Dominic S. Lee

Implementation and evaluation of the SPRINT protocol for tight glycaemic control in critically ill patients: a clinical practice change

IntroductionStress-induced hyperglycaemia is prevalent in critical care. Control of blood glucose levels to within a 4.4 to 6.1 mmol/L range or below 7.75 mmol/L can reduce mortality and improve clinical outcomes. The Specialised Relative Insulin Nutrition Tables (SPRINT) protocol is a simple wheel-based system that modulates insulin and nutritional inputs for tight glycaemic control.MethodsSPRINT was implemented as a clinical practice change in a general intensive care unit (ICU). The objective of this study was to measure the effect of the SPRINT protocol on glycaemic control and mortality compared with previous ICU control methods. Glycaemic control and mortality outcomes for 371 SPRINT patients with a median Acute Physiology And Chronic Health Evaluation (APACHE) II score of 18 (interquartile range [IQR] 15 to 24) are compared with a 413-patient retrospective cohort with a median APACHE II score of 18 (IQR 15 to 23).ResultsOverall, 53.9% of all measurements were in the 4.4 to 6.1 mmol/L band. Blood glucose concentrations were found to be log-normal and thus log-normal statistics are used throughout to describe the data. The average log-normal glycaemia was 6.0 mmol/L (standard deviation 1.5 mmol/L). Only 9.0% of all measurements were below 4.4 mmol/L, with 3.8% below 4 mmol/L and 0.1% of measurements below 2.2 mmol/L. On SPRINT, 80% more measurements were in the 4.4 to 6.1 mmol/L band and standard deviation of blood glucose was 38% lower compared with the retrospective control. The range and peak of blood glucose were not correlated with mortality for SPRINT patients (P >0.30). For ICU length of stay (LoS) of greater than or equal to 3 days, hospital mortality was reduced from 34.1% to 25.4% (-26%) (P = 0.05). For ICU LoS of greater than or equal to 4 days, hospital mortality was reduced from 34.3% to 23.5% (-32%) (P = 0.02). For ICU LoS of greater than or equal to 5 days, hospital mortality was reduced from 31.9% to 20.6% (-35%) (P = 0.02). ICU mortality was also reduced but the P value was less than 0.13 for ICU LoS of greater than or equal to 4 and 5 days.ConclusionSPRINT achieved a high level of glycaemic control on a severely ill critical cohort population. Reductions in mortality were observed compared with a retrospective hyperglycaemic cohort. Range and peak blood glucose metrics were no longer correlated with mortality outcome under SPRINT.

Blood Glucose Prediction Using Stochastic Modeling in Neonatal Intensive Care

Hyperglycemia is a common metabolic problem in premature, low-birth-weight infants. Blood glucose homeostasis in this group is often disturbed by immaturity of endogenous regulatory systems and the stress of their condition in intensive care. A dynamic model capturing the fundamental dynamics of the glucose regulatory system provides a measure of insulin sensitivity (SI). Forecasting the most probable future SI can significantly enhance real-time glucose control by providing a clinically validated/proven level of confidence on the outcome of an intervention, and thus, increased safety against hypoglycemia. A 2-D kernel model of SI is fitted to 3567 h of identified, time-varying SI from retrospective clinical data of 25 neonatal patients with birth gestational age 23 to 28.9 weeks. Conditional probability estimates are used to determine SI probability intervals. A lag-2 stochastic model and adjustments of the variance estimator are used to explore the bias-variance tradeoff in the hour-to-hour variation of SI. The model captured 62.6% and 93.4% of in-sample SI predictions within the (25th-75th) and (5th-95th) probability forecast intervals. This overconservative result is also present on the cross-validation cohorts and in the lag-2 model. Adjustments to the variance estimator found a reduction to 10%-50% of the original value provided optimal coverage with 54.7% and 90.9% in the (25th-75th) and (5th-95th) intervals. A stochastic model of SI provided conservative forecasts, which can add a layer of safety to real-time control. Adjusting the variance estimator provides a more accurate, cohort-specific stochastic model of SI dynamics in the neonate.

Classifying algorithms for SIFT-MS technology and medical diagnosis

Selected Ion Flow Tube-Mass Spectrometry (SIFT-MS) is an analytical technique for real-time quantification of trace gases in air or breath samples. SIFT-MS system thus offers unique potential for early, rapid detection of disease states. Identification of volatile organic compound (VOC) masses that contribute strongly towards a successful classification clearly highlights potential new biomarkers. A method utilising kernel density estimates is thus presented for classifying unknown samples. It is validated in a simple known case and a clinical setting before-after dialysis. The simple case with nitrogen in Tedlar bags returned a 100% success rate, as expected. The clinical proof-of-concept with seven tests on one patient had an ROC curve area of 0.89. These results validate the method presented and illustrate the emerging clinical potential of this technology.

Development of a model-based clinical sepsis biomarker for critically ill patients

Sepsis occurs frequently in the intensive care unit (ICU) and is a leading cause of admission, mortality, and cost. Treatment guidelines recommend early intervention, however positive blood culture results may take up to 48 h. Insulin sensitivity (S(I)) is known to decrease with worsening condition and could thus be used to aid diagnosis. Some glycemic control protocols are able to accurately identify insulin sensitivity in real-time. Hourly model-based insulin sensitivity S(I) values were calculated from glycemic control data of 36 patients with sepsis. The hourly S(I) is compared to the hourly sepsis score (ss) for these patients (ss=0-4 for increasing severity). A multivariate clinical biomarker was also developed to maximize the discrimination between different ss groups. Receiver operator characteristic (ROC) curves for severe sepsis (ss ≥ 2) are created for both S(I) and the multivariate clinical biomarker. Insulin sensitivity as a sepsis biomarker for diagnosis of severe sepsis achieves a 50% sensitivity, 76% specificity, 4.8% positive predictive value (PPV), and 98.3% negative predictive value (NPV) at an S(I) cut-off value of 0.00013 L/mU/min. Multivariate clinical biomarker combining S(I), temperature, heart rate, respiratory rate, blood pressure, and their respective hourly rates of change achieves 73% sensitivity, 80% specificity, 8.4% PPV, and 99.2% NPV. Thus, the multivariate clinical biomarker provides an effective real-time negative predictive diagnostic for severe sepsis. Examination of both inter- and intra-patient statistical distribution of this biomarker and sepsis score shows potential avenues to improve the positive predictive value.

Development of a Model-Based Clinical Sepsis Biomarker for Critically Ill Patients

Abstract Sepsis occurs frequently in the intensive care unit (ICU) and is a leading cause of admission, mortality, and cost. Treatment guidelines recommend early intervention, however positive blood culture results may take up to 48 hours. Insulin sensitivity (S I ) is known to decrease with worsening condition and could thus be used to aid diagnosis. Some glycemic control protocols are able to accurately identify insulin sensitivity in real-time. Receiver operator characteristic (ROC) curves and cut-off S I values for sepsis diagnosis were calculated for real-time model-based insulin sensitivity from glycemic control data of 36 patients with sepsis. Patients were identified as having sepsis based on a clinically validated sepsis score (ss) of 2 or higher (ss = 0–4 for increasing severity). A clinical biomarker was calculated from patient clinical data to maximize the discrimination between cohorts. Insulin sensitivity as a sepsis biomarker for diagnosis of severe sepsis achieves a 50% sensitivity, 76% specificity, 4.8% PPV, and 98.3% NPV at a S I cut-off value of 0.00013 L * mU min −1 . A clinical biomarker combining S I , temperature, heart rate, respiratory rate, blood pressure, and their respective hourly rates of change achieves 73% sensitivity, 80% specificity, 8.4% PPV, and 99.2% NPV. Thus, a clinical biomarker provides an effective real-time negative predictive diagnostic for severe sepsis. Examination of both inter- and intra-patient statistical distribution of this biomarker and sepsis score show potential avenues to improve the positive predictive value.

Physiologically-based minimal model of agitation-sedation dynamics

Agitation-sedation cycling in critically ill patients, characterized by oscillations between states of agitation and over-sedation, damages patient health and increases length of stay and cost. The model presented captures the essential dynamics of the agitation-sedation system, is physiologically representative, and is validated by accurately simulating patient response for 37 critical care patients. The model provides a platform to develop and test controllers that offer the potential of improved agitation management.

A new model validation tool using kernel regression and density estimation

In physiological system modelling for control or decision support, model validation is a critical element. A nonparametric approach for assessing the validity of deterministic dynamic models against empirical data is developed, based on kernel regression and kernel density estimation, yielding visual graphical assessment tools as well as numerical metrics of compatibility between the model and the data. Nonparametric regression has been suggested for assessing a parametric statistical model by constructing a confidence band for the proposed model and then checking whether the nonparametric regression curve lies within the band. However, for deterministic models, there is no confidence band that can be constructed. A reversal of roles is therefore suggested--construct a probability band for the nonparametric regression curve and check whether the proposed model lies within the band. This approach extends the utility of nonparametric regression for model assessment to deterministic models. Weighted kernel density estimation is incorporated to derive a density profile for the regression curve, creating a local graphical validation tool. In addition, the density profile is used to define and compute two numerical measures--average normalized density (AND) and relative average normalized density (RAND), representing global statistical validity measures. These tools are demonstrated using a biomedical system model for agitation-sedation and sedation management control.

Posterior Expectation of Regularly Paved Random Histograms

We present a novel method for averaging a sequence of histogram states visited by a Metropolis-Hastings Markov chain whose stationary distribution is the posterior distribution over a dense space of tree-based histograms. The computational efficiency of our posterior mean histogram estimate relies on a statistical data-structure that is sufficient for nonparametric density estimation of massive, multidimensional metric data. This data-structure is formalized as statistical regular paving (SRP). A regular paving (RP) is a binary tree obtained by selectively bisecting boxes along their first widest side. SRP augments RP by mutably caching the recursively computable sufficient statistics of the data. The base Markov chain used to propose moves for the Metropolis-Hastings chain is a random walk that data-adaptively prunes and grows the SRP histogram tree. We use a prior distribution based on Catalan numbers and detect convergence heuristically. The performance of our posterior mean SRP histogram is empirically assessed for large sample sizes simulated from several multivariate distributions that belong to the space of SRP histograms.

Do complex models increase prediction of complex behaviours? Predicting driving ability in people with brain disorders

Prediction of complex behavioural tasks via relatively simple modelling techniques, such as logistic regression and discriminant analysis, often has limited success. We hypothesized that to more accurately model complex behaviour, more complex models, such as kernel-based methods, would be needed. To test this hypothesis, we assessed the value of six modelling approaches for predicting driving ability based on performance on computerized sensory–motor and cognitive tests (SMCTests™) in 501 people with brain disorders. The models included three models previously used to predict driving ability (discriminant analysis, DA; binary logistic regression, BLR; and nonlinear causal resource analysis, NCRA) and three kernel methods (support vector machine, SVM; product kernel density, PK; and kernel product density, KP). At the classification level, two kernel methods were substantially more accurate at classifying on-road pass or fail (SVM 99.6%, PK 99.8%) than the other models (DA 76%, BLR 78%, NCRA 74%, KP 81%). However, accuracy decreased substantially for all of the kernel models when cross-validation techniques were used to estimate prediction of on-road pass or fail in an independent referral group (SVM 73–76%, PK 72–73%, KP 71–72%) but decreased only slightly for DA (74–75%) and BLR (75–76%). Cross-validation of NCRA was not possible. In conclusion, while kernel-based models are successful at modelling complex data at a classification level, this is likely to be due to overfitting of the data, which does not lead to an improvement in accuracy in independent data over and above the accuracy of other less complex modelling techniques.

Long term verification of glucose-insulin regulatory system model dynamics

Hyperglycaemia in critically ill patients increases the risk of further complications and mortality. A long-term verification of a model that captures the essential glucose- and insulin-kinetics is presented, using retrospective data gathered in an intensive care unit (ICU). The model uses only two patient specific parameters, for glucose clearance and insulin sensitivity. The optimization of these parameters is accomplished through a novel integration-based fitting approach, and a piecewise linearization of the parameters. This approach reduces the non-linear, non-convex optimization problem to a simple linear equation system. The method was tested on long-term blood glucose recordings from 17 ICU patients, resulting in an average error of 7%, which is in the range of the sensor error. One-hour predictions of blood glucose data proved acceptable with an error range between 711%. These results verify the models ability to capture long-term observed glucose-insulin dynamics in hyperglycaemic ICU patients.