BackgroundPerioperative mortality prediction in resource-limited settings is challenged by severe class imbalance (16.9:1) and heterogeneous complication pathways. Existing tools such as P-POSSUM require intraoperative variables unavailable before surgery and provide no uncertainty quantification. We present a prevalence-adaptive four-model Bayesian ensemble with entropy-based three-tier triage, trained on 697 real patients (39 deaths, raw prevalence 5.59%) from a 930-patient general surgical cohort, with class imbalance addressed by VAE augmentation (1,935-sample corpus; imbalance reduced from 16.9:1 to 1.94:1).
MethodsFour stochastic models -- a classifier Variational Autoencoder (VAE; AUC=0.9449), a Flipout Last Layer network (M1; AUC=0.9009), an all-probabilistic network (M2; AUC=0.9598), and a Monte Carlo Dropout network (Bayesian; AUC=0.9100) -- were trained on 67 preoperative and postoperative features. Class imbalance was addressed through VAE augmentation (1,935-sample corpus). Performance-normalised weights (Rokach 2010): VAE=0.2587, M1=0.2337, M2=0.2679, Bayesian=0.2398. A six-stage pipeline incorporated weighted base risk, a three-path majority-3 gate, Shannon entropy uncertainty quantification (gamma=0.10), and validated deployment thresholds. Monte Carlo inference used 100 passes with seed=42.
Pandey, A. K. et al. · CC-BY 4.0