A set of 9 variables, derived from a clinical knowledge-based model, can be used to predict inpatient gout flares among patients hospitalized for comorbid gout, according to research results published in the Annals of the Rheumatic Diseases.
Using retrospective data from a cohort of patients hospitalized with comorbid gout, researchers sought to identify the causes of inpatient gout flare and develop a prediction model for inpatient gout flares among patients with comorbid gout.
The final analysis included data from 625 patients (78% men), of whom 87 experienced inpatient gout flare episodes. Patients who experienced gout flares had significantly longer hospital stays compared with patients who did not experience flares (median, 8 vs 2 days; P <.001).
Investigators utilized 3 approaches, which resulted in 3 variable sets associated with inpatient gout flare incidences. Model A was a clinical knowledge-driven model, Model B was a statistics-driven model, and Model C was a decision tree model. Across all 3 models, the discrimination and calibration were tested using C-statistics and calibration slope, respectively.
Finally, 4 variables (no preadmission urate-lowering therapy (ULT), ULT adjustment, diuretics adjustment, and preadmission urate >0.36 mmol/L) were present across all 3 models.
Models A and B demonstrated comparable discrimination (C-statistics, 0.82 and 0.81; 95% CI, 0.77-0.86 and 0.76-0.86, respectively), whereas Model C demonstrated slightly inferior discrimination (0.76; 95% CI, 0.71-0.82). Model A provided the most optimal model fit (calibration slope, 0.93; 95% CI, 0.33-1.52), with predicted probabilities of flare between the lower and upper calibration plot deciles of 1% and 48%, respectively.
In terms of practicality, investigators noted that Models A and C were simpler, compared with Model B, and contained fewer predictors that were “readily available in [a] hospital setting and did not require additional blood tests,” the researchers wrote. Model A was ultimately selected because of superior performance and simplicity.
Researchers also performed a 1000-sample bootstrap validation of Model A. The optimism-corrected C-statistic was 0.80 (95% CI, 0.78-0.88), which was comparable to the original C-statistic of 0.82. The optimism-corrected calibration slope was 0.78 (95% CI, 0.52-1.02), indicating that the final model showed a similar discrimination, but slightly lower calibration.
Study limitations included the limits inherent to the predictive nature of the model as well as the use of a retrospective dataset, as well as the relatively small sample size.
“The proposed predictors are practical and noninvasive, as they are readily available in [a] typical hospital care setting,” the researchers concluded. “The model may help clinicians identify hospitalized patients with comorbid gout who are at risk of developing flare.”
Disclosure: Several study authors declared affiliations with the pharmaceutical industry. Please see the original reference for a full list of authors’ disclosures.
Jatuworapruk K, Grainger R, Dalbeth N, Taylor WJ. Development of a prediction model for inpatient gout flares in people with comorbid gout [published online December 6, 2019]. Ann Rheum Dis. doi:10/1136/annrheumdis-2019-216277