Assessment of Transnational Performance of the Universal Cardiac Surgical Risk Score Across International Sites

DATA AND ANALYSIS Each site records the clinical variables from which the scores are computed, the site's own STS-PROM and EuroSCORE II values (comparators only; they never enter the UCSRS score), the Frailty Test components, and the 30-day outcome components. Sites compute no derived quantities; every derived value is computed centrally by the same code.

Data are de-identified at the site. A site either sends a de-identified patient-level dataset to the coordinating center or runs the identical standardized analysis locally and shares only summary results (c-statistic, observed/expected ratio, calibration slope and intercept), which are combined by summary-level meta-analysis. Discrimination uses the c-statistic with the DeLong test; calibration uses calibration plots, the observed/expected ratio, and calibration slope and intercept.

MODEL UNDER EVALUATION UCSRS version 3.2 (specification 3.2.0) has a single log-odds baseline computed from the patient's own clinical variables, to which it adds a required frailty correction (Frailty Test, 0-6: chair-rise, cognition assessment, hemoglobin, albumin) and optional corrections for liver function (MELD), left ventricular geometry and coronary complexity (SYNTAX), and right-heart catheterization. STS-PROM and EuroSCORE II are calculated independently by each site on the official calculators and are used only as comparators.

VERSION HISTORY Version 1.0 was published (J Cardiothorac Surg 2026; DOI 10.1186/s13019-026-04433-x), with a baseline blending STS-PROM and EuroSCORE II. Version 2.0 (27 August 2026) removed STS-PROM from the baseline. Version 2.1 (28 August 2026) rebuilt the baseline in log-odds. Version 3.0 replaced the blend with a single logistic baseline from which EuroSCORE II was removed entirely. Version 3.1 (20 September 2026) reset the calibration constant. Version 3.2 (24 October 2026) counted aortic disease once, lowered the MELD slopes, applied the frailty ladder to the odds, added CABG + AVR + MVR as an operation, and re-solved the calibration constant. Each change is recorded in a dated specification of record. The protocol registered here is ATLAS Study Protocol version 4.3 (24 October 2026), which describes UCSRS version 3.2.

SIMULATION AND CALIBRATION The calibration constants were set before enrollment against published mortality from thirteen surgical registries. For each registry, a synthetic cohort was tuned until its mean EuroSCORE II matched that registry's published observed mortality divided by its published EuroSCORE II observed/expected ratio; this is the only point at which real-world information enters. Runs scored millions of synthetic patients: about 11 million exploratory simulations in version 2.1 (which set no constant); 30,000,000-patient runs across the 13 registry-anchored cohorts in versions 3.0 and 3.1; and, in version 3.2, a 325,000-patient solve, a 520,000-patient hold-out, a 29,999,996-patient locking run and two 2,600,000-patient hold-outs. At the version 3.2 constant the median observed/expected ratio across the 13 registries was 1.0011 in the locking run, with hold-outs of 0.9968 and 0.9936. A separate 1,000,000-patient high-risk validation lowered mean predicted risk from 52.3% (version 3.1) to 20.8% (version 3.2) and the share of patients at the 70% score cap from 40.1% to 0.003%.

STRUCTURAL ENUMERATION On 15 September 2026 a complete factorial enumeration of 318,162,816 input combinations (cells, not patients) was run on a development version of the scoring engine, with a provisional intercept and no calibration shift. It covered elective first-time and redo operations and all eleven procedure codes on a deliberately adverse grid (MELD of 10 or more, hemoglobin 13 g/dL or less, albumin 4.0 g/dL or less). It was used to examine structure only: relative behavior, ordering, dynamic range and saturation at the score cap. Its absolute values are uncalibrated and are not risk estimates. In this grid EuroSCORE II, which carries a single flag for previous cardiac surgery and a single flag for thoracic aortic disease, separated a second redo sternotomy from a first redo by 0.02 percentage points and separated aortic root and arch operations from ascending replacement by 0.18 points; the UCSRS engine distinguished these operations. The results are recorded in a dated results document.

LIMITS OF THE SIMULATIONS No simulated patient has an outcome, and no outcome model was used at any stage. Discrimination therefore cannot be assessed in simulation, and the calibration results show internal consistency, not superiority over STS-PROM or EuroSCORE II. No performance claim accompanies the current specification. The version 3.2 constants were set by the investigator before first enrollment, as a documented exception to an earlier commitment; they are not moved again before 5,000 patients are enrolled.

The coordinating investigator developed UCSRS. This is declared in all study documents and publications, and negative results will be published.