SpiDB project: Creation and validation of a spirometry database
Respiratory diseases affect the lungs and other parts of the human respiratory
system and could be caused by infections, direct or indirect consumption of
polluting substances found in the environment. The prevention and control of these diseases must be an absolute priority in decision-making in the health sector. A basic technique for lung function study is spirometry, which represents a necessary test for the evaluation and monitoring of respiratory diseases. At present, specific computerized systems use Machine Learning techniques to assist in diagnosing a respiratory disease from spirometry records. However, there are still no definitive methods to structure and classify these records to achieve an adequate diagnosis. Therefore, in this research, a comparison methodology was designed to diagnose respiratory health through spirometric patterns such as Normal, Obstructive, and Mixed. Through the use of pre-processing techniques, selection of characteristics, and implementation of four methods of subsets in order to identify the most relevant set of parameters. Two methods of dimension reduction, LDA and PCA were also used, and finally selected supervised classification techniques in K-nearest neighbors (KNN), Parzen classifier, Support
Vector Machines (SVC), Random Forests (RF) and Neural Networks (ANN). In search of the best methodology, performance is evaluated by determining the classification error, sensitivity (𝑆𝑒), specificity (𝑆𝑝), precision (𝐴𝑐𝑐), and
computational cost where classifiers such as RF, SVC, and ANN were the most appropriate, reaching values that allowed a very accurate clinical diagnosis, achieving an accuracy of 98%, indicating a high level of diagnosis in respiratory health.