Data sources

The Alberta Provincial Project for Outcomes Assessment in Coronary Heart Disease (APPROACH) is a registry of prospectively collected information on all patients undergoing cardiac catheterization and any subsequent cardiac procedures including cardiac surgery in the province of Alberta, Canada [16],[17]. The registry captures detailed individual patient demographic, medical, angiographic, surgical, and postoperative information. Preoperative patient demographic and medical variables and well as cardiac diagnostic and surgical procedural information are entered into the dataset by trained cardiac catheterization laboratory and operating room staff at the time of the procedure. Postoperative CVICU variables and complications are extracted from patient charts by trained chart abstracters using Society of Thoracic Surgeons (STS) database data definitions to identify complications occurring before index ICU discharge [18]. Abstracters also review and extract all previously entered pre- and intraoperative variables. A notable exception to these STS definitions is that neurologic complications were defined as an aggregate of all STS neurologic complications including any cerebrovascular accident, intracranial hemorrhage, coma ≥24 hours, encephalopathy, or paralysis. The latter composite definition was employed because the APPROACH dataset did not code for individual neurologic complications, only the incidence of any STS-defined neurological complication. Mortality is tracked through an Alberta Bureau of Vital Statistics data linkage [16]. This study, and a waiver of patient consent in this prospectively collected data registry, was approved by the University of Alberta Health Research Ethics Board (Pro00042669).

Study population

A total of 10,799 patients ≥18 years who underwent CABG and/or valvular surgery between January 2004 and December 2012 and were discharged from the CVICU alive from the University of Alberta were included in this analysis. Patients in our institution are admitted to the CVICU after CABG and valvular surgeries then discharged to a surgical postoperative ward when it was deemed clinically appropriate by the attending physician. For each surgical hospitalization, only patient, procedural, and postoperative variables recorded up to the time of the first CVICU discharge were evaluated in the prediction model. Excluded patients included: noncardiovascular surgical admissions, cardiac or pulmonary transplants, and isolated ventricular assist device or extracorporeal membrane oxygenation insertions.

Outcomes of interest

The primary outcome of interest was defined as any critical care unit readmission after CVICU discharge during the index surgical admission. Secondary outcomes of interest included hospital length of stay, in-hospital mortality, and one-year mortality.

Statistical analyses

Continuous variables were summarized using means and standard deviations and medians where appropriate. Categorical variables were summarized using percentages. Differences in categorical variables were tested using chi-square tests and differences in continuous variables were tested using Student t tests. The clinical prediction model was derived using the multivariable logistic regression model and CVICU readmission as the outcome variables. Independent variables of interest were entered in blocks. Individual covariates were entered and removed from the model with point estimates and model performance was evaluated with each step [19]. Perioperative covariates included in the model derivation included demographics (age, body mass index), medical history (hypertension, type 1 and type 2 diabetes, family history of heart disease, heart failure, New York Heart Association class, prior CABG, Canadian Cardiovascular Society class cerebrovascular disease, chronic liver disease, chronic lung disease, chronic renal disease, presurgical dialysis), preoperative coronary anatomy (one- and two-vessel disease, three-vessel disease, left main disease, not available), preoperative left ventricular ejection fraction (>50%, 35 to 50%, 20 to 34%, <20%, not available, not recorded due to hemodynamic instability), surgical priority (low risk, emergent, urgent in-hospital, urgent out of hospital), surgical procedure (isolated CABG, isolated valvular surgical repair or replacement, CABG and single-valve surgical repair or replacement, single-valve repair or replacement plus non-CABG surgery, multivalve repair or replacement), surgical incidence (first cardiac reoperation, ≥2 prior cardiac operations), cardiopulmonary bypass time, intraoperative red blood cell transfusion, intraoperative fresh frozen plasma, and postoperative CVICU complications (cardiac arrest, atrial fibrillation, cardiac tamponade, mechanical ventilation >24 hours, re-intubation, superficial sternal infection, deep sternal wound infection, sepsis, leg harvest site infection, urinary tract infection, acute respiratory distress syndrome, chest tube insertion, pneumonia, pleural effusion, positive sputum culture, pulmonary embolism, pulmonary edema, gastrointestinal bleed, mesenteric ischemia, renal failure, postoperative dialysis, delirium). The final derivative parsimonious model included all variables that remained independently significantly associated with CVICU readmission. For the purposes of validating the model, and estimating the population parameters, a bootstrapping technique was used. One thousand bootstrap samples with replacement were used to derive robust estimates of standard errors and confidence intervals for estimates of the odds ratio for all predictor variables that were independently associated with CVICU readmission. Scores were created by using the method described in the development of the Framingham risk scores [20]. Statistical analyses were performed using IBM SPSS statistics version 21 software (IBM Corp, Armonk, NY, USA).

Predictors of readmission

A total of 12 variables were independent predictors of intensive care unit readmission after cardiac surgery (Table 4). The final model included four preoperative variables (age ≥70 years, chronic lung disease, ejection fraction (EF) 35 to 50%, EF 20 to 34%, and EF <20%), two intraoperative variables (single-valve repair or replacement plus non-CABG surgery, multivalve repair or replacement), and seven postoperative complications (cardiac arrest, pneumonia, pleural effusion, deep sternal wound infection, leg graft harvest site infection, gastrointestinal bleed, and neurologic complications). The percentage of the total adjusted Wald χ2 for pre-, intra-, and postoperative independent variables were 13.8%, 3.1%, and 83.1%, respectively. The model demonstrated robust discrimination (c index = 0.799) and calibration (Hosmer-Lemeshow χ2 = 11.17, P = 0.192; Figure 1). The relationship between observed and expected readmission to intensive care across risk deciles is shown in Figure 1. The number of screened patients needed to identify (NNI) one readmission was calculated post hoc across each risk decile and the results were as follows: decile 1 (score 0 to 0.5, NNI = 80), decile 2 (score 1.0, NNI = 65), decile 3 (score 1.5 to 2.0, NNI = 57 ), decile 4 (score 2.5, NNI = 51 ), decile 5 (score 3.0, NNI = 46 ), decile 6 (score 3.5, NNI = 41), decile 7 (score 4.0 to 5.0 NNI = 36), decile 8 (score 5.5-8.5, NNI = 30), decile 9 (score 9.0 to 10.5, NNI = 19), decile 10 (score >11, NNI = 5).

Variable	Derivation cohort		Internal validation
	Wald χ2	OR (95% CI)	OR (95% CI)
Patient variables
Age ≥70 years, per 10 years	20.56	1.23 (1.11, 2.66)	1.23 (1.12, 1.36)
Chronic lung disease	14.40	1.49 (1.21, 1.83)	1.49 (1.21,1.87)
Ejection fraction (EF) a
EF 20-34%	6.03	1.64 (1.11, 2.44)	1.64 (1.05,2.37)
EF <20%	13.54	3.06 (1.69, 5.55	3.06 (1.49,5.35)
Surgical variables
Single-valve repair or replacement + non-CABG surgeryb	5.86	1.41 (1.07, 1.86)	1.41 (1.05,1.82)
Repair or replacements of ≥2 valvesb	6.2	2.13 (1.18, 3.86)	2.13 (1.00,3.88)
Postoperative CVICU variables
Cardiac arrest	65.06	4.04 (2.88, 5.66)	4.03 (2.69, 6.15)
Pneumonia	77.84	3.08 (2.40, 3.96)	3.08 (2.37, 4.09)
Pleural effusion	45.65	2.86 (2.11, 3.88)	2.86 (2.03, 3.83)
Deep sternal wound infection	48.29	6.58 (3.87, 11.18)	6.57 (3.33,13.11)
Leg graft harvest site infection	4.93	1.56 (1.05, 2.31)	1.56 (1.02,2.46)
Gastrointestinal bleed	61.45	4.67 (3.18, 6.86)	4.66 (2.93,7.22)
Neurologic complicationc	24.16	2.22 (1.61, 3.05)	2.21 (1.57, 3.17)

^aReference ejection fraction >50%; ^breference isolated coronary artery bypass; non-CABG procedures most commonly included left atrial appendage ligations, maze procedures, and atrial septal defect or patent foramen ovale closures; ^cneurologic complication defined as: cerebrovascular accident, intracranial hemorrhage, coma ≥24 hours, encephalopathy, or paralysis. CABG: coronary artery bypass grafting; CI: confidence interval; CVICU: cardiovascular intensive care unit; OR: odds ratio.

The stability of the 12 model variables was assessed in 1,000 bootstrap samples. We used the bootstrapping method to derive robust estimates of the standard errors of the odds ratios of the variables in the derivation model. All variables in the derivation model were retained in the bootstrapped model (Table 4). The confidence intervals estimated by the bootstrapped samples showed that the variables in the derivation model appropriately reflect the statistically independent predictors of the probability of readmission to the CVICU.

Risk score

The perioperative independent variables in the multivariable model were used to create the APPROACH CVICU readmission score (Figure 2A). The relationship between risk score and the predicted probability of intensive care readmission is also shown in Figure 2B which shows a corresponding rise in the predicted risk of CVICU readmission as the APPROACH CVICU risk score rises.

Limitations and strengths

The limitations of this analysis merit consideration. First, the study only included patients undergoing CABG and valvular surgeries; thus, the prediction model may not apply to patients undergoing other cardiac surgical procedures. The model was constructed on the most common cardiac surgeries in an effort to maximize external generalizability because procedures such as transplants, ventricular assist devices, and adult congenital surgeries are often conducted only in specialized centers. Second, the model requires validation in an external dataset; however, an internal validation with bootstrapping was applied to the derivation model. Third, no CVICU physiologic data or postoperative medication data were available in this dataset, though it should be noted that the discrimination and calibration of the model were excellent with available variables. Fourth, although the postoperative complications were extracted by trained abstracters using standardized definition into a registry module that codes for CVICU complications the dataset’s coding has not been independently validated.