Validation of a model-based virtual trials method for tight glycemic control in intensive care

Abstract

BackgroundIn-silico virtual patients and trials offer significant advantages in cost, time and safety for designing effective tight glycemic control (TGC) protocols. However, no such method has fully validated the independence of virtual patients (or resulting clinical trial predictions) from the data used to create them. This study uses matched cohorts from a TGC clinical trial to validate virtual patients and in-silico virtual trial models and methods.MethodsData from a 211 patient subset of the Glucontrol trial in Liege, Belgium. Glucontrol-A (N = 142) targeted 4.4-6.1 mmol/L and Glucontrol-B (N = 69) targeted 7.8-10.0 mmol/L. Cohorts were matched by APACHE II score, initial BG, age, weight, BMI and sex (p > 0.25). Virtual patients are created by fitting a clinically validated model to clinical data, yielding time varying insulin sensitivity profiles (SI(t)) that drives in-silico patients.Model fit and intra-patient (forward) prediction errors are used to validate individual in-silico virtual patients. Self-validation (tests A protocol on Group-A virtual patients; and B protocol on B virtual patients) and cross-validation (tests A protocol on Group-B virtual patients; and B protocol on A virtual patients) are used in comparison to clinical data to assess ability to predict clinical trial results.ResultsModel fit errors were small (<0.25%) for all patients, indicating model fitness. Median forward prediction errors were: 4.3, 2.8 and 3.5% for Group-A, Group-B and Overall (A+B), indicating individual virtual patients were accurate representations of real patients. SI and its variability were similar between cohorts indicating they were metabolically similar.Self and cross validation results were within 1-10% of the clinical data for both Group-A and Group-B. Self-validation indicated clinically insignificant errors due to model and/or clinical compliance. Cross-validation clearly showed that virtual patients enabled by identified patient-specific SI(t) profiles can accurately predict the performance of independent and different TGC protocols.ConclusionsThis study fully validates these virtual patients and in silico virtual trial methods, and clearly shows they can accurately simulate, in advance, the clinical results of a TGC protocol, enabling rapid in silico protocol design and optimization. These outcomes provide the first rigorous validation of a virtual in-silico patient and virtual trials methodology.

First pilot trial of the STAR-Liege protocol for tight glycemic control in critically ill patients

“…Virtual trial [16,17,23] was performed using the eight new stochastic models of Table 2 to better assess the increased variability in insulin sensitivity observed.…”

“…The potential impact of control performance of using a more specialized stochastic model was investigated in virtual trial re-simulations of the clinical trial. Tables 7 and 8 Differences between the clinical data and simulations using the original stochastic model arise from different nursing interventions and ability of the simulation environment to replicate the clinical trial [23]. During the pilot trial, slightly less insulin than specified by the controller was administered to patients in some cases if the nurses chose to override the recommendations (9 of 205, 4.4%), which explains the slightly higher BG levels for the clinical data.…”

“…Previous and current blood glucose measurements are used to identify a patient-specific current SI parameter value for the prior time interval [21]. This step accounts for inter-patient variability [17,22,23].…”

“…The insulin rate required to achieve this BG target is computed with a bisection method using a clinically validated insulin-glucose system model [23][24][25] and the median (50 th percentile)…”

“…Forecasting performance is assessed by the number of clinical results falling in an interquartile range (denoted as IQR, 50% confidence interval band) and 90% confidence interval band (expect ~50% and ~90% respectively). The different stochastic models are also assessed in clinically validated virtual trial re-simulations [23] of the trial to determine the potential impact on interventions given and glycemic outcomes.…”