Examining discordance in spirometry reference equations: A retrospective study

Abstract

This study aimed to evaluate discordance, binary classification, and model fit between race-predicted and race-neutral spirometry prediction equations. Spirometry data from 9506 patients (18-95 years old) self-identifying as White, Black, or Hispanic were analyzed, focusing on the lower limit of normal (LLN). Best-fit prediction equations were developed from 3771 patients with normal spirometry, using Bayesian Information Criterion (BIC) to compare models with and without race as a covariate. Results showed that including race as a covariate improved model fit, reducing BIC by at least ten units compared to Race-Neutral equations. Discordance between race-specific and race-neutral equations for detecting airway obstruction and restrictive spirometry patterns ranged from 4% to 13%. Using race-neutral equations resulted in false discovery rates (FDR) of 14% for Hispanics and 45% for Blacks and false negative rates (FNR) of 21% for Hispanics and 27% for Blacks in diagnosing airway obstruction. These findings indicate that removing race as a covariate in spirometry equations increases FDR and FNR, leading to higher misclassification rates. The 4%-13% discordance in interpreting airway obstruction and restrictive patterns has significant clinical implications, underscoring the need for careful consideration in developing spirometry reference equations.

Keywords: accuracy; ethnicity; lung function; prediction equations; pulmonary; race; reference equations.

FIGURE 1

Density plot of root mean square error (RMSE) values between race‐specific and race‐neutral equations for FEV₁, (Blacks vs. Whites). Race‐specific equation (Blue Line): This equation includes race as a covariate, with race coded as 0 for Mexican and 1 for White. The race‐neutral equation (Green Line) does not include race as a covariate and treats all subjects as one race. Left panel: The density plot shows that the race‐specific reference equation tends to have lower RMSE values compared to the race‐neutral reference equation, as indicated by the blue line being shifted to the left of the green line. The peak density for the race‐specific equation is higher and more concentrated around lower RMSE values, suggesting better model performance compared to the race‐neutral reference equation. The spread of the RMSE values is narrower for the race‐specific equation, indicating more consistent performance. Permutation test results: RMSE Difference = 0.025 L, two‐sided p‐value = 0.004. Thus, this difference is unlikely to have occurred by chance. However, the bootstrapped test results did not show a statistically significant difference between the equations (two‐sided, p = 0.567), implying that this small difference may not be practically meaningful. Thus, in real‐world applications, the choice between race‐specific and race‐neutral equations may not be so important for FEV₁ when comparing Blacks versus Whites. Right panel: The blue line (race‐specific equation) is shifted to the right compared to the green line (race‐neutral equation), indicating that the race‐specific equation generally has higher correlation coefficients (permutation test results, correlation difference = −0.013, two‐sided p‐value = 0.008). This implies that the correlations between actual values and the predicted values on the testing set were closer using the race‐specific equation than the race‐neutral equation. However, the bootstrap results did not show a statistically significant difference between the two equations (p = 0.598), also implying that this small difference may not be practically meaningful. On the other hand, airway obstruction is defined by a combination of FEV₁/FVC and FVC.

FIGURE 2

Density plot of root mean square error (RMSE) values between race‐specific and race‐neutral equations for FVC (Blacks vs. Whites). Race‐specific equation (Blue Curve): This equation includes race as a covariate, with race coded as 0 for Mexican and 1 for White. The race‐neutral model (Green Curve) does not include race as a covariate and treats all subjects as one race. Left panel: The density plot shows that the race‐specific equation tends to have lower RMSE values compared to the race‐neutral equation, as indicated by the blue line being shifted to the left of the green line. The peak density for the race‐specific equation is and more concentrated around lower RMSE values, whereas the density plots for the race‐neutral equation is wider. This suggests that the race‐specific equation has a better and more consistent model performance compared to the race‐neutral equation. Permutation test results: RMSE Difference = 0.038 L, two‐sided p‐value = 0.0011. Thus, the difference between equations is unlikely to have occurred by chance. However, the bootstrapped test results did not show a statistically significant difference between the equations (two‐sided, p = 0.527), implying that this small difference may not be practically meaningful. Thus, in real‐world applications, the choice between race‐specific and race‐neutral equations for FVC may not be so important when comparing Blacks versus Whites. Right panel: The blue line (race‐specific equation) is shifted to the right compared to the green line (race‐neutral equation), indicating that the race‐specific equation generally has higher correlation coefficients (permutation test results, correlation difference = −0.019, two‐sided p‐value = 0.0008). This implies that the correlations between actual values and the predicted values on the testing set were closer using race‐specific equations. However, the bootstrap results did not show a statistically significant difference between the two equations (p = 0.490), implying that this small difference may not be practically meaningful.