BAYESIAN-OPTIMISED TABULAR ATTENTION LEARNING FOR FOLLOW-UP-INFORMED HEART FAILURE MORTALITY PREDICTION

Abstract

Heart failure mortality prediction remains challenging because routinely available clinical variables capture heterogeneous and interaction-dependent risk patterns. This study proposes a Bayesian-optimised deep neural framework for follow-up-informed mortality prediction using structured heart failure records. Five compact neural architectures were evaluated, including multilayer, residual, wide-and-deep, autoencoder-based, and lightweight tabular attention models. Hyperparameters were optimised using Optuna with a Tree-structured Parzen Estimator strategy, followed by repeated cross-validation, pooled out-of-fold evaluation, hold-out validation, calibration assessment, and diagnostic visualisation. The UCI Heart Failure Clinical Records dataset was used, with 299 patient records and death_event as the target variable. The proposed TabAttentionLite model achieved the strongest overall performance, with repeated cross-validation ROC-AUC of 0.8930 and pooled out-of-fold ROC-AUC of 0.8898. Hold-out validation produced ROC-AUC of 0.8570, indicating useful discrimination on unseen samples from the same dataset. These findings suggest that lightweight attention-based feature interaction modelling can improve follow-up-informed mortality-risk discrimination compared with the evaluated deep neural alternatives. Because follow-up time was included as a predictor, the model should be interpreted as a follow-up-informed prediction framework rather than a baseline-only clinical decision model.