Using machine learning to develop Iraqi-specific spirometric reference equations

Spirometry is essential for diagnosing and managing respiratory diseases. Accurate interpretation relies on reference equations that reflect population-specific lung function. Global equations, such as those from the Global Lung Function Initiative (GLI), may not suit all populations, including Iraqis.
To develop sex-specific spirometric reference equations for healthy Iraqi adults using machine learning (ML) and compare their performance with the GLI 2012 and 2022 equations.
This cross-sectional study included 3,959 healthy, nonsmoking Iraqi adults aged ≥18 years. Spirometry was performed per ATS/ERS guidelines. Five ML models (linear regression, random forest, support vector machine, gradient boosting machine (GBM), and k-nearest neighbors) were trained using age and height. Data were split into training (70%) and validation (30%) sets. Performance was assessed using RMSE, R2, and z-score calibration. GBM was selected as the best model.
GBM outperformed all other models and GLI equations. In females, R2 was 0.4473 for FEV1 and 0.4519 for FVC; in males, 0.3509 and 0.3674, respectively. GLI equations underestimated lung volumes, while GBM predictions were well calibrated with mean z-scores near zero.
GBM-derived equations show improved accuracy and calibration over GLI standards for Iraqi adults, offering a more suitable tool for spirometry interpretation.
This study used advanced computer methods called machine learning to create new equations that show what normal lung function looks like in healthy Iraqi adults. These equations help doctors interpret breathing tests more accurately than the international standards currently used, which do not fit the Iraqi population well. By training the models on data from healthy Iraqi men and women, the study found that these local equations provide a better match and reduce the chance of misdiagnosis. This approach could also be used in other countries to make breathing tests fairer and more accurate for their own populations.

Sex-specific machine-learning equations for FEV1, FVC, and FEV1/FVC were derived from a community-based cohort of healthy, never-smoking Iraqi adults.
Models showed lower error and better calibration than GLI-2012/2022 for FEV1 and FVC, with comparable calibration for FEV1/FVC.
Predicted LLNs from the ML models aligned more closely with observed distributions, improving diagnostic agreement versus GLI in both sexes.
External validation preserved performance across age and height strata, confirming transportability within the sampled population.
Sex-stratified age–height grids and LLNs are provided to support clinical interpretation and local surveillance.

Sex-specific machine-learning equations for FEV1, FVC, and FEV1/FVC were derived from a community-based cohort of healthy, never-smoking Iraqi adults.
Models showed lower error and better calibration than GLI-2012/2022 for FEV1 and FVC, with comparable calibration for FEV1/FVC.
Predicted LLNs from the ML models aligned more closely with observed distributions, improving diagnostic agreement versus GLI in both sexes.
External validation preserved performance across age and height strata, confirming transportability within the sampled population.
Sex-stratified age–height grids and LLNs are provided to support clinical interpretation and local surveillance.