Diagnosis of Respiratory Changes in Cystic Fibrosis Using a Soft Voting Ensemble with Bayesian Networks and Machine Learning Algorithms

Purpose: Advances in the treatment of cystic fibrosis (CF) have allowed patients to reach adulthood. The forced oscillation technique (FOT) is a new method for providing an exam that is simple to perform and simultaneously provides a detailed respiratory system evaluation. The purpose of this study was to use machine learning (ML) algorithms to increase the accuracy and interpretability of FOT parameters in the investigation and diagnosis of respiratory changes in adults with CF. Methods: The database was created based on 150 measurements in 50 volunteers (23 in the control group and 27 in the test group). The following supervised ML algorithms were selected for the tests: K-nearest neighbors (KNN), random forest (RF), AdaBoost with decision trees (ADAB), and light gradient boosting (LGB). These data were also subjected to a Bayesian network synthesized by a genetic algorithm (BNGA) in an attempt to maintain good accuracy and increase the interpretability of the results. A soft vote ensemble strategy was employed to enhance the diagnostic accuracy. Results: The first part of this study showed the best FOT parameter: the reactance X_m(AUC = 0.85), indicating moderate accuracy. In the second part, the original FOT parameters were used as input in the chosen algorithms. BNGA had the best performance alone (AUC = 0.88), while the soft voting ensemble achieved AUC = 0.90. When cross-product and feature selection methods were applied, the RF and BNGA were the algorithms with the best results (AUC = 0.88), and the soft voting ensemble achieved an AUC = 0.94. Conclusion: This study provides high diagnostic accuracy with improved interpretability of the FOT parameters, which assists doctors in the medical diagnosis of respiratory changes in CF.