Predicting total lung capacity from spirometry: a machine learning approach.

Area of Science:

• Pulmonary Medicine
• Machine Learning
• Medical Diagnostics

Background:

• Spirometry can indicate restrictive ventilatory impairment, but total lung capacity (TLC) measurements are needed for diagnosis.
• Accurate TLC estimation from spirometry is crucial for efficient patient management.

Purpose of the Study:

• To train a supervised machine learning model for accurate TLC estimation from spirometry data.
• To identify patients who would benefit most from complete pulmonary function testing.

Main Methods:

• Trained three tree-based machine learning models on over 51,000 spirometry data points with TLC measurements.
• Evaluated model performance on an independent test set of 1,402 patients.
• Utilized the best model to identify restrictive ventilatory impairment and compared it to standard spirometry patterns.

Main Results:

• The CatBoost model demonstrated superior performance, estimating TLC with a mean squared error of 560.1 mL.
• The algorithm achieved 83% sensitivity, 92% specificity, and a 75% F1-score for predicting restrictive ventilatory impairment.
• Restrictive ventilatory impairment was present in 16.7% of the test set.

Conclusions:

• Machine learning models can accurately estimate TLC from spirometry data.
• This approach can improve diagnostic efficiency for restrictive lung diseases.
• Potential for developing smart home-based spirometry solutions for patient self-monitoring and decision support.