A regularized system identification approach to subject-specific physiological modeling with limited data

This paper investigates a novel regularized system identification approach to physiological modeling using limited data. The proposed approach operates in two steps: 1) limited data from individual subjects are consolidated and leveraged to determine a population-average physiological model; then, 2) a subject-specific model for an individual subject is derived from a regularized system identification procedure whose objective is to reconcile the model's capability to predict individual-specific behavior and to retain typical population-representative trends. This is achieved by embedding a regularizing condition into the cost function for system identification that enforces parsimony in parametric deviation from the population-average model. A few unique advantages of the proposed approach are that 1) it offers superior predictive accuracy in both measured as well as unmeasured physiological system responses when compared to a standard system identification approach; and 2) it provides high-sensitivity parameters in the model associated with each individual subject, thus potentially eliminating the necessity for post-hoc parametric sensitivity analysis. Merits and limitations of the proposed regularized approach are illustrated with a real world case study on physiological modeling of hemodynamics in response to burn injury and resuscitation.