Generative AI-Based Imputation to Preserve Data Fidelity and Enhance Outcome Prediction: A Multi-Institutional Study in Cardiac Surgery

Missing data in large-scale perioperative datasets can impair the accuracy, fairness, and transportability of predictive models in cardiac surgery, and no single imputation approach is uniformly optimal for heterogeneous clinical data. We present a multi-institutional benchmarking framework that jointly evaluates imputation fidelity and downstream clinical utility across classical methods (mean/mode, KNN, random-forest-based, MICE) and generative AI approaches (transformer- and diffusion-based models). Using two independent cohorts from Mass General Brigham and Maine Hospital (14,000$ adult cardiac surgery patients, 2014-2025), we simulate missingness at controlled rates (5%, 10%, 20% MCAR) across key perioperative variables and quantify reconstruction accuracy separately for categorical and continuous features. We then assess how imputation choices propagate to prediction of clinically relevant outcomes, including postoperative atrial fibrillation, ICU readmission, unresponsiveness status, in-hospital and 30-day mortality, and continuous perioperative targets (cardiopulmonary bypass time, aortic cross-clamp time, length of stay after surgery, and initial ICU hours). Across institutions, deep-learning imputers often provide strong fidelity, but top-performing methods vary by data type and missingness intensity; moreover, simple baselines that can appear competitive on fidelity (and are still commonly used) may yield weaker discrimination for some binary endpoints, motivating evaluation beyond reconstruction alone. By linking fidelity to clinical prediction across multiple imputation families, our framework provides practical guidance for perioperative analytics and is transferable to other clinical datasets where missingness threatens decision support.