Accurate and precise prediction of insulin sensitivity variance in critically ill patients. (January 2018)

• Record Type: Journal Article
• Authors: Langdon, Ruby; Docherty, Paul D.; Mansell, Erin J.; Chase, J. Geoffrey
• Is Part Of: Biomedical signal processing and control. Volume 39(2018)
• Page Start: 327
• Page End: 335
• Publication Date: 2018-01
• Subjects: AIC Akaike information criterion -- GC glycaemic control -- ICU intensive care unit -- SI insulin sensitivity; Glycaemic control -- Akaike information criterion -- Critical care -- Model parameterisation -- Stochastic mapping; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2017.08.010
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Stochastic prediction distributions should match the occurrence distribution. Model parameterisation levels for optimal prediction precision were determined. Optimal precision could be used to increase confidence in glucose control algorithms. Slightly under parameterised glucose models had the best prediction precision. Abstract: Background: Glycaemic control (GC) in critical care can reduce mortality and improve clinical outcomes. Model based GC allows personalised and effective predictive control. Improving accuracy and precision of patient specific models, parameter identification, and resulting blood glucose predictions would enable more effective model-based GC algorithms. Methods: Glycaemic data from 30 critically ill patient episodes was used to fit a model of glucose dynamics. In this model, insulin sensitivity ( SI ) is identified through time using b-spline basis functions. The number of basis functions, M, relative to the number of data points in each data set, N, determines the level of parameterisation of the model. By fitting the model to all 30 data sets, the identified SI profiles could be used to build stochastic maps of SI changes over time. These maps show how SI is expected to change based on the current SI and the overall behaviour of this cohort. Thus, for a given time step into the future, SI prediction distributions can be found. Stochastic maps were made for varying basis function degree ( d ) ( d ∊ {0, 1, 2}), and ratio of M to N ( M : N ∊ {0.3, 0.65, 0.85, 1}). Using small subsets of each data set, many trials were carried out to compare SI prediction distributions to the distribution of the true SI values at the next measurement, for each combination of M and d . The main aim was to observe how the model parameterisation affected the predictive ability of the model. Outcomes from an Akaike Information Criterion (AIC) analysis were compared to the prediction analysis. AIC is a method of model comparison that assesses which of the given models provides the best trade-off between goodness-of-fit and model complexity, based on the expected measurement noise. For each basis function order, an AIC analysis was used to select the best M : N ratio of the four options tested. Results: Increasing the parameterisation of the model resulted in lower model-data residuals and thus wider stochastic maps, as SI changed at a faster rate to enable the model to more closely fit to the data. Therefore, increasing M : N resulted in wider prediction distributions. In all cases, when M : N = 1, the prediction distributions were too wide, and when M : N = 0.3 the prediction distributions were too narrow. Increasing the basis function order resulted in tighter prediction distributions, and allowed more accurate predictions to be made with a higher M : N ratio. The ratios that gave the most accurate predictions were M : N = 0.65 when d = 0, M : N = 0.85 when d = 1, and M : N = 1 when d = 2. In contrast, the AIC analysis found that an M : N ratio of approximately 0.45 was optimal in all cases. Conclusions: This study presented a new strategy for observing how the level of parameterisation and basis function order affects accuracy of future predictions of SI variability. Applications based on the findings of this research with a larger cohort could lead to improved predictive capability in GC algorithms. Accurate predictions would ultimately allow model-based GC algorithms to be implemented more effectively, increasing clinician confidence and improving patient outcomes.