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

Abstract

Introduction: Interpretation of spirometry data depends on the availability of reference equations that reflect the physiological norms of the assessed population. Although GAMLSS models provide clinically acceptable models, they may lack simplicity and ease of application. This study evaluated the efficiency of machine learning (ML)-based spirometry reference equations as an alternative for Jordanian adults.

Method: In this cross-sectional study, ML models were trained using age and height to predict FEV₁, FVC and FEV₁/FVC. Model development was based on the same datasets previously used to construct GAMLSS-based Jordanian equations, which included 1,948 participants (54.2% females). External validation was performed on a newly recruited sample of healthy, non-smoking adults (n = 487, 46.6% females).

Results: ML predicted and lower limits of normal (LLNs) values were compared with those from the Jordanian GAMLSS, GLI equations, using z-score distributions, residual plots, and clinical diagnostic agreement. For both sexes, ML models consistently produced comparable mean squared errors (MSE) to the Jordanian GAMLSS equations and lower MSE values and z-scores closer to zero when compared with global reference equations. Agreement analyses revealed that the ML and Jordanian models more reliably classified individuals within ± 0.5 and ± 1.0 z-score thresholds, emphasizing their superior calibration. ML and Jordanian models were the only ones to classify all the healthy study sample as normal spirometry.

Conclusion: ML-derived spirometry equations demonstrated strong alignment with the observed data and outperformed global standards in representing Jordanian adults. These findings support the use of reference equations customized for specific regions in respiratory diagnostics.