IJCS | Volume 33, Nº2, March / April 2020

160 equation was used for construction of a weighted risk score; when transformed into odds ratios, the values were rounded to compose the score. Validation The preliminary risk score (obtained in the derivation cohort) was applied in the validation cohort, and the following performance statistics were applied: C statistics (area under the receiver operating characteristic [ROC] curve), the chi-square of goodness-of-fit of Hosmer- Lemeshow (HL) and consequently Pearson›s correlation coefficient between the observed events and those predicted by the model. Values for the area under the ROC curve greater than 0.70 indicate that the model has good discriminatory power. A chi-square of HL with p > 0.05 indicates good calibration of the model. APearson correlation coefficient r ≥ 0.7 indicates a strong association between observed and predicted values. Obtaining the final risk score After obtaining adequate performance in the validation, both databases (derivation and validation) were combined for the formulation of the final score. The variables were the same as those previously studied to achieve more accurate statistics for the coefficients. Performance statistics were obtained as described above. In addition to the final score, a logistic model (formula below) was generated, which allows direct estimates of the probability of outcome occurrence. The use of a mathematical model is considered by some authors to be more appropriate for obtaining event estimates since a complex formula would limit the use among physicians. In individuals considered to be at high risk in the additive model, the use of the logistic model is the most adequate in determining the individual occurrence of the clinical outcome. 15 P (event) = 1 {1 + exp (- ([30 + (31x1 +... + kak)) Data were processed and analyzed using the Statistical Package for the Social Sciences (SPSS), version 22.0. Ethical considerations This research project was submitted to the Research Ethics Committee of the ICIFUC (registration number 2345902). Results Characteristics In the total sample (n = 1,054), 272 patients had POAF (25.7%). When considering only myocardial revascularization surgeries (63.8%), the POAF rate was 20.3%. In cases of valve intervention alone (23.9%), the frequency of POAFwas 34.3%and, in combined surgeries (7.3%), the highest prevalence was observed – 36.6%. Mean age of the patients was 60.1 ± 12.1 years old, and 26.6% of the patients were 70 years old or older. Most of the patients (65.2%) were men (Table 1). POAF was associated with longer hospital stay compoaredwithpatientswithout POAF (medianof 10days vs. 7 days, respectively, p < 0.05) and increased in-hospital mortality (5.5%vs. 1.0%, respectively; p< 0.001). Inaddition, with a mean follow-up of five years, we observed higher late mortality rate for patients with POAF compared with thosewithout POAF (6.5%vs 1.4%, respectively, p = 0.002). Development of the risk model (derivation cohort) The multivariate analysis of the predictors in the derivation cohort (n = 448) is described in Table 2. Based on their statistical significance, the predictors selected for the construction of the score included age (≥ 70 years), mitral valvedisease, thenon-useor discontinuationof beta-blocker therapy and a positivewater balance greater than 1500mL. Points were assigned to each variable according to the odds ratio obtained (Table 2). The area under the ROC curve of the obtained model was 0.77 (95% confidence interval [CI] 0.73 to 0.81). Validation of the risk model External validation was performed in 606 patients of the validation cohort. The risk model had an accuracy of 0.78 (95% CI 0.73 to 0.82) measured by the area under the ROC curve, thus exhibiting a good discriminatory ability. Good correlation was noted between expected and observed POAF: r = 0.99, with chi-square = 1.73 (p = 0.94) (Hosmer-Lemeshow test). Risk model in the total sample (n = 1,054) The model was then edited using a combination of the developed score and data from the derivation Ronsoni et al. POAF risk score Int J Cardiovasc Sci. 2020; 33(2):158-166 Original Article