RCRI vs Gupta: comparing two perioperative cardiac risk scores

What the RCRI and the Gupta score each predict

The RCRI and the Gupta score predict overlapping but different outcomes from very different data. The RCRI, published by Lee and colleagues in Circulation in 1999, came from 4,315 patients aged 50 or older having elective major noncardiac surgery at one Boston teaching hospital. Its outcome, major cardiac complications, combined myocardial infarction, pulmonary edema, ventricular fibrillation or primary cardiac arrest, and complete heart block, and infarction was detected with CK-MB. The Gupta score, published by Gupta and colleagues in Circulation in 2011, came from the American College of Surgeons' National Surgical Quality Improvement Program (NSQIP), a prospective database from more than 250 hospitals. Of 211,410 patients in the 2007 data, 1,371 (0.65%) had a perioperative myocardial infarction or cardiac arrest, the only outcome it predicts. The model was then validated on 257,385 patients in the 2008 data. The Canadian guideline calls it the NSQIP MICA risk index.

How the RCRI and Gupta inputs differ

The RCRI and Gupta inputs differ in both number and kind. The RCRI asks six yes/no questions: is the surgery intraperitoneal, intrathoracic or suprainguinal vascular; does the patient have ischemic heart disease, congestive heart failure or cerebrovascular disease; are they on insulin before surgery; and is their serum creatinine above 2.0 mg/dL. Each yes is 1 point, so it can be added up at the bedside. The Gupta model uses the five predictors that multivariate analysis found in the NSQIP data: type of surgery, dependent functional status, abnormal creatinine, American Society of Anesthesiologists (ASA) class and increasing age. These are combined in a regression equation, so the result needs the online calculator rather than a tally. Two inputs set the scores apart. Gupta includes age, which the RCRI does not; in Lee's cohort advanced age did not correlate with complications. And Gupta includes ASA class and functional status, while the RCRI names specific cardiac, cerebrovascular, diabetic and renal conditions.

Because the RCRI needs no software, the RCRI calculator is simply six tick boxes that add up to the class.

Which score discriminates better

On discrimination, the Gupta score outperformed the RCRI in the data it was built on. Gupta and colleagues reported a C statistic, the area under the ROC curve, of 0.884 in the 2007 NSQIP derivation data and 0.874 in the 2008 validation data. When they applied the RCRI to the same 2008 NSQIP patients, its C statistic was 0.747, and they concluded that their calculator's predictive performance surpasses the RCRI. The RCRI's own figures vary with the setting. In Lee's prospective validation cohort its ROC area was 0.806. Ford and colleagues' 2010 systematic review of 24 studies and 792,740 patients pooled an AUC of 0.75 after mixed noncardiac surgery, falling to 0.64 after vascular surgery, and found the RCRI did not predict death well. A C statistic measures ranking, not calibration: it shows whether higher-risk patients get higher scores, not whether the predicted percentage matches the true rate, which is where the Canadian guideline's concern lies.

Why the Canadian guideline prefers the RCRI

The 2017 Canadian Cardiovascular Society (CCS) guideline, written by Duceppe and colleagues, prefers the RCRI despite the Gupta score's higher C statistic. Its recommendation 6 reads: when evaluating cardiac risk, we suggest clinicians use the RCRI over the other available clinical risk prediction scores, a conditional recommendation based on low-quality evidence. The panel accepted that the NSQIP MICA and the ACS NSQIP risk indices showed superior discrimination, but judged it highly probable that both underestimated cardiac risk, because NSQIP patients did not have systematic troponin measurements after surgery. The guideline notes that without cardiac biomarker screening more than half of all perioperative myocardial infarctions go undetected, which likely explains the low infarction counts in the NSQIP studies. It adds that neither NSQIP index had been externally validated in a study that systematically monitored troponin. The RCRI, by contrast, has been tested in many cohorts, and the panel pooled its risks from troponin-monitored external validation studies.

These pooled figures are higher than Lee's original rates (0.4%, 0.9%, 6.6% and 11.0% in the validation cohort). The guideline explains that the original study monitored CK-MB and excluded emergency surgery, while the later studies measured troponin, which is much more sensitive, and some included emergency operations.

A worked example with both scores

A worked example shows what each score needs and returns for the same patient. Take a 72-year-old who lives independently, had a stroke 3 years ago, has no heart disease or diabetes, has a creatinine of 1.1 mg/dL and is booked for an open hemicolectomy. For the RCRI, intraperitoneal surgery scores 1 and the prior stroke scores 1, so the total is 2 points, class III. Lee's validation cohort reported major cardiac complications in 6.6% of class III patients, and the CCS guideline's troponin-based estimate for 2 points is 10.1% for myocardial infarction, cardiac arrest or death at 30 days. For the Gupta score, the same patient's age, independent functional status, normal creatinine and colectomy all go into the model, but it also needs the ASA class, which the anesthesia team assigns. Its output is a single percentage for myocardial infarction or cardiac arrest only, so it cannot be set beside the RCRI rates as if they measured the same thing.

Tick the two boxes in the RCRI calculator to see the same class III result with both published risk figures.