Comparison of the performance of four clinical prediction rules for mortality in patients with COVID-19.

backgroundClinical prediction rules integrate clinical and laboratory variables to estimate outcomes, facilitating decision-making and optimizing resources, especially in high-demand settings. We aimed to validate and compare the performance of four mortality prediction scores -ISARIC-4C, CALL, SEIMC, q-CSI- in a Peruvian cohort of unvaccinated hospitalized COVID-19 pneumonia patients during the initial pandemic wave.

methodsWe performed a retrospective cohort study based on a secondary analysis of data from a previous study (March-December 2020). To ensure a robust and standardized head-to-head comparison, we utilized a complete-case analysis (n = 1,074). Selection bias was rigorously assessed by comparing the analytic sample with excluded patients. Each score's performance was evaluated using sensitivity, specificity, predictive values, likelihood ratios, area under the receiver operating characteristic curve (AUROC), and robust calibration metrics, including the calibration intercept (α) and the calibration slope (β). The ISARIC-4C score was used as an international reference standard for benchmarking.

resultsAmong 3,074 hospitalized patients, 1,074 had complete data for all four scores; no clinically significant differences were found between this group and excluded participants, indicating a representative sample. The cohort was mainly male (67.9%) with a median age of 59 years. The q-CSI score showed the best discrimination (AUROC 0.85, 95% CI: 0.83-0.87), significantly better than ISARIC-4C (0.82, 95% CI: 0.80-0.85), SEIMC (0.78, 95% CI: 0.75-0.81), and CALL (0.69, 95% CI: 0.66-0.72) (p < 0.0001). Direct comparison favored q-CSI over ISARIC-4C (p = 0.0016). While the Hosmer-Lemeshow (HL) test indicated a lack of fit for some models, robust calibration analysis confirmed accurate scaling (β for all scores was approx. 1.0) and adequate centering (α values with p > 0.05), supporting their clinical reliability.

conclusionsIn this Peruvian cohort, the q-CSI score exhibited the best predictive performance and highest feasibility for in-hospital mortality among patients with COVID-19 pneumonia. While the HL test indicated a lack of fit, the analysis of the calibration α and β confirmed that the models are globally well-calibrated, supporting their utility for risk stratification. However, local adjustment is still necessary prior to clinical use in our setting. These findings provide a valuable baseline for resource optimization in resource-limited setting during pandemic waves.