Development of a composite variability metric for clinical application

Abstract

Methods: Our methodology involves taking advantage of an existing body of knowledge in the form of ODEs and using this to reason more effectively with the real-time data available at the bedside. A standard first-order Euler approximation of the system of ODEs is mapped directly to a DBN. Delta nodes capture the changes in the model variables in each time step. Model parameters are modeled as continuous gaussian nodes and allowed to vary over time to allow them adjust to the individual patient. The framework allows a clear distinction to be made between true values and measured values, thus accounting for data uncertainty. Results: We applied our methodology to track and predict glycemia levels in critically ill patients in receipt of intravenous glucose and insulin. By using the DBN framework, we account for uncertainty in the reactions of unstable patients and uncertainty in measurement errors. We evaluated our approach on real data for 12 critically ill patients. We compared our results to an existing approach whereby model parameters were reestimated using a nonlinear optimization algorithm. In 10 of 12 cases, the DBN approach outperformed the existing approach, very substantially so in some cases. Conclusions: Much knowledge of human physiology is formalized as systems of differential equations. We have devised a methodology for incorporating this knowledge in a DBN framework and demonstrated its effectiveness on a nonlinear system where evidence is sparse and infrequent. The methodology is used to predict a critically ill patient's plasma glucose levels in response to insulin and glucose infusions. With the data available, which are sporadic and may be inaccurate and incomplete, the DBN approach outperforms a previous approach demonstrating that the DBN method is effective at reestimating model parameters and reasoning with sparse and potentially unreliable data.