Work

Temporal Validation of Machine-Learning Models for Predicting Postoperative Atrial Fibrillation After Valve or Coronary Artery Bypass Surgery

Year: 2026

Type: preprint

Abstract: Background Postoperative atrial fibrillation (POAF) is common after valve or coronary artery bypass grafting (CABG) surgery, but prediction models often lack temporal validation and calibration assessment. We hypothesized that left atrial volume index (LAVI), as a marker of chronic atrial substrate, would be more transportable than perioperative stress variables and that regularized models would generalize better than complex algorithms under temporal shift. Methods We used a case-enriched retrospective development cohort of 347 patients (189 POAF events; May 2022–May 2025) and a prospective temporal cohort of 167 consecutive patients (44 events; June–December 2025). Eleven classifiers were assessed by repeated stratified five-fold cross-validation with fold-contained preprocessing, then frozen and applied to the temporal cohort. We evaluated discrimination, calibration, decision-curve net benefit, explainability, independent predictors, and six-month prognosis. Results LAVI was the strongest perioperative predictor (adjusted odds ratio per SD, 3.19; 95% CI, 1.66–6.12; P < 0.001). Random forest had the highest internal AUROC (0.827; 95% CI, 0.783–0.870), whereas the frozen preoperative ridge model performed best temporally (AUROC, 0.723; 95% CI, 0.639–0.808). The case-enriched development sample caused absolute-risk overprediction (mean predicted 0.442 vs observed 0.263; integrated calibration index [ICI], 0.173); intercept-only recalibration improved ICI to 0.048. Platt scaling fitted in the temporal cohort yielded an apparent ICI of 0.009 and was treated as exploratory updating. DCA indicates that within the decision threshold range of 0.10–0.50, the clinical benefit is greater than that of the "all-treatment/no-treatment" strategy. POAF was associated with six-month AF recurrence (7/43 vs 1/112; odds ratio, 21.58) and a higher composite-event hazard (adjusted hazard ratio, 3.22; 95% CI, 1.00–10.34; P = 0.050). Conclusions LAVI captured a transportable atrial-substrate signal, while the internally best nonlinear model did not retain superiority under temporal shift. The preoperative ridge model requires recalibration and independent multicenter validation before deployment. POAF identified persistent arrhythmic vulnerability through six months.

Source: Research Square

DOI: 10.21203/rs.3.rs-11295554/v1

Authors: Yalu Yu, Jiaqi Xia, Cheng Yang, Qin Jiang

Institutions: University of Electronic Science and Technology of China, Macau University of Science and Technology, Sichuan Academy of Medical Sciences & Sichuan Provincial People's Hospital

Language: English

Topic: Atrial Fibrillation Management and Outcomes

Subfield: Cardiology and Cardiovascular Medicine

Field: Medicine

Domain: Health Sciences

SDG: Good health and well-being

Open Access status: green

Funder: Science Fund for Distinguished Young Scholars of Sichuan Province