New COPD Risk Tool ACCEPT 3.0 Cuts Overprediction Across 18 Countries

Researchers have unveiled a recalibrated version of a widely used algorithm for predicting flare-ups of chronic obstructive pulmonary disease, reporting that adding country-specific background risk substantially improved the accuracy of its forecasts. The new tool, called ACCEPT 3.0, was validated and recalibrated using data from NOVELTY, a multi-country observational study of patients treated in routine primary and specialist care. The work, published in Diagnostic and Prognostic Research, addresses a persistent problem in clinical risk prediction: models that perform well where they were built often misjudge risk when applied elsewhere.

The ACute COPD Exacerbation Prediction Tool, or ACCEPT, is an algorithm designed to predict a patient’s 12-month risk of chronic obstructive pulmonary disease exacerbations, the episodes of worsening symptoms that are a major cause of illness and death in COPD. The tool combines a patient’s history of exacerbations, often considered the best predictor of future risk, with other predictors routinely available at the point of care, including age, sex, smoking status, body mass index, lung function, domiciliary oxygen therapy, symptom burden and current medication use. The earlier version, ACCEPT 2.0, had been built using pooled data from three randomised clinical trials and recalibrated in the ECLIPSE cohort, and it has been regarded as one of the few reliable tools for managing COPD exacerbation risk prediction.

However, transportability of prediction models from one setting to another remains a well-known challenge. Recent evidence indicates that models are unlikely to transport well, especially when the outcome prevalence differs significantly across jurisdictions. In a large-scale validation study, prediction models for cardiovascular disease risk failed to generalise to new populations, creating a risk of harm when the net benefit of using the model fell below that of not using it, and this risk was reduced when models were adjusted for the background cardiovascular risk of the new population. For COPD, the problem may be amplified because exacerbations are defined in part by subjective, symptom-based criteria that depend on local practice patterns and cultural norms. Evidence from international randomised trials has documented a two- to three-fold difference in mean exacerbation rates across jurisdictions, driven by system-level factors such as quality of care, local clinical practices, medication availability and costs, and air quality.

To build the new version, the team used data from NOVELTY, a multi-national prospective observational study of roughly 12,000 asthma and COPD patients under routine primary and secondary care, described as the largest real-life study dedicated to asthma and COPD. The analysis sample comprised 5,181 patients aged 40 or older with physician-diagnosed COPD and at least 30 days of follow-up, drawn from 18 countries: Argentina, Australia, Brazil, Canada, Colombia, Denmark, France, Germany, Italy, Japan, Mexico, the Netherlands, Norway, South Korea, Spain, Sweden, the UK and the USA. Country-specific sample sizes ranged from 39 in Norway to 976 in the USA. The mean age at baseline was 66.5 years, 59.4% of the sample were male, and the mean follow-up was 357 days. Most patients had moderate-to-severe COPD, with 39.5% classified as GOLD 2 and 23.3% as GOLD 3 severity grades.

The primary outcome was the occurrence of any moderate-to-severe exacerbation over 12 months, defined according to the European Respiratory Society/American Thoracic Society definition as episodes of worsening respiratory symptoms requiring treatment with systemic steroids, antibiotics, or both. Exacerbations resulting in an emergency department visit or hospitalisation were considered severe; otherwise they were classified as moderate. The secondary outcome was the occurrence of any severe exacerbation over 12 months. The observed 12-month risk of moderate-to-severe exacerbations varied widely across countries, from 0.11 in South Korea to 0.46 in the UK and the Netherlands, with an overall mean of 0.26, while severe exacerbation risk ranged from 0.034 in South Korea to 0.21 in Colombia, with an overall mean of 0.08.

When externally validated in NOVELTY, ACCEPT 2.0 showed good discrimination but systematically overpredicted risk. The area under the receiver operating characteristic curve was 0.73 (95% CI 0.72–0.75) for moderate-to-severe exacerbations and 0.76 (95% CI 0.73–0.78) for severe exacerbations. Yet the mean predicted risks of 0.45 and 0.16 were substantially higher than the observed risks, yielding observed-to-expected ratios of 0.57 (95% CI 0.54–0.60) and 0.51 (95% CI 0.46–0.56) respectively. The calibration plots showed overprediction across the entire range of predicted risks. The authors suggest this likely reflected the fact that ACCEPT was developed in samples highly enriched for exacerbations, thereby overestimating risk in a real-world sample such as NOVELTY.

To correct this, the researchers recalibrated ACCEPT 2.0 by applying country-specific random intercepts and a fixed correction slope to predicted risks. They fitted a mixed-effects Cox model with observed survival time as the dependent variable and the transformed 12-month risk from ACCEPT 2.0 as the independent variable, with country included as a random-effect term to account for country-level variation in outcome risk. The random intercept for each country acts as a correction factor adjusting for systematic calibration error, while the coefficient of the transformed predicted risks serves as a single slope-correction factor identical across countries. For the secondary outcome, the same set of country-specific intercepts was used with a re-estimated slope, after exploratory analysis showed that re-estimating country effects separately for severe exacerbations provided no further improvement.

The recalibrated ACCEPT 3.0 showed similar discrimination, with an AUROC of 0.75 (95% CI 0.74–0.77) for moderate-to-severe exacerbations and 0.76 (95% CI 0.73–0.78) for severe exacerbations, but substantially improved calibration. The average predicted risks of 0.26 and 0.08 were close to the observed risks, producing observed-to-expected ratios of 0.99 (95% CI 0.94–1.05) and 0.99 (95% CI 0.90–1.09) for the primary and secondary outcomes respectively. In decision curve analysis, ACCEPT 3.0 showed consistently higher net benefit than ACCEPT 2.0 across the range of treatment thresholds, particularly for severe exacerbations. At thresholds of 5%, 10% and 20%, ACCEPT 3.0 achieved net benefits of 0.21, 0.18 and 0.12 for moderate-to-severe exacerbations and 0.044, 0.020 and 0.006 for severe exacerbations, compared with 0.21, 0.17 and 0.085 and 0.033, 0.008 and 0.003 for ACCEPT 2.0. These thresholds were chosen because a recent study reported a threshold of 12% on 12-month exacerbation risk for low-dose azithromycin therapy, with lower values representing potential thresholds for inhaled medications and higher values representing thresholds for more expensive therapies such as biologics.

Importantly, the team also developed a function mapping a given background exacerbation risk to an intercept adjustment term, extending the model to any setting where the average risk is known, not just the 18 countries in NOVELTY. The relationship between country-specific observed risks and the estimated random effects was approximately linear, with a p-value of the slope below 0.0001 and an adjusted R-squared of 0.877. This means, for example, that a hospital system could estimate its exacerbation risk from previous-year records and use the published algorithm to estimate its local intercept term. ACCEPT 3.0 is available on a publicly accessible webapp and through a web-based Application Programming Interface, and the webapp also integrates treatment effect estimates to guide personalised preventive treatment, for example for roflumilast and azithromycin.

The study has limitations. Exacerbation history and COPD treatment during the pre-index period were assessed by patient recall, and the low sample size in some countries, such as Denmark with 51, Mexico with 58 and Norway with 39 participants, may produce high uncertainty in country-specific corrections for those settings. The modified regression coefficients are based only on the countries in this study, and the authors recommend dedicated country-specific validation studies as the next step. The analysis did not account for the competing risk of death, which was low at 1.8% of the sample, nor time-varying effects, and recruitment of patients likely to complete three years of follow-up may have introduced a selection bias toward a healthier cohort. In a sensitivity analysis restricted to the 3,705 patients with spirometry-confirmed airflow obstruction, the same pattern held: AUROCs of 0.75 and 0.75 after recalibration, with calibration and net benefit both enhanced. The authors conclude that background exacerbation risks should be accounted for when adjusting predicted exacerbation risks in a new setting, and that ACCEPT 3.0 is expected to provide higher clinical utility across a range of treatment thresholds.

Subject of Research: Recalibration and external validation of the ACCEPT COPD exacerbation risk prediction model using multi-country observational data

Article Title: ACCEPT 3.0: Recalibrating ACute COPD Exacerbation Prediction Tool (ACCEPT) for global use

Article References: Lim, L. H. M., Mehareen, J., Adibi, A., Janson, C., Burgel, P.-R., Lee, T. Y., Müllerová, H., Varghese, D., Marshall, J., Bhatt, S. P., Sin, D., Chen, W., Sadatsafavi, M., & On behalf of the NOVELTY Scientific Community and NOVELTY study investigators (2026). ACCEPT 3.0: Recalibrating ACute COPD Exacerbation Prediction Tool (ACCEPT) for global use. Diagnostic and Prognostic Research, 10(1), Article 27. https://doi.org/10.1186/s41512-026-00242-x

Image Credits: AI Generated

DOI: 10.1186/s41512-026-00242-x

Keywords: COPD, exacerbation prediction, ACCEPT 3.0, risk prediction model, recalibration, NOVELTY study, external validation, net benefit, AUROC, country-specific risk, clinical decision making, prediction model transportability