-means, hierarchical clustering analysis, PCA (Principal Components Analysis) and mixture model. The classification of the data used in the teaching-training phase is apparent in the counseling-learning method. However, there is no specific classification in the non-counseling learning method. Among the performance evaluation methods of the models created, exclusion, repeated exclusion, triple separation, cross-validation, and bootstrap are the most commonly used ones. In this study, being inspired by the principle of the sound transition velocity in different environments, a simulation trial was performed for the diagnosis of Pneumothorax and Effusion patients. The average transition velocities and standard deviations of sound were calculated using sections on the CT images. Accordingly, the simulation data was obtained which included 130, 500, 1000, 5000, and 10000 samples. The fact that the distribution of the samples used in the validation phase is similar to the distribution in the population where the actual data come from is important for a more accurate evaluation. Therefore, unbalanced groups need to be balanced in terms of the number of samples. The SMOTE method is one of the most commonly used methods for this balancing process. In the analyzes of the groups performed before and after balancing, the model performances were evaluated using the evaluation criteria of sensitivity, specificity, positive predictive value, negative predictive value, F-criterion, and accuracy rate. Keywords: Machine Learning; Supervised and Unsupervised Machine Learning; SMOTE; Sound Transition Velocity; Pneumothorax and Effusion