J.G. Chase

Model-Based Insulin and Nutrition Administration for Tight Glycaemic Control in Critical Care

OBJECTIVE Present a new model-based tight glycaemic control approach using variable insulin and nutrition administration. BACKGROUND Hyperglycaemia is prevalent in critical care. Current published protocols use insulin alone to reduce blood glucose levels, require significant added clinical effort, and provide highly variable results. None directly address both the practical clinical difficulties and significant patient variation seen in general critical care, while also providing tight control. METHODS The approach presented manages both nutritional inputs and exogenous insulin infusions using tables simplified from a model-based, computerised protocol. Unique delivery aspects include bolus insulin delivery for safety and variable enteral nutrition rates. Unique development aspects include the use of simulated virtual patient trials created from retrospective data. The model, protocol development, and first 50 clinical case results are presented. RESULTS High qualitative correlation to within +/-10% between simulated virtual trials and published clinical results validates the overall approach. Pilot tests covering 7358 patient hours produced an average glucose of 5.9 +/- 1.1 mmol/L. Time in the 4-6.1 mmol/L band was 59%, with 84% in 4.0-7.0 mmol/L, and 92% in 4.0-7.75 mmol/L. The average feed rate was 63% of patient specific goal feed and the average insulin dose was 2.6U/hour. There was one hypoglycaemic measurement of 2.1 mmol/L. No departures from protocol or clinical interventions were required at any time. SUMMARY Modulating both low dose insulin boluses and nutrition input rates challenges the current practice of using only insulin in larger doses to reduce hyperglycaemic levels. Clinical results show very tight control in safe glycaemic bands. The approach could be readily adopted in any typical ICU.

Fast normalized cross correlation for motion tracking using basis functions

Digital image-based elasto-tomography (DIET) is an emerging method for non-invasive breast cancer screening. Effective clinical application of the DIET system requires highly accurate motion tracking of the surface of an actuated breast with minimal computation. Normalized cross correlation (NCC) is the most robust correlation measure for determining similarity between points in two or more images providing an accurate foundation for motion tracking. However, even using fast Fourier transform (FFT) methods, it is too computationally intense for rapidly managing several large images. A significantly faster method of calculating the NCC is presented that uses rectangular approximations in place of randomly placed landmark points or the natural marks on the breast. These approximations serve as an optimal set of basis functions that are automatically detected, dramatically reducing computational requirements. To prove the concept, the method is shown to be 37-150 times faster than the FFT-based NCC with the same accuracy for simulated data, a visco-elastic breast phantom experiment and human skin. Clinically, this approach enables thousands of randomly placed points to be rapidly and accurately tracked providing high resolution for the DIET system.

STAR Development and Protocol Comparison

Accurate glycemic control (AGC) is difficult due to excessive hypoglycemia risk. Stochastic TARgeted (STAR) glycemic control forecasts changes in insulin sensitivity to calculate a range of glycemic outcomes for an insulin intervention, creating a risk framework to improve safety and performance. An improved, simplified STAR framework was developed to reduce light hypoglycemia and clinical effort, while improving nutrition rates and performance. Blood glucose (BG) levels are targeted to 80-145 mg/dL, using insulin and nutrition control for 1-3 h interventions. Insulin changes are limited to +3U/h and nutrition to ±30% of goal rate (minimum 30%). All targets and rate change limits are clinically specified and generalizable. Clinically validated virtual trials were run on using clinical data from 371 patients (39841 h) from the Specialized Relative Insulin and Nutrition Tables (SPRINT) cohort. Cohort and per-patient results are compared to clinical SPRINT data, and virtual trials of three published protocols. Performance was measured as time within glycemic bands, and safety by patients with severe (BG <; 40 mg/dL) and mild (%BG <; 72 mg/dL) hypoglycemia. Pilot trial results from the first ten patients (1486 h) are included to support the in-silico findings. In both virtual and clinical trials, mild hypoglycemia was below 2% versus 4% for SPRINT. Severe hypoglycemia was reduced from 14 (SPRINT) to 6 (STAR), and 0 in the pilot trial. AGC was tighter than both SPRINT clinical data and in-silico comparison protocols, with 91% BG within the specified target (80-145 mg/dL) in virtual trials and 89.4% in pilot trials. Clinical effort (measurements) was reduced from 16.2/day to 11.8/day (13.5/day in pilot trials). This STAR framework provides safe AGC with significant reductions in hypoglycemia and clinical effort due to stochastic forecasting of patient variation - a unique risk-based approach. Initial pilot trials validate the in-silico design methods and resulting protocol, all of which can be generalized to suit any given clinical environment.

Model-based PEEP optimisation in mechanical ventilation

BackgroundAcute Respiratory Distress Syndrome (ARDS) patients require mechanical ventilation (MV) for breathing support. Patient-specific PEEP is encouraged for treating different patients but there is no well established method in optimal PEEP selection.MethodsA study of 10 patients diagnosed with ALI/ARDS whom underwent recruitment manoeuvre is carried out. Airway pressure and flow data are used to identify patient-specific constant lung elastance (Elung ) and time-variant dynamic lung elastance (Edrs ) at each PEEP level (increments of 5cmH2O), for a single compartment linear lung model using integral-based methods. Optimal PEEP is estimated using Elung versus PEEP, Edrs -Pressure curve and Edrs Area at minimum elastance (maximum compliance) and the inflection of the curves (diminishing return). Results are compared to clinically selected PEEP values. The trials and use of the data were approved by the New Zealand South Island Regional Ethics Committee.ResultsMedian absolute percentage fitting error to the data when estimating time-variant Edrs is 0.9% (IQR = 0.5-2.4) and 5.6% [IQR: 1.8-11.3] when estimating constant Elung . Both Elung and Edrs decrease with PEEP to a minimum, before rising, and indicating potential over-inflation. Median Edrs over all patients across all PEEP values was 32.2 cmH2O/l [IQR: 26.1-46.6], reflecting the heterogeneity of ALI/ARDS patients, and their response to PEEP, that complicates standard approaches to PEEP selection. All Edrs -Pressure curves have a clear inflection point before minimum Edrs , making PEEP selection straightforward. Model-based selected PEEP using the proposed metrics were higher than clinically selected values in 7/10 cases.ConclusionContinuous monitoring of the patient-specific Elung and Edrs and minimally invasive PEEP titration provide a unique, patient-specific and physiologically relevant metric to optimize PEEP selection with minimal disruption of MV therapy.

Model-based cardiac diagnosis of pulmonary embolism

A minimal cardiac model has been shown to accurately capture a wide range of cardiovascular system dynamics commonly seen in the intensive care unit (ICU). However, standard parameter identification methods for this model are highly non-linear and non-convex, hindering real-time clinical application. An integral-based identification method that transforms the problem into a linear, convex problem, has been previously developed, but was only applied on continuous simulated data with random noise. This paper extends the method to handle discrete sets of clinical data, unmodelled dynamics, a significantly reduced data set theta requires only the minimum and maximum values of the pressure in the aorta, pulmonary artery and the volumes in the ventricles. The importance of integrals in the formulation for noise reduction is illustrated by demonstrating instability in the identification using simple derivative-based approaches. The cardiovascular system (CVS) model and parameter identification method are then clinically validated on porcine data for pulmonary embolism. Errors for the identified model are within 10% when re-simulated and compared to clinical data. All identified parameter trends match clinically expected changes. This work represents the first clinical validation of these models, methods and approach to cardiovascular diagnosis in critical care.

Integral-based identification of patient specific parameters for a minimal cardiac model

A minimal cardiac model has been developed which accurately captures the essential dynamics of the cardiovascular system (CVS). However, identifying patient specific parameters with the limited measurements often available, hinders the clinical application of the model for diagnosis and therapy selection. This paper presents an integral-based parameter identification method for fast, accurate identification of patient specific parameters using limited measured data. The integral method turns a previously non-linear and non-convex optimization problem into a linear and convex identification problem. The model includes ventricular interaction and physiological valve dynamics. A healthy human state and four disease states, valvular stenosis, pulmonary embolism, cardiogenic shock and septic shock are used to test the method. Parameters for the healthy and disease states are accurately identified using only discretized flows into and out of the two cardiac chambers, the minimum and maximum volumes of the left and right ventricles, and the pressure waveforms through the aorta and pulmonary artery. These input values can be readily obtained non-invasively using echo-cardiography and ultra-sound, or invasively via catheters that are often used in Intensive Care. The method enables rapid identification of model parameters to match a particular patient condition in clinical real time (3-5 min) to within a mean value of 4-10% in the presence of 5-15% uniformly distributed measurement noise. The specific changes made to simulate each disease state are correctly identified in each case to within 10% without false identification of any other patient specific parameters. Clinically, the resulting patient specific model can then be used to assist medical staff in understanding, diagnosis and treatment selection.

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.

Model-based identification and diagnosis of a porcine model of induced endotoxic shock with hemofiltration

A previously validated cardiovascular system (CVS) model and parameter identification method for cardiac and circulatory disease states are extended and further validated in a porcine model (N=6) of induced endotoxic shock with hemofiltration. Errors for the identified model are within 10% when the model is re-simulated and compared to the clinical data. All identified parameter trends over time in the experiments match clinically expected changes both individually and over the cohort. This work represents a further clinical validation of these model-based cardiovascular diagnosis and therapy guidance methods for use with monitoring endotoxic disease states.

Structural Identifiability and Practical Applicability of an Alveolar Recruitment Model for ARDS Patients

Patient-specific mathematical models of respiratory mechanics can offer substantial insight into patient state and pulmonary dynamics that are not directly measurable. Thus, they offer significant potential to evaluate and guide patient-specific lung protective ventilator strategies for acute respiratory distress syndrome (ARDS) patients. To assure bedside applicability, the model must be computationally efficient and identifiable from the limited available data, while also capturing dominant dynamics and trends observed in ARDS patients. In this study, an existing static recruitment model is enhanced by considering alveolar distension and implemented in a novel time-continuous dynamic respiratory mechanics model. The model was tested for structural identifiability and a hierarchical gradient descent approach was used to fit the model to low-flow test responses of 12 ARDS patients. Finally, a comprehensive practical identifiability analysis was performed to evaluate the impact of data quality on the model parameters. Identified parameter values were physiologically plausible and very accurately reproduced the measured pressure responses. Structural identifiability of the model was proven, but practical identifiability analysis of the results showed a lack of convexity on the error surface indicating that successful parameter identification is currently not assured in all test sets. Overall, the model presented is physiologically and clinically relevant, captures ARDS dynamics, and uses clinically descriptive parameters. The patient-specific models show the ability to capture pulmonary dynamics directly relevant to patient condition and clinical guidance. These characteristics currently cannot be directly measured or established without such a validated model.

Unique parameter identification for cardiac diagnosis in critical care using minimal data sets

Lumped parameter approaches for modelling the cardiovascular system typically have many parameters of which a significant percentage are often not identifiable from limited data sets. Hence, significant parts of the model are required to be simulated with little overall effect on the accuracy of data fitting, as well as dramatically increasing the complexity of parameter identification. This separates sub-structures of more complex cardiovascular system models to create uniquely identifiable simplified models that are one to one with the measurements. In addition, a new concept of parameter identification is presented where the changes in the parameters are treated as an actuation force into a feed back control system, and the reference output is taken to be steady state values of measured volume and pressure. The major advantage of the method is that when it converges, it must be at the global minimum so that the solution that best fits the data is always found. By utilizing continuous information from the arterial/pulmonary pressure waveforms and the end-diastolic time, it is shown that potentially, the ventricle volume is not required in the data set, which was a requirement in earlier published work. The simplified models can also act as a bridge to identifying more sophisticated cardiac models, by providing an initial set of patient specific parameters that can reveal trends and interactions in the data over time. The goal is to apply the simplified models to retrospective data on groups of patients to help characterize population trends or un-modelled dynamics within known bounds. These trends can assist in improved prediction of patient responses to cardiac disturbance and therapy intervention with potentially smaller and less invasive data sets. In this way a more complex model that takes into account individual patient variation can be developed, and applied to the improvement of cardiovascular management in critical care.