Deep learning for detecting and early predicting chronic obstructive pulmonary disease from spirogram time series.

Area of Science:

• Pulmonary Medicine
• Artificial Intelligence
• Biomedical Data Science

Background:

• Chronic Obstructive Pulmonary Disease (COPD) diagnosis relies on spirometry, but current methods struggle with early risk prediction.
• Subtle patterns in spirometry data hold potential for forecasting future COPD development.

Purpose of the Study:

• To introduce DeepSpiro, a novel deep learning model for early prediction of future COPD risk.
• To evaluate DeepSpiro's efficacy in detecting COPD and forecasting disease progression.

Main Methods:

• DeepSpiro integrates SpiroSmoother, SpiroEncoder, SpiroExplainer, and SpiroPredictor for robust analysis.
• The model analyzes Volume-Flow time series, focusing on key data patches for prediction.
• Heterogeneous data integration and volume attention mechanisms enhance prediction accuracy.

Main Results:

• DeepSpiro achieved an AUC of 0.8328 for COPD detection in the UK Biobank dataset.
• The model demonstrated statistically significant predictive performance for future COPD risk (p < 0.001).
• Predictions were made for 1-5 year horizons, indicating long-term forecasting capabilities.

Conclusions:

• DeepSpiro offers a powerful tool for early COPD risk assessment and prediction.
• The deep learning approach effectively captures subtle patterns for improved diagnostic and prognostic capabilities.
• DeepSpiro can aid in managing the long-term progression of COPD disease.