Development and validation of an interpretable machine learning model for early prediction of postoperative atrial fibrillation following on-pump cardiac surgery

Abstract

Background Postoperative atrial fibrillation (POAF) complicates 20%–40% of cardiac surgeries, increasing morbidity, length of stay, and healthcare costs. Early risk stratification could enable targeted prophylactic interventions. This study aimed to develop and validate an interpretable machine learning model for predicting incident POAF following on-pump cardiac surgery using routinely available perioperative parameters. Methods We conducted a retrospective cohort study of 1,054 adults undergoing on-pump cardiac surgery between January and December 2025. The development cohort (n = 870) was randomly partitioned into training (70%) and internal validation (30%) sets, with a temporal validation cohort (n = 184) recruited from a later period at the same institution. Least absolute shrinkage and selection operator (LASSO) regression identified optimal predictors from 56 candidate variables. Six machine learning algorithms were evaluated, with random forest selected based on discriminative performance. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) analysis. Results POAF incidence was 27.1%. LASSO regression identified 11 optimal predictors, including operation duration, age, oxygenation indices, lactate, and cholinesterase (ChE). The random forest model achieved area under the curve values of 0.725 (training), 0.711 (internal validation), and 0.740 (temporal validation), indicating acceptable discrimination with stable performance across cohorts. SHAP analysis revealed ChE and age as the predominant risk drivers, with exploratory, model-derived nonlinear thresholds identified: ChE inflection at approximately 4,000 U/L, accelerated age risk beyond 70 years, and J-shaped lactate risk escalation above 5 mmol/L. Conclusions This study presents an interpretable machine learning model for POAF prediction developed and temporally validated in a single-center cohort. The identification of ChE as a key predictive biomarker is hypothesis-generating and warrants further mechanistic and prospective investigation. The model may serve as a candidate decision-support tool for early ICU risk stratification, pending prospective multicenter validation of its generalizability and clinical utility.