LSTM-Based Time Series Forecasting of Pulmonary Function Test for COPD Early Diagnosis

Chronic Obstructive Pulmonary Disease (COPD) is a serious lung disease that makes breathing difficult and cannot be easily detected. Even though early diagnosis technology for COPD using machine learning has been developed, Pulmonary Function Test (PFT) data-based time series prediction studies are still lacking. We use PFT data with insufficient measurement intervals, propose a Long Short-Term Memory (LSTM) to predict PFT values for the future 1Q from the past 2Q, and classify whether COPD occurs or not. The data were interpolated to resolve the imbalanced time period. To confirm the validity of the augmented data, Multivariate Analysis of Variance (MANOVA) was performed, and through the rigorous MANOVA, we proved that there was no significant difference between the original and interpolated data. Mean Absolute Percentage Error (MAPE), recalls, and F1 scores, which are the harmonic mean of precision and recall for classification, were measured for two test scenarios: only the original data and the augmented data. Finally, we found the interpolated data decreased MAPE by almost 7%, however, improved recall and F1 score by almost 22% and 12% for obstructive pulmonary disease, compared with the original data. Besides, we can predict COPD within 3 months, irrelevant to smokers and non-smokers.