Methods & Research Background

Study design, model selection rationale, validation, and limitations.

1. Study Overview

Derivation Cohort

Çam ve Sakura

n=707 patients · 27 MACE events

External Validation Cohort

Siyami Ersek

n=378 patients · 38 MACE events

Endpoint: Major Adverse Cardiovascular Events (MACE) within 30 days of non-cardiac surgery, including cardiac death, non-fatal MI, and non-fatal stroke.

This is a retrospective, observational, two-center Turkish cohort study. ML models were trained on the derivation cohort and performance was assessed on the completely independent Siyami Ersek external validation cohort.

2. Model Selection Rationale

Ten ML classifiers were trained and externally validated. Two are presented as patient-facing probabilities based on the following hierarchy:

Model / Score	Role	AUROC	Brier	Cal. slope	NB@10%
HistGradientBoostingDisplayed	Study ML (default)	0.694	0.087	0.813	0.027
GradientBoostingDisplayed	Sensitivity ML	0.707	0.088	0.991	0.015
NaiveBayes	Highest AUROC — not deployed	0.738	0.090	2.139	0.000
AUB-HAS2Displayed	Best clinical benchmark	0.690	0.092	0.557	—
RCRIDisplayed	Legacy comparator	0.583	0.094	0.879	—

HistGradientBoosting (default ML): Best Brier score (0.087) and strongest net benefit at 10% and 15% decision thresholds — the most clinically actionable model for perioperative decisions.

GradientBoosting (sensitivity ML): Higher AUROC (0.707) and near-ideal calibration slope (0.991) — provides a second independent probability estimate for sensitivity analysis.

We do not claim ML superiority over AUB-HAS2. Both approaches are presented together for comparison.

3. Why NaiveBayes Is Not DeployedAUROC 0.738 — highest

NaiveBayes achieved the highest AUROC in external validation (0.738), which might suggest it is the best model. However, AUROC alone is insufficient to judge clinical utility.

Calibration slope: 2.139

Ideal = 1.0. A slope of 2.1 means the model dramatically overestimates risk for high-risk patients and underestimates for low-risk — predicted probabilities cannot be trusted as absolute risks.

Net benefit @ 10–15%: ≈0

Decision curve analysis shows zero net benefit at the clinically relevant 5–15% threshold range. Using this model would provide no clinical benefit over treating all or no patients.

O:E ratio: 2.033

Observed-to-expected ratio of 2.0 indicates severe miscalibration — the model predicts twice as many events as actually occurred.

Lesson: A high AUROC only means the model distinguishes high-risk from low-risk patients in rank order. It says nothing about whether the absolute probabilities are trustworthy. For shared decision-making and risk communication, calibration and net benefit matter more.

4. ECG-AI / PreOpNet Findings

PreOpNet is NOT included in this calculator.

ECG upload and PreOpNet predictions are explicitly excluded from patient-facing risk estimation in this tool.

Our study also investigated a digitized printed-ECG AI model (PreOpNet). Key findings:

ECG-AI probabilities derived from digitized printed ECGs were not calibrated absolute risk estimates.
External MACE discrimination was weak.
Clinical + ECG-AI models showed no incremental value over clinical variables alone.

These findings may be summarized on a separate research information page if needed, but ECG-AI predictions must not be presented as patient-facing MACE probabilities.

5. Clinical Score Definitions

AUB-HAS2 Score (0–6)

·+1 Heart disease history

·+1 Angina or dyspnea symptoms

·+1 Age ≥75 years

·+1 Hemoglobin <12 g/dL

·+1 Vascular surgery

·+1 Emergency surgery

0–1: Low risk

2–3: Intermediate risk

>3: High risk

RCRI Score (0–6)

·+1 High-risk surgery

·+1 Ischemic heart disease

·+1 Heart failure

·+1 Cerebrovascular disease

·+1 Insulin-treated diabetes

·+1 Creatinine >2 mg/dL

0: Low risk

1: Low risk

2: Elevated risk

≥3: High risk

Probabilities shown in the calculator use local derivation-cohort logistic calibration mappings when available, rather than originally published estimates. References: AUB-HAS2 (PMC7660845), RCRI (PubMed 10477528).

6. Limitations

→Two-center, retrospective observational study — findings may not generalize across all surgical populations or healthcare systems.
→Limited event count in derivation cohort (27 MACE events in 707 patients), which constrains model stability and generalizability.
→External calibration drift observed — absolute probabilities were systematically higher than actual event rates in the validation cohort.
→Several important variables (troponin, pro-BNP) had high missingness and were excluded as mandatory inputs.
→Research prototype only — this calculator is intended for research and educational purposes and does not replace guideline-based perioperative risk assessment.
→ML pipeline imputes missing values using training-cohort medians; predictions with many missing variables may be unstable.