UBC Theses and Dissertations
Generalizability of risk stratification algorithms for acute exacerbation of chronic obstructive pulmonary disease Ho, Joseph (Khoa Nguyen)
Background: Contemporary management guidelines for chronic obstructive pulmonary disease (COPD) rely on exacerbation history to risk-stratify patients and guide therapy for the prevention of future exacerbations. However, exacerbation history alone may not reliably predict future exacerbations due to random variability in frequency. To address this problem, multivariable prediction models have been developed to improve predictive accuracy. Objective: The objective of this thesis was to assess the generalizability of COPD exacerbation risk stratification algorithms and assess whether the inclusion of race improves the performance of such algorithms. Methods: I evaluated three algorithms: the Acute COPD Exacerbation Prediction Tool (ACCEPT), a prediction model by Bertens et al., and exacerbation history alone, using data from three COPD clinical trials representing different levels of exacerbation risk. I examined discrimination, calibration, and clinical utility as measures for model performance. I then recalibrated the models using the setting-specific exacerbation risk for comparison. I explored race as a variable that could convey information on background risk and assessed whether adjusting for race with a random-effects approach could improve model performance. Results: Both prediction models had better discrimination compared to exacerbation history alone with Δ area under the curves (AUCs) ranging from 0.05 to 0.10 (P-values <0.001). However, no algorithm was superior in clinical utility, and all had the risk of harm. When the models were recalibrated, clinical utility was significantly improved, and the risk of harm was substantially mitigated. The crude exacerbation risk ratios (RRs) of race varied between 0.96 to 1.57. However, in the random-effects model, the shrunken RRs ranged between 0.99 to 1.07. Using the adjusted RRs to update ACCEPT, I showed that the inclusion of race in ACCEPT did not significantly improve model performance compared to the base ACCEPT. The ΔAUCs were <0.01 in all samples with P-values > 0.17. There were also no notable improvements to calibration, clinical utility, or goodness-of-fit (P-value 0.15) after race-adjustment. Conclusions: Risk stratification algorithms for COPD exacerbations are not universally applicable across all settings. However, the flexibility of clinical prediction models allows them to be updated to accommodate setting differences.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International