Training machine learning-based spirometry reference equations: a comparison with GAMLSS and GLI reference equations.

Area of Science:

• Respiratory Medicine
• Biostatistics
• Machine Learning Applications

Background:

• Accurate spirometry interpretation relies on population-specific reference equations.
• Existing Generalized Additive Models for Location, Scale and Shape (GAMLSS) models, while effective, can be complex.
• There is a need for simpler, efficient alternatives for spirometry reference equations.

Purpose of the Study:

• To evaluate machine learning (ML)-based spirometry reference equations as an alternative for Jordanian adults.
• To compare the efficiency of ML models against existing GAMLSS and Global Lung Initiative (GLI) equations.
• To assess the clinical utility of ML-derived spirometry norms.

Main Methods:

• A cross-sectional study utilizing existing datasets for ML model training (n=1,948).
• ML models were developed using age and height to predict FEV₁, FVC, and FEV₁/FVC.
• External validation was performed on a new cohort (n=487) of healthy, non-smoking adults.

Main Results:

• ML models demonstrated comparable Mean Squared Errors (MSE) to Jordanian GAMLSS equations.
• ML models showed lower MSE and z-scores closer to zero compared to global reference equations.
• ML and Jordanian models exhibited superior calibration and correctly classified all healthy individuals as normal.

Conclusions:

• ML-derived spirometry equations align well with observed data for Jordanian adults.
• ML models outperform global standards in representing the local population.
• Region-specific reference equations, like those developed using ML, are crucial for accurate respiratory diagnostics.