Structured vital sign prediction in hospital environments via an Al-Biruni earth radius optimization-driven unified...

Accurate prediction in structured hospital monitoring data is challenging because inpatient datasets are high-dimensional and often contain redundant features and suboptimal hyperparameter settings. This problem is important because unreliable prediction can limit the effectiveness of hospital monitoring and clinical decision support. To address this, this study proposes a unified optimization framework that integrates the Al-Biruni Earth Radius (BER) metaheuristic with the Feature-Transformed Learning Model (FTLM) for both binary feature selection and continuous hyperparameter optimization. BER is first applied in a discrete search space to identify informative subsets of vital-sign, demographic, clinical, and temporal variables from the Patient Vital Signs and Event Tracking dataset, and then in a continuous space to tune FTLM hyperparameters under the same computational budget used for competing optimizers, including GWO, PSO, BA, WAO, SBO, SCA, FA, GA, and SAO. At baseline, FTLM achieved a mean squared error (MSE) of 0.012028 and [Formula: see text] of 0.782413. After BER-based feature selection, performance improved to an MSE of [Formula: see text] and [Formula: see text] of 0.860654, with correlation of 0.848181 and Nash-Sutcliffe efficiency of 0.879577. Following BER-driven hyperparameter optimization, FTLM attained an MSE of [Formula: see text], RMSE of [Formula: see text], correlation coefficient of 0.955593543, [Formula: see text] of 0.961124043, and Willmott Index of 0.963281686, achieving the strongest empirical performance among the evaluated optimizers under the same experimental setting. To further assess generalizability, an external validation experiment was conducted on an independent Human Vital Sign Dataset containing 200,000 samples, where BER + FTLM again achieved the strongest empirical performance among the evaluated optimizers. These findings show that BER provides stable convergence, reduced variance, and strong predictive alignment for structured clinical data modeling.