Predicting One Year Mortality after Cardiac Surgery Complicated by Prolonged Critical Illness: Derivation and Validation of a Population-based Risk Model

Abstract Objectives Patients experiencing prolonged critical illness after cardiac surgery represent a resource intensive group with a high risk of mortality. We sought to derive and validate a multivariable model that accurately predicts one-year mortality in people who have been critically ill for at least one week after cardiac surgery. Design This was a retrospective population-based cohort study using linked administrative data Setting Eleven hospitals providing cardiac surgical care in the Canadian province of Ontario. Participants All adult patients aged ≥18 years undergoing one of the five most common major cardiac surgical procedures between April 1, 2009 to March 31, 2014. Interventions None Measurements and Main Results Our primary exposure was presence in an intensive care unit on the seventh postoperative day (POD7) and our primary outcome was all-cause mortality occurring after POD7 and within one year from date of surgery. Candidate predictors included patient demographics, surgical details, preoperative medical conditions, postoperative status and life supportive therapies utilized on POD7. Descriptive statistics were used to compare predictor variables between participants who did or did not die in the year after surgery. The prediction model was derived in the full data set using logistic regression and the pre-specified set of predictor variables. 2,031 individuals remained in an ICU on POD7 (4.8% of all cardiac surgery patients). 521 people died (25.6%) in the year after surgery; 353 died prior to hospital discharge (17.3% of total cohort, 67.8% of deaths). Requirement for multiple vasoactive or inotropic medications was the strongest predictor of mortality, followed by need for invasive ventilation, 3 or more preoperative comorbidities, need for a single inotropic or vasoactive medication, and requirement for dialysis prior to surgery. The derivation area under the curve (AUC) was 0.80 and the model was well-calibrated with a Hosmer-Lemeshow P-value of 0.5, and good calibration across risk deciles. Conclusions A prespecified multivariable model using clinically relevant, routinely available variables was able to accurately predict death amongst those with prolonged critical illness after cardiac surgery.