Performance of multivariable risk prediction algorithms in predicting COPD exacerbations: a population-based study
Jeenat Mehareen, Laura Huey Mien Lim, Amin Adibi, Joseph Emil Amegadzie, Yuan Xia, Mary A De Vera, Michael R Law, Don D Sin, Surya P Bhatt, Jennifer K Quint, Mohsen SadatsafaviIntroduction
Efficient preventive management of acute exacerbation of chronic obstructive pulmonary disease (COPD) is predicated on accurate risk stratification. We compared the performance of exacerbation history (current standard of care) versus a revised version of a multivariable risk scoring tool (Acute COPD Exacerbation Prediction Tool (ACCEPT)) using primary-care UK data.
Methods
We used validated case definitions to identify diagnosed patients with COPD ≥40 years old from the UK Clinical Practice Research Datalink Aurum (2004–2020). For each patient, a single annual COPD visit was randomly selected as the index date. The outcome was the occurrence of ≥1 moderate/severe exacerbation(s) within a year of the index date. We conducted time-to-event analyses of the latest version of ACCEPT (ACCEPT 2.0) and developed a recalibrated version (ACCEPT 3.0-UK). Model performance was evaluated using discrimination (time-dependent area under the receiver operating characteristic curve (AUC)), calibration and net benefit.
Results
The final cohort included 158 384 patients (55.0% male; mean age 71.5 years). ACCEPT 2.0 achieved an AUC of 0.77 for predicting moderate/severe exacerbations, outperforming both any and frequent exacerbator categories (AUC: 0.69 and 0.67, respectively). However, it overpredicted exacerbation events (observed-to-expected (O/E) ratio: 0.81 (95% CI 0.80 to 0.81)). Recalibration resolved this overprediction, yielding O/E ratio of 1.00 (95% CI 0.99 to 1.00), while maintaining discrimination (AUC: 0.77). ACCEPT 3.0-UK was net beneficial and superior to exacerbation history across a wide range of risk thresholds.
Conclusion
ACCEPT 3.0-UK has substantially higher performance than exacerbation history, quantifies predicted risks for shared decision-making and is likely to confer clinical utility for risk stratification in primary care.