Evolutionary Optimization of Extended Kalman Filter Hyperparameters for Adaptive Data-Driven ABP Prediction Framework Inspired by Cardiovascular Digital-Twin Principles

Continuous arterial blood pressure (ABP) prediction is relevant to cardiovascular monitoring and development of adaptive patient-specific digital twins. This study presents an adaptive data-driven ABP prediction framework inspired by cardiovascular digital-twin principles for causal one-step-ahead prediction using a single physiological signal modality. The proposed methodology combines a seven-sample causal ABP representation, a compact 7–5–1 multilayer perceptron, recursive Extended Kalman Filter (EKF) adaptation of all neural-network weights and biases, and offline evolutionary optimization of EKF covariance hyperparameters. Model development and evaluation used invasive ABP recordings from five CHARIS subjects available in PhysioNet, with five non-overlapping 100 s windows per subject. A rotating 3/1/1 subject-wise protocol assigned three subjects to supervised MLP training, one independent subject to covariance-hyperparameter selection, and one completely unseen subject to final testing in each fold. A fixed MLP achieved a mean test RMSE of 3.4123 mmHg. Recursive EKF adaptation using identical manually selected covariance hyperparameters across folds reduced mean RMSE to 3.0033 mmHg, corresponding to an average improvement of 11.98%. Offline optimization of $P_0$, $Q_0$, and $R_0$ further reduced mean RMSE to 2.1546 mmHg with Differential Evolution, 2.1548 mmHg with Particle Swarm Optimization, and 2.1546 mmHg with Genetic Algorithm, representing an overall improvement of approximately 35.9% relative to the fixed-MLP baseline. Despite differences in selected covariance values, all three evolutionary methods produced nearly identical final predictive performance. These results show that recursive neural-parameter adaptation and systematic covariance tuning provide complementary improvements in causal ABP prediction while preserving strict subject-level separation among training, hyperparameter selection, and final evaluation.