A physiological Intensive Control Insulin-Nutrition-Glucose (ICING) model validated in critically ill patients

A methodology of phenotyping ICU patients from EHR data: High-fidelity, personalized, and interpretable phenotypes estimation

OBJECTIVE

Computing phenotypes that provide high-fidelity, time-dependent characterizations and yield personalized interpretations is challenging, especially given the complexity of physiological and healthcare systems and clinical data quality. This paper develops a methodological pipeline to estimate unmeasured physiological parameters and produce high-fidelity, personalized phenotypes anchored to physiological mechanics from electronic health record (EHR).

METHODS

A methodological phenotyping pipeline is developed that computes new phenotypes defined with unmeasurable computational biomarkers quantifying specific physiological properties in real time. Working within the inverse problem framework, this pipeline is applied to the glucose-insulin system for ICU patients using data assimilation to estimate an established mathematical physiological model with stochastic optimization. This produces physiological model parameter vectors of clinically unmeasured endocrine properties, here insulin secretion, clearance, and resistance, estimated for individual patient. These physiological parameter vectors are used as inputs to unsupervised machine learning methods to produce phenotypic labels and discrete physiological phenotypes. These phenotypes are inherently interpretable because they are based on parametric physiological descriptors. To establish potential clinical utility, the computed phenotypes are evaluated with external EHR data for consistency and reliability and with clinician face validation.

RESULTS

The phenotype computation was performed on a cohort of 109 ICU patients who received no or short-acting insulin therapy, rendering continuous and discrete physiological phenotypes as specific computational biomarkers of unmeasured insulin secretion, clearance, and resistance on time windows of three days. Six, six, and five discrete phenotypes were found in the first, middle, and last three-day periods of ICU stays, respectively. Computed phenotypic labels were predictive with an average accuracy of 89%. External validation of discrete phenotypes showed coherence and consistency in clinically observable differences based on laboratory measurements and ICD 9/10 codes and clinical concordance from face validity. A particularly clinically impactful parameter, insulin secretion, had a concordance accuracy of 83%±27%.

CONCLUSION

The new physiological phenotypes computed with individual patient ICU data and defined by estimates of mechanistic model parameters have high physiological fidelity, are continuous, time-specific, personalized, interpretable, and predictive. This methodology is generalizable to other clinical and physiological settings and opens the door for discovering deeper physiological information to personalize medical care.

Classification-based deep neural network vs mixture density network models for insulin sensitivity prediction problem

Model-based glycemic control (GC) protocols are used to treat stress-induced hyperglycaemia in intensive care units (ICUs). The STAR (Stochastic-TARgeted) glycemic control protocol - used in clinical practice in several ICUs in New Zealand, Hungary, Belgium, and Malaysia - is a model-based GC protocol using a patient-specific, model-based insulin sensitivity to describe the patient's actual state. Two neural network based methods are defined in this study to predict the patient's insulin sensitivity parameter: a classification deep neural network and a Mixture Density Network based method. Treatment data from three different patient cohorts are used to train the network models. Accuracy of neural network predictions are compared with the current model- based predictions used to guide care. The prediction accuracy was found to be the same or better than the reference. The authors suggest that these methods may be a promising alternative in model-based clinical treatment for patient state prediction. Still, more research is needed to validate these findings, including in-silico simulations and clinical validation trials.

How Electronic Medical Record Integration Can Support More Efficient Critical Care Clinical Trials

Large volumes of data are collected on critically ill patients, and using data science to extract information from the electronic medical record (EMR) and to inform the design of clinical trials represents a new opportunity in critical care research. Using improved methods of phenotyping critical illnesses, subject identification and enrollment, and targeted treatment group assignment alongside newer trial designs such as adaptive platform trials can increase efficiency while lowering costs. Some tools such as the EMR to automate data collection are already in use. Refinement of data science approaches in critical illness research will allow for better clinical trials and, ultimately, improved patient outcomes.

Interpretable physiological forecasting in the ICU using constrained data assimilation and electronic health record data

OBJECTIVE

Prediction of physiological mechanics are important in medical practice because interventions are guided by predicted impacts of interventions. But prediction is difficult in medicine because medicine is complex and difficult to understand from data alone, and the data are sparse relative to the complexity of the generating processes. Computational methods can increase prediction accuracy, but prediction with clinical data is difficult because the data are sparse, noisy and nonstationary. This paper focuses on predicting physiological processes given sparse, non-stationary, electronic health record data in the intensive care unit using data assimilation (DA), a broad collection of methods that pair mechanistic models with inference methods.

METHODS

A methodological pipeline embedding a glucose-insulin model into a new DA framework, the constrained ensemble Kalman filter (CEnKF) to forecast blood glucose was developed. The data include tube-fed patients whose nutrition, blood glucose, administered insulins and medications were extracted by hand due to their complexity and to ensure accuracy. The model was estimated using an individual's data as if they arrived in real-time, and the estimated model was run forward producing a forecast. Both constrained and unconstrained ensemble Kalman filters were estimated to compare the impact of constraints. Constraint boundaries, model parameter sets estimated, and data used to estimate the models were varied to investigate their influence on forecasting accuracy. Forecasting accuracy was evaluated according to mean squared error between the model-forecasted glucose and the measurements and by comparing distributions of measured glucose and forecast ensemble means.

RESULTS

The novel CEnKF produced substantial gains in robustness and accuracy while minimizing the data requirements compared to the unconstrained ensemble Kalman filters. Administered insulin and tube-nutrition were important for accurate forecasting, but including glucose in IV medication delivery did not increase forecast accuracy. Model flexibility, controlled by constraint boundaries and estimated parameters, did influence forecasting accuracy.

CONCLUSION

Accurate and robust physiological forecasting with sparse clinical data is possible with DA. Introducing constrained inference, particularly on unmeasured states and parameters, reduced forecast error and data requirements. The results are not particularly sensitive to model flexibility such as constraint boundaries, but over or under constraining increased forecasting errors.

A methodology of phenotyping ICU patients from EHR data: high-fidelity, personalized, and interpretable phenotypes estimation

Objective

Computing phenotypes that provide high-fidelity, time-dependent characterizations and yield personalized interpretations is challenging, especially given the complexity of physiological and healthcare systems and clinical data quality. This paper develops a methodological pipeline to estimate unmeasured physiological parameters and produce high-fidelity, personalized phenotypes anchored to physiological mechanics from electronic health record (EHR).

Methods

A methodological phenotyping pipeline is developed that computes new phenotypes defined with unmeasurable computational biomarkers quantifying specific physiological properties in real time. Working within the inverse problem framework, this pipeline is applied to the glucose-insulin system for ICU patients using data assimilation to estimate an established mathematical physiological model with stochastic optimization. This produces physiological model parameter vectors of clinically unmeasured endocrine properties, here insulin secretion, clearance, and resistance, estimated for individual patient. These physiological parameter vectors are used as inputs to unsupervised machine learning methods to produce phenotypic labels and discrete physiological phenotypes. These phenotypes are inherently interpretable because they are based on parametric physiological descriptors. To establish potential clinical utility, the computed phenotypes are evaluated with external EHR data for consistency and reliability and with clinician face validation.

Results

The phenotype computation was performed on a cohort of 109 ICU patients who received no or short-acting insulin therapy, rendering continuous and discrete physiological phenotypes as specific computational biomarkers of unmeasured insulin secretion, clearance, and resistance on time windows of three days. Six, six, and five discrete phenotypes were found in the first, middle, and last three-day periods of ICU stays, respectively. Computed phenotypic labels were predictive with an average accuracy of 89%. External validation of discrete phenotypes showed coherence and consistency in clinically observable differences based on laboratory measurements and ICD 9/10 codes and clinical concordance from face validity. A particularly clinically impactful parameter, insulin secretion, had a concordance accuracy of 83% ± 27%.

Conclusion

The new physiological phenotypes computed with individual patient ICU data and defined by estimates of mechanistic model parameters have high physiological fidelity, are continuous, time-specific, personalized, interpretable, and predictive. This methodology is generalizable to other clinical and physiological settings and opens the door for discovering deeper physiological information to personalize medical care.

A simple modeling framework for prediction in the human glucose-insulin system

Forecasting blood glucose (BG) levels with routinely collected data is useful for glycemic management. BG dynamics are nonlinear, complex, and nonstationary, which can be represented by nonlinear models. However, the sparsity of routinely collected data creates parameter identifiability issues when high-fidelity complex models are used, thereby resulting in inaccurate forecasts. One can use models with reduced physiological fidelity for robust and accurate parameter estimation and forecasting with sparse data. For this purpose, we approximate the nonlinear dynamics of BG regulation by a linear stochastic differential equation: we develop a linear stochastic model, which can be specialized to different settings: type 2 diabetes mellitus (T2DM) and intensive care unit (ICU), with different choices of appropriate model functions. The model includes deterministic terms quantifying glucose removal from the bloodstream through the glycemic regulation system and representing the effect of nutrition and externally delivered insulin. The stochastic term encapsulates the BG oscillations. The model output is in the form of an expected value accompanied by a band around this value. The model parameters are estimated patient-specifically, leading to personalized models. The forecasts consist of values for BG mean and variation, quantifying possible high and low BG levels. Such predictions have potential use for glycemic management as part of control systems. We present experimental results on parameter estimation and forecasting in T2DM and ICU settings. We compare the model's predictive capability with two different nonlinear models built for T2DM and ICU contexts to have a sense of the level of prediction achieved by this model.

Model-based subcutaneous insulin for glycemic control of pre-term infants in the neonatal intensive care unit

Hyperglycaemia is a common problem in neonatal intensive care units (NICUs). Achieving good control can result in better outcomes for patients. However, good control is difficult, where poor control and resulting hypoglycaemia reduces outcomes and confounds results. Clinically validated models can provide good control, and subcutaneous insulin delivery can provide more options for insulin therapy for clinicians. However, this combination has only been significantly utilised in adult outpatient diabetes, but could hold benefit for treating NICU infants. This research combines a well-validated NICU metabolic model with subcutaneous insulin kinetics models to assess the feasibility of a model-based approach. Clinical data from 12 very/extremely pre-mature infants was collected for an average study duration of 10.1 days. Blood glucose, interstitial and plasma insulin, as well as subcutaneous and local insulin were modelled, and patient-specific insulin sensitivity profiles were identified for each patient. Modelling error was low, where the cohort median [IQR] mean percentage error was 0.8 [0.3 3.4] %. For external validation, insulin sensitivity was compared to previous NICU cohorts using the same metabolic model, where overall levels of insulin sensitivity were similar. Overall, the combined system model accurately captured observed glucose and insulin dynamics, showing the potential for a model-based approach to glycaemic control using subcutaneous insulin in this cohort. The results justify further model validation and clinical trial research to explore a model-based protocol.

A physical activity-intensity driven glycemic model for type 1 diabetes

BACKGROUND AND OBJECTIVE

The glucose response to physical activity for a person with type 1 diabetes (T1D) depends upon the intensity and duration of the physical activity, plasma insulin concentrations, and the individual physical fitness level. To accurately model the glycemic response to physical activity, these factors must be considered.

METHODS

Several physiological models describing the glycemic response to physical activity are proposed by incorporating model terms proportional to the physical activity intensity and duration describing endogenous glucose production (EGP), glucose utilization, and glucose transfer from the plasma to tissues. Leveraging clinical data of T1D during physical activity, each model fit is assessed.

RESULTS

The proposed model with terms accommodating EGP, glucose transfer, and insulin-independent glucose utilization allow for an improved simulation of physical activity glycemic responses with the greatest reduction in model error (mean absolute percentage error: 16.11 ± 4.82 vs. 19.49 ± 5.87, p = 0.002).

CONCLUSIONS

The development of a physiologically plausible model with model terms representing each major contributor to glucose metabolism during physical activity can outperform traditional models with physical activity described through glucose utilization alone. This model accurately describes the relation of plasma insulin and physical activity intensity on glucose production and glucose utilization to generate the appropriately increasing, decreasing or stable glucose response for each physical activity condition. The proposed model will enable the in silico evaluation of automated insulin dosing algorithms designed to mitigate the effects of physical activity with the appropriate relationship between the reduction in basal insulin and the corresponding glycemic excursion.

Estimating Enhanced Endogenous Glucose Production in Intensive Care Unit Patients with Severe Insulin Resistance

BACKGROUND

Critically ill ICU patients frequently experience acute insulin resistance and increased endogenous glucose production, manifesting as stress-induced hyperglycemia and hyperinsulinemia. STAR (Stochastic TARgeted) is a glycemic control protocol, which directly manages inter- and intra- patient variability using model-based insulin sensitivity (SI). The model behind STAR assumes a population constant for endogenous glucose production (EGP), which is not otherwise identifiable.

OBJECTIVE

This study analyses the effect of estimating EGP for ICU patients with very low SI (severe insulin resistance) and its impact on identified, model-based insulin sensitivity identification, modeling accuracy, and model-based glycemic clinical control.

METHODS

Using clinical data from 717 STAR patients in 3 independent cohorts (Hungary, New Zealand, and Malaysia), insulin sensitivity, time of insulin resistance, and EGP values are analyzed. A method is presented to estimate EGP in the presence of non-physiologically low SI. Performance is assessed via model accuracy.

RESULTS

Results show 22%-62% of patients experience 1+ episodes of severe insulin resistance, representing 0.87%-9.00% of hours. Episodes primarily occur in the first 24 h, matching clinical expectations. The Malaysian cohort is most affected. In this subset of hours, constant model-based EGP values can bias identified SI and increase blood glucose (BG) fitting error. Using the EGP estimation method presented in these constrained hours significantly reduced BG fitting errors.

CONCLUSIONS

Patients early in ICU stay may have significantly increased EGP. Increasing modeled EGP in model-based glycemic control can improve control accuracy in these hours. The results provide new insight into the frequency and level of significantly increased EGP in critical illness.

The Impact of Exogenous Insulin Input on Calculating Hepatic Clearance Parameters

OBJECTIVE

Model-based metabolic tests require accurate identification of subject-specific parameters from measured assays. Insulin assays are used to identify insulin kinetics parameters, such as general and first-pass hepatic clearances. This study assesses the impact of intravenous insulin boluses on parameter identification precision.

METHOD

Insulin and C-peptide data from two intravenous glucose tolerance test (IVGTT) trials of healthy adults (N = 10 × 2; denoted A and B), with (A) and without (B) insulin modification, were used to identify insulin kinetics parameters using a grid search. Monte Carlo analysis (N = 1000) quantifies variation in simulation error for insulin assay errors of 5%. A region of parameter values around the optimum was identified whose errors are within variation due to assay error. A smaller optimal region indicates more precise practical identifiability. Trial results were compared to assess identifiability and precision.

RESULTS

Trial B, without insulin modification, has optimal parameter regions 4.7 times larger on average than Trial A, with 1-U insulin bolus modification. Ranges of optimal parameter values between trials A and B increase from 0.04 to 0.12 min^-1 for hepatic clearance and from 0.07 to 0.14 for first-pass clearance on average. Trial B's optimal values frequently lie outside physiological ranges, further indicating lack of distinct identifiability.

CONCLUSIONS

A small 1-U insulin bolus improves identification of hepatic clearance parameters by providing a smaller region of optimal parameter values. Adding an insulin bolus in metabolic tests can significantly improve identifiability and outcome test precision. Assay errors necessitate insulin modification in clinical tests to ensure identifiability and precision.

Determining Losses in Jet Injection Subcutaneous Insulin Delivery: A Model-Based Approach

OBJECTIVE

Accurate, safe glycemic management requires reliable delivery of insulin doses. Insulin can be delivered subcutaneously for action over a longer period of time. Needle-free jet injectors provide subcutaneous (SC) delivery without requiring needle use, but the volume of insulin absorbed varies due to losses associated with the delivery method. This study employs model-based methods to determine the expected proportion of active insulin present from a needle-free SC dose.

METHODS

Insulin, C-peptide, and glucose assay data from a frequently sampled insulin-modified oral glucose tolerance test trial with 2U SC insulin delivery, paired with a well-validated metabolic model, predict metabolic outcomes for N = 7 healthy adults. Subject-specific nonlinear hepatic clearance profiles are modeled over time using third-order basis splines with knots located at assay times. Hepatic clearance profiles are constrained within a physiological rate of change, and relative to plasma glucose profiles. Insulin loss proportions yielding optimal insulin predictions are then identified, quantifying delivery losses.

RESULTS

Optimal parameter identification suggests losses of up to 22% of the nominal 2U SC dose. The degree of loss varies between subjects and between trials on the same subject. Insulin fit accuracy improves where loss greater than 5% is identified, relative to where delivery loss is not modeled.

CONCLUSIONS

Modeling shows needle-free SC jet injection of a nominal dose of insulin does not necessarily provide metabolic action equivalent to total dose, and this availability significantly varies between trials. By quantifying and accounting for variability of jet injection insulin doses, better glycemic management outcomes using SC jet injection may be achieved.

Estimating Increased EGP During Stress Response in Critically Ill Patients

BACKGROUND

Stress-induced hyperglycemia is frequently experienced by critically ill patients and the use of glycemic control (GC) has been shown to improve patient outcomes. For model-based approaches to GC, it is important to understand and quantify model parameter assumptions. This study explores endogenous glucose production (EGP) and the use of a population-based parameter value in the intensive care unit context.

METHOD

Hourly insulin sensitivity (SI) was fit to clinical data from 145 patients on the Specialized Relative Insulin and Nutrition Titration GC protocol for at least 24 hours. Constraint of SI at a lower bound was used to explore likely EGP variability due to stress response. Minimum EGP was estimated during times when the model SI was constrained, and time and duration of events were examined.

RESULTS

Constrained events occur for 1.6% of patient hours. About 70% of constrained events occur in the first 12 hours and most events (~80%) occur when there is no exogenous nutrition given. Enhanced EGP values ranged from 1.16 mmol/min (current population value) to 2.75 mmol/min, with most being below 1.5 mmol/min (21% increase).

CONCLUSION

The frequency of constrained events is low and the current population value of 1.16 mmol/min is sufficient for more than 98% of patient hours, however, some patients experience significantly raised EGP probably due to an extreme stress response early in patient stay.

The development of a glucose prediction model in critically ill patients

PURPOSE

The aim of the current study is to develop a prediction model for glucose levels applicable for all patients admitted to the ICU with an expected ICU stay of at least 24 h. This model will be incorporated in a closed-loop glucose system to continuously and automatically control glucose values.

METHODS

Data from a previous single-center randomized controlled study was used. All patients received a FreeStyle Navigator II subcutaneous CGM system from Abbott during their ICU stay. The total dataset was randomly divided into a training set and a validation set. A glucose prediction model was developed based on historical glucose data. Accuracy of the prediction model was determined using the Mean Squared Difference (MSD), the Mean Absolute Difference (MAD) and a Clarke Error Grid (CEG).

RESULTS

The dataset included 94 ICU patients with a total of 134,673 glucose measurements points that were used for modelling. MSD was 0.410 ± 0.495 for the model, the MAD was 5.19 ± 2.63 and in the CEG 99.8% of the data points were in the clinically acceptable regions.

CONCLUSION

In this study a glucose prediction model for ICU patients is developed. This study shows that it is possible to accurately predict a patient's glucose 30 min ahead based on historical glucose data. This is the first step in the development of a closed-loop glucose system.

High Inter-Patient Variability in Sepsis Evolution: A Hidden Markov Model Analysis

BACKGROUND

Severe sepsis and septic shock are common in the intensive care unit (ICU) and contribute significantly to cost and mortality. Early treatment is critical but is confounded by the difficulty of real-time diagnosis. This study uses hidden Markov models (HMMs) to examine whether the time evolution of sepsis can add diagnostic accuracy or value using a proven set of bio-signals.

METHODS

Clinical data (N=36 patients; 6071 hours), including an hourly personalised insulin sensitivity metric. A two hidden state HMM is created to discriminate diagnosed cases (Severe Sepsis, Septic Shock) from controls (SIRS, Sepsis) states. Diagnostic performance is measured by ROC curves, likelihood ratios (LHRs), sensitivity/specificity, and diagnostic odds-ratios (DOR), for a best-case resubstitution estimate and a worst-case 80/20% repeated holdout analysis.

RESULTS

The HMM delivered near perfect results (95% Sensitivity; 96% Specificity) for best-case resubstitution estimates, but was comparatively poor (59% Sensitivity; 61% Specificity) for worst-case repeated holdout estimations. Adding the time evolution of sepsis did not add to the accuracy of diagnosis from using the signals alone without time history.

CONCLUSIONS

These potentially surprising results indicate significant inter-patient variability in the time evolution of sepsis, preventing effective diagnosis in the context of the bio-signals, data, and HMM topology used. Efforts for improved real-time, early sepsis diagnosis should concentrate on the robustness and efficacy of the bio-signals and data used, as well as the level of model complexity, to create more effective real-time classifiers.

Insulin sensitivity in critically ill patients: are women more insulin resistant?

Background

Glycaemic control (GC) in intensive care unit is challenging due to significant inter- and intra-patient variability, leading to increased risk of hypoglycaemia. Recent work showed higher insulin resistance in female preterm neonates. This study aims to determine if there are differences in inter- and intra-patient metabolic variability between sexes in adults, to gain in insight into any differences in metabolic response to injury. Any significant difference would suggest GC and randomised trial design should consider sex differences to personalise care.

Methods

Insulin sensitivity (SI) levels and variability are identified from retrospective clinical data for men and women. Data are divided using 6-h blocks to capture metabolic evolution over time. In total, 91 male and 54 female patient GC episodes of minimum 24 h are analysed. Hypothesis testing is used to determine whether differences are significant (P < 0.05), and equivalence testing is used to assess whether these differences can be considered equivalent at a clinical level. Data are assessed for the raw cohort and in 100 Monte Carlo simulations analyses where the number of men and women are equal.

Results

Demographic data between females and males were all similar, including GC outcomes (safety from hypoglycaemia and high (> 50%) time in target band). Females had consistently significantly lower SI levels than males, and this difference was not clinically equivalent. However, metabolic variability between sexes was never significantly different and always clinically equivalent. Thus, inter-patient variability was significantly different between males and females, but intra-patient variability was equivalent.

Conclusion

Given equivalent intra-patient variability and significantly greater insulin resistance, females can receive the same benefit from safe, effective GC as males, but may require higher insulin doses to achieve the same glycaemia. Clinical trials should consider sex differences in protocol design and outcome analyses.

Insulin sensitivity in critically ill patients: are women more insulin resistant?

BACKGROUND

Glycaemic control (GC) in intensive care unit is challenging due to significant inter- and intra-patient variability, leading to increased risk of hypoglycaemia. Recent work showed higher insulin resistance in female preterm neonates. This study aims to determine if there are differences in inter- and intra-patient metabolic variability between sexes in adults, to gain in insight into any differences in metabolic response to injury. Any significant difference would suggest GC and randomised trial design should consider sex differences to personalise care.

METHODS

Insulin sensitivity (SI) levels and variability are identified from retrospective clinical data for men and women. Data are divided using 6-h blocks to capture metabolic evolution over time. In total, 91 male and 54 female patient GC episodes of minimum 24 h are analysed. Hypothesis testing is used to determine whether differences are significant (P < 0.05), and equivalence testing is used to assess whether these differences can be considered equivalent at a clinical level. Data are assessed for the raw cohort and in 100 Monte Carlo simulations analyses where the number of men and women are equal.

RESULTS

Demographic data between females and males were all similar, including GC outcomes (safety from hypoglycaemia and high (> 50%) time in target band). Females had consistently significantly lower SI levels than males, and this difference was not clinically equivalent. However, metabolic variability between sexes was never significantly different and always clinically equivalent. Thus, inter-patient variability was significantly different between males and females, but intra-patient variability was equivalent.

CONCLUSION

Given equivalent intra-patient variability and significantly greater insulin resistance, females can receive the same benefit from safe, effective GC as males, but may require higher insulin doses to achieve the same glycaemia. Clinical trials should consider sex differences in protocol design and outcome analyses.

Increased insulin resistance in intensive care: longitudinal retrospective analysis of glycaemic control patients in a New Zealand ICU

BACKGROUND

Critical care populations experience demographic shifts in response to trends in population and healthcare, with increasing severity and/or complexity of illness a common observation worldwide. Inflammation in critical illness impacts glucose-insulin metabolism, and hyperglycaemia is associated with mortality and morbidity. This study examines longitudinal trends in insulin sensitivity across almost a decade of glycaemic control in a single unit.

METHODS

A clinically validated model of glucose-insulin dynamics is used to assess hour-hour insulin sensitivity over the first 72 h of insulin therapy. Insulin sensitivity and its hour-hour percent variability are examined over 8 calendar years alongside severity scores and diagnostics.

RESULTS

Insulin sensitivity was found to decrease by 50-55% from 2011 to 2015, and remain low from 2015 to 2018, with no concomitant trends in age, severity scores or risk of death, or diagnostic category. Insulin sensitivity variability was found to remain largely unchanged year to year and was clinically equivalent (95% confidence interval) at the median and interquartile range. Insulin resistance was associated with greater incidence of high insulin doses in the effect saturation range (6-8 U/h), with the 75th percentile of hourly insulin doses rising from 4-4.5 U/h in 2011-2014 to 6 U/h in 2015-2018.

CONCLUSIONS

Increasing insulin resistance was observed alongside no change in insulin sensitivity variability, implying greater insulin needs but equivalent (variability) challenge to glycaemic control. Increasing insulin resistance may imply greater inflammation and severity of illness not captured by existing severity scores. Insulin resistance reduces glucose tolerance, and can cause greater incidence of insulin saturation and resultant hyperglycaemia. Overall, these results have significant clinical implications for glycaemic control and nutrition management.

Assessment of Glycemic Control Protocol (STAR) Through Compliance Analysis Amongst Malaysian ICU Patients

Purpose

This paper presents an assessment of an automated and personalized stochastic targeted (STAR) glycemic control protocol compliance in Malaysian intensive care unit (ICU) patients to ensure an optimized usage.

Patients and Methods

STAR proposes 1–3 hours treatment based on individual insulin sensitivity variation and history of blood glucose, insulin, and nutrition. A total of 136 patients recorded data from STAR pilot trial in Malaysia (2017–quarter of 2019*) were used in the study to identify the gap between chosen administered insulin and nutrition intervention as recommended by STAR, and the real intervention performed.

Results

The results show the percentage of insulin compliance increased from 2017 to first quarter of 2019* and fluctuated in feed administrations. Overall compliance amounted to 98.8% and 97.7% for administered insulin and feed, respectively. There was higher average of 17 blood glucose measurements per day than in other centres that have been using STAR, but longer intervals were selected when recommended. Control safety and performance were similar for all periods showing no obvious correlation to compliance.

Conclusion

The results indicate that STAR, an automated model-based protocol is positively accepted among the Malaysian ICU clinicians to automate glycemic control and the usage can be extended to other hospitals already. Performance could be improved with several propositions.

Accuracy of Blood Glucose Measurement and Blood Glucose Targets

BACKGROUND

To summarize new evidence regarding the methodological aspects of blood glucose control in the intensive care unit (ICU).

METHODS

We reviewed the literature on blood glucose control in the ICU up to August 2019 through Ovid Medline and Pubmed.

RESULTS

Since the publication of the Leuven studies, the benefits of glycemic control have been recognized. However, the methodology of blood glucose control, notably the blood glucose measurement accuracy and the insulin titration protocol, plays an important but underestimated role. This may partially explain the negative results of the large, pragmatic multicenter trials and made everyone realize that tight glycemic control with less-frequent glucose measurements on less accurate blood glucose meters is neither feasible nor advisable in daily practice. Blood gas analyzers remain the gold standard. New generation point-of-care blood glucose meters may be an alternative when using whole blood of critically ill patients in combination with a clinically validated insulin dosing algorithm.

CONCLUSION

When implementing blood glucose management in an ICU one needs to take into account the interaction between aimed glycemic target and blood glucose measurement methodology.

Kernel density estimates for sepsis classification

OBJECTIVE

Severe sepsis is a leading cause of intensive care unit (ICU) admission, length of stay, mortality, and cost. systemic inflammatory response syndrome (SIRS) and organ failure due to infection define it, but also make it hard to diagnose. Early diagnosis reduces morbidity, mortality and cost, and diagnosis is often significantly delayed due to a lack of effective biomarkers. This research employs kernel density estimation (KDE) methods fusing a personalized, model-based insulin sensitivity (SI) metric with standard bedside measures of: temperature, heart rate, respiratory rate, systolic and diastolic blood pressure, and SIRS, as these measures are available hourly or more frequently.

METHODS

Model-based SI is a derived metric, identified using clinical data and a clinically validated metabolic model. The KDE classifier discriminates severe sepsis and septic shock from moderate sepsis using accepted consensus sepsis scores. A best case in-sample estimate, a worst case independent cross validation estimate, and an accepted .632 bootstrap estimate are calculated to assess performance using multi-level likelihood ratios, and sensitivity and specificity. Performance is assessed against clinically and statistically defined thresholds denoted for the minimum acceptable level as: "high accuracy, often providing useful information, and clinical significance," and similar definitions for greater or lesser quality.

RESULTS

The .632 bootstrap estimate performs near clinically defined levels of high accuracy, often providing useful information, and clinical significance based on sensitivity, specificity, and multilevel likelihood ratios.

CONCLUSION AND SIGNIFICANCE

The classifier created and this overall approach is useful for clinical decision making in diagnosing severe sepsis and septic shock in real time, for both case and control hours. However, improvements could be made with larger clinical data sets.

Risk and reward: extending stochastic glycaemic control intervals to reduce workload

Background

STAR is a model-based, personalised, risk-based dosing approach for glycaemic control (GC) in critically ill patients. STAR provides safe, effective control to nearly all patients, using 1–3 hourly measurement and intervention intervals. However, the average 11–12 measurements per day required can be a clinical burden in many intensive care units. This study aims to significantly reduce workload by extending STAR 1–3 hourly intervals to 1 to 4-, 5-, and 6-hourly intervals, and evaluate the impact of these longer intervals on GC safety and efficacy, using validated in silico virtual patients and trials methods. A Standard STAR approach was used which allowed more hyperglycaemia over extended intervals, and a STAR Upper Limit Controlled approach limited nutrition to mitigate hyperglycaemia over longer intervention intervals.

Results

Extending STAR from 1–3 hourly to 1–6 hourly provided high safety and efficacy for nearly all patients in both approaches. For STAR Standard, virtual trial results showed lower % blood glucose (BG) in the safe 4.4–8.0 mmol/L target band (from 83 to 80%) as treatment intervals increased. Longer intervals resulted in increased risks of hyper- (15% to 18% BG > 8.0 mmol/L) and hypo- (2.1% to 2.8% of patients with min. BG < 2.2 mmol/L) glycaemia. These results were achieved with slightly reduced insulin (3.2 [2.0 5.0] to 2.5 [1.5 3.0] U/h) and nutrition (100 [85 100] to 90 [75 100] % goal feed) rates, but most importantly, with significantly reduced workload (12 to 8 measurements per day). The STAR Upper Limit Controlled approach mitigated hyperglycaemia and had lower insulin and significantly lower nutrition administration rates.

Conclusions

The modest increased risk of hyper- and hypo-glycaemia, and the reduction in nutrition delivery associated with longer treatment intervals represent a significant risk and reward trade-off in GC. However, STAR still provided highly safe, effective control for nearly all patients regardless of treatment intervals and approach, showing this unique risk-based dosing approach, modulating both insulin and nutrition, to be robust in its design. Clinical pilot trials using STAR with different measurement timeframes should be undertaken to confirm these results clinically.

STAR-Liège Clinical Trial Interim Results: Safe and Effective Glycemic Control for All

While the benefits of glycemic control for critically ill patients are increasingly demonstrated, the ability to deliver safe, effective control to intermediate target ranges is widely debated due to the increased risk of hypoglycemia. This study analyzes interim clinical trial results of the fully computerized model-based Stochastic TARgeted (STAR) glycemic control framework at the University Hospital of Liège, Belgium. Patients with dysglycemia were randomly assigned to the full version of STAR, modulating both insulin and nutrition inputs, or STAR-IO, an insulin only version of STAR. Both arms target the normoglycemic 80-145 mg/dL (4.4-8.0 mmol/L) band. Results are further compared to retrospective data from 20 patients under the standard unit protocol targeting a higher 100-150 mg/dL (5.6-8.3 mmol/L) band. Much higher time in target band is provided under the full version of STAR, with similar safety and significantly lower incidence of mild hyperglycemia (blood glucose > 145 mg/dL or 8.0 mmol/L) and severe hyperglycemia (blood glucose > 180 mg/dL or 10.0 mmol/L). As a result, lower median blood glucose levels are safely and consistently achieved with lower glycemic variability, suggesting STAR's potential to improve clinical outcomes. These interim results show the possibility to achieve safe, effective control for all patients using STAR, and suggest glycemic control to lower targets could be beneficial.

Multi-input stochastic prediction of insulin sensitivity for tight glycaemic control using insulin sensitivity and blood glucose data

BACKGROUND

Glycaemic control in the intensive care unit is dependent on effective prediction of patient insulin sensitivity (SI). The stochastic targeted (STAR) controller uses a 2D stochastic model for prediction, with current SI as an input and future SI as an output.

METHODS

This paper develops an extension of the STAR 2D stochastic model into 3D by adding blood glucose (G) as an input. The performance of the 2D and 3D stochastic models is compared over a retrospective cohort of 65,269 data points across 1525 patients.

RESULTS

Under five-fold cross-validation, the 3D model was found to better match the expected potion of data points within, above and below various credible intervals, suggesting it provided a better representation of the underlying probability field. The 3D model was also found to provide an 18.1% narrower 90% credible interval on average, and a narrower 90% credible interval in 96.4% of cases, suggesting it provided more accurate predictions of future SI. Additionally, the 3D stochastic model was found to avoid the undesirable tendency of the 2D model to overestimate SI for patients with high G, and underestimate SI for patients with low G.

CONCLUSIONS

Overall, the 3D stochastic model is shown to provide clear potential benefits over the 2D model for minimal clinical cost or effort, though further exploration into whether these improvements in SI prediction translate into improved clinical outcomes is required.

3D kernel-density stochastic model for more personalized glycaemic control: development and in-silico validation

Background

The challenges of glycaemic control in critically ill patients have been debated for 20 years. While glycaemic control shows benefits inter- and intra-patient metabolic variability results in increased hypoglycaemia and glycaemic variability, both increasing morbidity and mortality. Hence, current recommendations for glycaemic control target higher glycaemic ranges, guided by the fear of harm. Lately, studies have proven the ability to provide safe, effective control for lower, normoglycaemic, ranges, using model-based computerised methods. Such methods usually identify patient-specific physiological parameters to personalize titration of insulin and/or nutrition. The Stochastic-Targeted (STAR) glycaemic control framework uses patient-specific insulin sensitivity and a stochastic model of its future variability to directly account for both inter- and intra-patient variability in a risk-based insulin-dosing approach.

Results

In this study, a more personalized and specific 3D version of the stochastic model used in STAR is compared to the current 2D stochastic model, both built using kernel-density estimation methods. Fivefold cross validation on 681 retrospective patient glycaemic control episodes, totalling over 65,000 h of control, is used to determine whether the 3D model better captures metabolic variability, and the potential gain in glycaemic outcome is assessed using validated virtual trials. Results show that the 3D stochastic model has similar forward predictive power, but provides significantly tighter, more patient-specific, prediction ranges, showing the 2D model over-conservative > 70% of the time. Virtual trial results show that overall glycaemic safety and performance are similar, but the 3D stochastic model reduced median blood glucose levels (6.3 [5.7, 7.0] vs. 6.2 [5.6, 6.9]) with a higher 61% vs. 56% of blood glucose within the 4.4–6.5 mmol/L range.

Conclusions

This improved performance is achieved with higher insulin rates and higher carbohydrate intake, but no loss in safety from hypoglycaemia. Thus, the 3D stochastic model developed better characterises patient-specific future insulin sensitivity dynamics, resulting in improved simulated glycaemic outcomes and a greater level of personalization in control. The results justify inclusion into ongoing clinical use of STAR.

The parameter Houlihan: A solution to high-throughput identifiability indeterminacy for brutally ill-posed problems

One way to interject knowledge into clinically impactful forecasting is to use data assimilation, a nonlinear regression that projects data onto a mechanistic physiologic model, instead of a set of functions, such as neural networks. Such regressions have an advantage of being useful with particularly sparse, non-stationary clinical data. However, physiological models are often nonlinear and can have many parameters, leading to potential problems with parameter identifiability, or the ability to find a unique set of parameters that minimize forecasting error. The identifiability problems can be minimized or eliminated by reducing the number of parameters estimated, but reducing the number of estimated parameters also reduces the flexibility of the model and hence increases forecasting error. We propose a method, the parameter Houlihan, that combines traditional machine learning techniques with data assimilation, to select the right set of model parameters to minimize forecasting error while reducing identifiability problems. The method worked well: the data assimilation-based glucose forecasts and estimates for our cohort using the Houlihan-selected parameter sets generally also minimize forecasting errors compared to other parameter selection methods such as by-hand parameter selection. Nevertheless, the forecast with the lowest forecast error does not always accurately represent physiology, but further advancements of the algorithm provide a path for improving physiologic fidelity as well. Our hope is that this methodology represents a first step toward combining machine learning with data assimilation and provides a lower-threshold entry point for using data assimilation with clinical data by helping select the right parameters to estimate.

Model-based glycemic control in a Malaysian intensive care unit: performance and safety study

Background: Stress-induced hyperglycemia is common in critically ill patients. A few forms of model-based glycemic control have been introduced to reduce this phenomena and among them is the automated STAR protocol which has been used in the Christchurch and Gyulá hospitals' intensive care units (ICUs) since 2010. Methods: This article presents the pilot trial assessment of STAR protocol which has been implemented in the International Islamic University Malaysia Medical Centre (IIUMMC) Hospital ICU since December 2017. One hundred and forty-two patients who received STAR treatment for more than 20 hours were used in the assessment. The initial results are presented to discuss the ability to adopt and adapt the model-based control framework in a Malaysian environment by analyzing its performance and safety. Results: Overall, 60.7% of blood glucose measurements were in the target band. Only 0.78% and 0.02% of cohort measurements were below 4.0 mmol/L and 2.2 mmol/L (the limitsfor mild and severe hypoglycemia, respectively). Treatment preference-wise, the clinical staff were favorable of longer intervention options when available. However, 1 hourly treatments were still used in 73.7% of cases. Conclusion: The protocol succeeded in achieving patient-specific glycemic control while maintaining safety and was trusted by nurses to reduce workload. Its lower performance results, however, give the indication for modification in some of the control settings to better fit the Malaysian environment.

Modelling intestinal glucose absorption in premature infants using continuous glucose monitoring data

BACKGROUND

Model-based glycaemic control protocols have shown promise in neonatal intensive care units (NICUs) for reducing both hyperglycaemia and insulin-therapy driven hypoglycaemia. However, current models for the appearance of glucose from enteral feeding are based on values from adult intensive care cohorts. This study aims to determine enteral glucose appearance model parameters more reflective of premature infant physiology.

METHODS

Peaks in CGM data associated with enteral milk feeds in preterm and term infants are used to fit a two compartment gut model. The first compartment describes glucose in the stomach, and the half life of gastric emptying is estimated as 20 min from literature. The second compartment describes glucose in the small intestine, and absorption of glucose into the blood is fit to CGM data. Two infant cohorts from two NICUs are used, and results are compared to appearances derived from data in highly controlled studies in literature.

RESULTS

The average half life across all infants for glucose absorption from the gut to the blood was 50 min. This result was slightly slower than, but of similar magnitude to, results derived from literature. No trends were found with gestational or postnatal age. Breast milk fed infants were found to have a higher absorption constant than formula fed infants, a result which may reflect known differences in gastric emptying for different feed types.

CONCLUSIONS

This paper presents a methodology for estimation of glucose appearance due to enteral feeding, and model parameters suitable for a NICU model-based glycaemic control context.

Nutrition delivery, workload and performance in a model-based ICU glycaemic control system

BACKGROUND AND OBJECTIVE

Hyperglycaemia is commonplace in the adult intensive care unit (ICU), and has been associated with increased morbidity and mortality. Effective glycaemic control (GC) can reduce morbidity and mortality, but has proven difficult. STAR is a model-based GC protocol that uniquely maintains normoglycaemia by changing both insulin and nutrition interventions, and has been proven effective in controlling blood glucose (BG) in the ICU. However, most ICU GC protocols only change insulin interventions, making the variable nutrition aspect of STAR less clinically desirable. This paper compares the performance of STAR modulating only insulin, with three simpler alternative nutrition protocols in clinically evaluated virtual trials.

METHODS

Alternative nutrition protocols are fixed nutrition rate (100% caloric goal), CB (Cahill et al. best) stepped nutrition rate (60%, 80% and 100% caloric goal for the first 3 days of GC, and 100% thereafter) and SLQ (STAR lower quartile) stepped nutrition rate (65%, 75% and 85% caloric goal for the first 3 days of GC, and 85% thereafter). Each nutrition protocol is simulated with the STAR insulin protocol on a 221 patient virtual cohort, and GC performance, safety and total intervention workload are assessed.

RESULTS

All alternative nutrition protocols considerably reduced total intervention workload (14.6-19.8%) due to reduced numbers of nutrition changes. However, only the stepped nutrition protocols achieved similar GC performance to the current variable nutrition protocol. Of the two stepped nutrition protocols, the SLQ nutrition protocol also improved GC safety, almost halving the number of severe hypoglycaemic cases (5 vs. 9, P = 0.42).

CONCLUSIONS

Overall, the SLQ nutrition protocol was the best alternative to the current variable nutrition protocol, but either stepped nutrition protocol could be adapted by STAR to reduce workload and make it more clinically acceptable, while maintaining its proven performance and safety.

Mechanistic machine learning: how data assimilation leverages physiologic knowledge using Bayesian inference to forecast the future, infer the present, and phenotype

We introduce data assimilation as a computational method that uses machine learning to combine data with human knowledge in the form of mechanistic models in order to forecast future states, to impute missing data from the past by smoothing, and to infer measurable and unmeasurable quantities that represent clinically and scientifically important phenotypes. We demonstrate the advantages it affords in the context of type 2 diabetes by showing how data assimilation can be used to forecast future glucose values, to impute previously missing glucose values, and to infer type 2 diabetes phenotypes. At the heart of data assimilation is the mechanistic model, here an endocrine model. Such models can vary in complexity, contain testable hypotheses about important mechanics that govern the system (eg, nutrition's effect on glucose), and, as such, constrain the model space, allowing for accurate estimation using very little data.

Interpretation of Retrospective BG Measurements

BACKGROUND

This study investigates blood glucose (BG) measurement interpolation techniques to represent intermediate BG dynamics, and the effect resampling of retrospective BG data has on key glycemic control (GC) performance results. GC protocols in the ICU have varying BG measurement intervals ranging from 0.5 to 4 hours. Sparse data pose problems, particularly in comparing GC performance or model fitting, and thus interpolation is required.

METHODS

Retrospective data from SPRINT in Christchurch Hospital Intensive Care Unit (ICU) (2005-2007) were used to analyze several interpolation techniques. Piecewise linear, spline, and cubic interpolation functions, which force interpolation through measured data, as well as 1st and 2nd Order B-spline basis functions, are used to identify the interpolated trace. Dense data were thinned to increase sparsity and obtain measurements (Hidden Measurements) for comparison after interpolation. Performance is assessed based on error in capturing hidden measurements. Finally, the effect of minutely versus hourly sampling of the interpolated trace on key GC performance statistics was investigated using retrospective data received from STAR GC in Christchurch Hospital ICU, New Zealand (2011-2015).

RESULTS

All of the piecewise functions performed considerably better than the fitted interpolation functions. Linear piecewise interpolation performed the best having a mean RMSE 0.39 mmol/L, within 2 standard deviations of the BG sensor error. Minutely sampled BG resulted in significantly different key GC performance values when compared to raw sparse BG measurements.

CONCLUSION

Linear piecewise interpolation provides the best estimate of intermediate BG dynamics and all analyses comparing GC protocol performance should use minutely linearly interpolated BG data.

Identifiability Analysis of Three Control-Oriented Models for Use in Artificial Pancreas Systems

OBJECTIVE

Our aim is to analyze the identifiability of three commonly used control-oriented models for glucose control in patients with type 1 diabetes (T1D).

METHODS

Structural and practical identifiability analysis were performed on three published control-oriented models for glucose control in patients with type 1 diabetes (T1D): the subcutaneous oral glucose minimal model (SOGMM), the intensive control insulin-nutrition-glucose (ICING) model, and the minimal model control-oriented (MMC). Structural identifiability was addressed with a combination of the generating series (GS) approach and identifiability tableaus whereas practical identifiability was studied by means of (1) global ranking of parameters via sensitivity analysis together with the Latin hypercube sampling method (LHS) and (2) collinearity analysis among parameters. For practical identifiability and model identification, continuous glucose monitor (CGM), insulin pump, and meal records were selected from a set of patients (n = 5) on continuous subcutaneous insulin infusion (CSII) that underwent a clinical trial in an outpatient setting. The performance of the identified models was analyzed by means of the root mean square (RMS) criterion.

RESULTS

A reliable set of identifiable parameters was found for every studied model after analyzing the possible identifiability issues of the original parameter sets. According to an importance factor ([Formula: see text]), it was shown that insulin sensitivity is not the most influential parameter from the dynamical point of view, that is, is not the parameter impacting the outputs the most of the three models, contrary to what is assumed in the literature. For the test data, the models demonstrated similar performance with most RMS values around 20 mg/dl (min: 15.64 mg/dl, max: 51.32 mg/dl). However, MMC failed to identify the model for patient 4. Also, considering the three models, the MMC model showed the higher parameter variability when reidentified every 6 hours.

CONCLUSION

This study shows that both structural and practical identifiability analysis need to be considered prior to the model identification/individualization in patients with T1D. It was shown that all the studied models are able to represent the CGM data, yet their usefulness in a hypothetical artificial pancreas could be a matter of debate. In spite that the three models do not capture all the dynamics and metabolic effects as a maximal model (ie, our FDA-accepted UVa/Padova simulator), SOGMM and ICING appear to be more appealing than MMC regarding both the performance and parameter variability after reidentification. Although the model predictions of ICING are comparable to the ones of the SOGMM model, the large parameter set makes the model prone to overfitting if all parameters are identified. Moreover, the existence of a high nonlinear function like [Formula: see text] prevents the use of tools from the linear systems theory.

Improving glycemic control in critically ill patients: personalized care to mimic the endocrine pancreas

There is considerable physiological and clinical evidence of harm and increased risk of death associated with dysglycemia in critical care. However, glycemic control (GC) currently leads to increased hypoglycemia, independently associated with a greater risk of death. Indeed, recent evidence suggests GC is difficult to safely and effectively achieve for all patients. In this review, leading experts in the field discuss this evidence and relevant data in diabetology, including the artificial pancreas, and suggest how safe, effective GC can be achieved in critically ill patients in ways seeking to mimic normal islet cell function. The review is structured around the specific clinical hurdles of: understanding the patient’s metabolic state; designing GC to fit clinical practice, safety, efficacy, and workload; and the need for standardized metrics. These aspects are addressed by reviewing relevant recent advances in science and technology. Finally, we provide a set of concise recommendations to advance the safety, quality, consistency, and clinical uptake of GC in critical care. This review thus presents a roadmap toward better, more personalized metabolic care and improved patient outcomes.

Performance of Stochastic Targeted Blood Glucose Control Protocol by virtual trials in the Malaysian intensive care unit

BACKGROUND AND OBJECTIVE

Blood glucose variability is common in healthcare and it is not related or influenced by diabetes mellitus. To minimise the risk of high blood glucose in critically ill patients, Stochastic Targeted Blood Glucose Control Protocol is used in intensive care unit at hospitals worldwide. Thus, this study focuses on the performance of stochastic modelling protocol in comparison to the current blood glucose management protocols in the Malaysian intensive care unit. Also, this study is to assess the effectiveness of Stochastic Targeted Blood Glucose Control Protocol when it is applied to a cohort of diabetic patients.

METHODS

Retrospective data from 210 patients were obtained from a general hospital in Malaysia from May 2014 until June 2015, where 123 patients were having comorbid diabetes mellitus. The comparison of blood glucose control protocol performance between both protocol simulations was conducted through blood glucose fitted with physiological modelling on top of virtual trial simulations, mean calculation of simulation error and several graphical comparisons using stochastic modelling.

RESULTS

Stochastic Targeted Blood Glucose Control Protocol reduces hyperglycaemia by 16% in diabetic and 9% in nondiabetic cohorts. The protocol helps to control blood glucose level in the targeted range of 4.0-10.0 mmol/L for 71.8% in diabetic and 82.7% in nondiabetic cohorts, besides minimising the treatment hour up to 71 h for 123 diabetic patients and 39 h for 87 nondiabetic patients.

CONCLUSION

It is concluded that Stochastic Targeted Blood Glucose Control Protocol is good in reducing hyperglycaemia as compared to the current blood glucose management protocol in the Malaysian intensive care unit. Hence, the current Malaysian intensive care unit protocols need to be modified to enhance their performance, especially in the integration of insulin and nutrition intervention in decreasing the hyperglycaemia incidences. Improvement in Stochastic Targeted Blood Glucose Control Protocol in terms of u_en model is also a must to adapt with the diabetic cohort.

Accuracy and optimization of a subcutaneous insulin model for less acute critical care patients

Extending safe, effective glycemic control to the general wards requires a simple approach using subcutaneous (SC) insulin. However, this approach can increase relative risk compared to intravenous insulin due to the increased variability of SC insulin appearance. This paper evaluates the accuracy of a SC plasma insulin model and optimizes its parameters using measured plasma insulin data from 6 less acute critical care patients treated with SC insulin. The SC plasma insulin model used captures the dynamics of regular SC insulin well. However, there appears to be a positive bias leading to an overall median [IQR] residual error of -28.3 [-37 - 19] mU/L. The optimized model reduced the RMS residual error by 20-70% for each patient. The distinct inter- and intra-patient, and cohort variation seen in this data highlights the importance to of understanding how SC insulin appearance dynamics may be affected by the subject condition.

Next-generation, personalised, model-based critical care medicine: a state-of-the art review of in silico virtual patient models, methods, and cohorts, and how to validation them

Critical care, like many healthcare areas, is under a dual assault from significantly increasing demographic and economic pressures. Intensive care unit (ICU) patients are highly variable in response to treatment, and increasingly aging populations mean ICUs are under increasing demand and their cohorts are increasingly ill. Equally, patient expectations are growing, while the economic ability to deliver care to all is declining. Better, more productive care is thus the big challenge. One means to that end is personalised care designed to manage the significant inter- and intra-patient variability that makes the ICU patient difficult. Thus, moving from current "one size fits all" protocolised care to adaptive, model-based "one method fits all" personalised care could deliver the required step change in the quality, and simultaneously the productivity and cost, of care. Computer models of human physiology are a unique tool to personalise care, as they can couple clinical data with mathematical methods to create subject-specific models and virtual patients to design new, personalised and more optimal protocols, as well as to guide care in real-time. They rely on identifying time varying patient-specific parameters in the model that capture inter- and intra-patient variability, the difference between patients and the evolution of patient condition. Properly validated, virtual patients represent the real patients, and can be used in silico to test different protocols or interventions, or in real-time to guide care. Hence, the underlying models and methods create the foundation for next generation care, as well as a tool for safely and rapidly developing personalised treatment protocols over large virtual cohorts using virtual trials. This review examines the models and methods used to create virtual patients. Specifically, it presents the models types and structures used and the data required. It then covers how to validate the resulting virtual patients and trials, and how these virtual trials can help design and optimise clinical trial. Links between these models and higher order, more complex physiome models are also discussed. In each section, it explores the progress reported up to date, especially on core ICU therapies in glycemic, circulatory and mechanical ventilation management, where high cost and frequency of occurrence provide a significant opportunity for model-based methods to have measurable clinical and economic impact. The outcomes are readily generalised to other areas of medical care.

Nutrition delivery of a model-based ICU glycaemic control system

BACKGROUND

Hyperglycaemia is commonplace in the adult intensive care unit (ICU), associated with increased morbidity and mortality. Effective glycaemic control (GC) can reduce morbidity and mortality, but has proven difficult. STAR is a proven, effective model-based ICU GC protocol that uniquely maintains normo-glycaemia by changing both insulin and nutrition interventions to maximise nutrition in the context of GC in the 4.4-8.0 mmol/L range. Hence, the level of nutrition it provides is a time-varying estimate of the patient-specific ability to take up glucose.

METHODS

First, the clinical provision of nutrition by STAR in Christchurch Hospital, New Zealand (N = 221 Patients) is evaluated versus other ICUs, based on the Cahill et al. survey of 158 ICUs. Second, the inter- and intra- patient variation of nutrition delivery with STAR is analysed. Nutrition rates are in terms of percentage of caloric goal achieved.

RESULTS

Mean nutrition rates clinically achieved by STAR were significantly higher than the mean and best ICU surveyed, for the first 3 days of ICU stay. There was large inter-patient variation in nutrition rates achieved per day, which reduced overtime as patient-specific metabolic state stabilised. Median intra-patient variation was 12.9%; however, the interquartile range of the mean per-patient nutrition rates achieved was 74.3-98.2%, suggesting patients do not deviate much from their mean patient-specific nutrition rate. Thus, the ability to tolerate glucose intake varies significantly between, rather than within, patients.

CONCLUSIONS

Overall, STAR's protocol-driven changes in nutrition rate provide higher nutrition rates to hyperglycaemic patients than those of 158 ICUs from 20 countries. There is significant inter-patient variability between patients to tolerate and uptake glucose, where intra-patient variability over stay is much lower. Thus, a best nutrition rate is likely patient specific for patients requiring GC. More importantly, these overall outcomes show high nutrition delivery and safe, effective GC are not exclusive and that restricting nutrition for GC does not limit overall nutritional intake compared to other ICUs.

Levels and Diagnostic Value of Model-based Insulin Sensitivity in Sepsis: A Preliminary Study

Background and Aims:

Currently, there is a lack of real-time metric with high sensitivity and specificity to diagnose sepsis. Insulin sensitivity (SI) may be determined in real-time using mathematical glucose-insulin models; however, its effectiveness as a diagnostic test of sepsis is unknown. Our aims were to determine the levels and diagnostic value of model-based SI for identification of sepsis in critically ill patients.

Materials and Methods:

In this retrospective, cohort study, we analyzed SI levels in septic (n = 18) and nonseptic (n = 20) patients at 1 (baseline), 4, 8, 12, 16, 20, and 24 h of their Intensive Care Unit admission. Patients with diabetes mellitus Type I or Type II were excluded from the study. The SI levels were derived by fitting the blood glucose levels, insulin infusion and glucose input rates into the Intensive Control of Insulin-Nutrition-Glucose model.

Results:

The median SI levels were significantly lower in the sepsis than in the nonsepsis at all follow-up time points. The areas under the receiver operating characteristic curve of the model-based SI at baseline for discriminating sepsis from nonsepsis was 0.814 (95% confidence interval, 0.675–0.953). The optimal cutoff point of the SI test was 1.573 × 10⁻⁴ L/mu/min. At this cutoff point, the sensitivity was 77.8%, specificity was 75%, positive predictive value was 73.7%, and negative predictive value was 78.9%.

Conclusions:

Model-based SI ruled in and ruled out sepsis with fairly high sensitivity and specificity in our critically ill nondiabetic patients. These findings can be used as a foundation for further, prospective investigation in this area.

Continuous Glucose Monitoring Measures Can Be Used for Glycemic Control in the ICU: An In-Silico Study

BACKGROUND

Continuous glucose monitoring (CGM) technology has become more prevalent in the intensive care unit (ICU), offering potential benefits of increased safety and reduced workload in glycemic control (GC). The drift and higher point accuracy errors of CGM devices over traditional intermittent blood glucose (BG) measures have so far limited their application in the ICU. This study delineates the trade-offs of performance, safety and workload that CGM sensors provide in GC protocols.

METHODS

Clinical data from 236 patients were used for clinically validated virtual trials. A CGM-enabled version of the STAR GC protocol was used to evaluate the use of guard rails and rolling windows. Safety was assessed through percentage of patients who had a severe hypoglycemic episode (BG < 40 mg/dl) as well as percentage of resampled BG < 72 mg/dl. Performance was assessed as percentage of resampled measurements in the 80-126 mg/dl and the 80-144 mg/dl target bands. Workload was measured by number of manual BG measures per day.

RESULTS

CGM-enabled versions of STAR decreased the number of required blood draws by up to 74%, while maintaining performance (76.6% BG measurements in the 80-126 mg/dl range vs 62.8% clinically, 87.9% in the 80-144 mg/dl range vs 83.7% clinically) and maintaining patient safety (1.13% of patients experienced a severe hypoglycemic event vs 0.85% clinically, 1.37% of BG measurements were less than 72 mg/dl vs 0.51% clinically).

CONCLUSION

CGM sensor traces were reproduced in virtual trials to guide GC. Existing GC protocols such as STAR may need to be adjusted only slightly to gain the benefits of the increased temporal measurements of CGM sensors, through which workload may be significantly decreased while maintaining GC performance and safety.

Untangling glycaemia and mortality in critical care

Background

Hyperglycaemia is associated with adverse outcomes in the intensive care unit, and initial studies suggested outcome benefits of glycaemic control (GC). However, subsequent studies often failed to replicate these results, and they were often unable to achieve consistent, safe control, raising questions about the benefit or harm of GC as well as the nature of the association of glycaemia with mortality and clinical outcomes. In this study, we evaluated if non-survivors are harder to control than survivors and determined if glycaemic outcome is a function of patient condition and eventual outcome or of the glycaemic control provided.

Methods

Clinically validated, model-based, hour-to-hour insulin sensitivity (SI) and its hour-to-hour variability (%ΔSI) were identified over the first 72 h of therapy in 145 patients (119 survivors, 26 non-survivors). In hypothesis testing, we compared distributions of SI and %ΔSI in 6-hourly blocks for survivors and non-survivors. In equivalence testing, we assessed if differences in these distributions, based on blood glucose measurement error, were clinically significant.

Results

SI level was never equivalent between survivors and non-survivors (95% CI of percentage difference in medians outside ±12%). Non-survivors had higher SI, ranging from 9% to 47% higher overall in 6-h blocks, and this difference became statistically significant as glycaemic control progressed. %ΔSI was equivalent between survivors and non-survivors for all 6-hourly blocks (95% CI of difference in medians within ±12%) and decreased in general over time as glycaemic control progressed.

Conclusions

Whereas non-survivors had higher SI levels, variability was equivalent to that of survivors over the first 72 h. These results indicate survivors and non-survivors are equally controllable, given an effective glycaemic control protocol, suggesting that glycaemia level and variability, and thus the association between glycaemia and outcome, are essentially determined by the control provided rather than by underlying patient or metabolic condition.

Electronic supplementary material

The online version of this article (doi:10.1186/s13054-017-1725-y) contains supplementary material, which is available to authorized users.