Enhanced Model Predictive Control (eMPC) Strategy for Automated Glucose Control.

Development of an effective artificial pancreas (AP) controller to deliver insulin autonomously to people with type 1 diabetes mellitus is a difficult task. In this paper, three enhancements to a clinically validated AP model predictive controller (MPC) are proposed that address major challenges facing automated blood glucose control, and are then evaluated by both in silico tests and clinical trials. First, the core model of insulin-blood glucose dynamics utilized in the MPC is expanded with a medically inspired personalization scheme to improve controller responses in the face of inter- and intra-individual variations in insulin sensitivity. Next, the asymmetric nature of the short-term consequences of hypoglycemia versus hyperglycemia is incorporated in an asymmetric weighting of the MPC cost function. Finally, an enhanced dynamic insulin-on-board algorithm is proposed to minimize the likelihood of controller-induced hypoglycemia following a rapid rise of blood glucose due to rescue carbohydrate load with accompanying insulin suspension. Each advancement is evaluated separately and in unison through in silico trials based on a new clinical protocol, which incorporates induced hyper- and hypoglycemia to test robustness. The advancements are also evaluated in an advisory mode (simulated) testing of clinical data. The combination of the three proposed advancements show statistically significantly improved performance over the nonpersonalized controller without any enhancements across all metrics, displaying increased time in the 70-180 mg/dL safe glycemic range (76.9 versus 68.8%) and the 80-140 mg/dL euglycemic range (48.1 versus 44.5%), without a statistically significant increase in instances of hypoglycemia. The proposed advancements provide safe control action for AP applications, personalizing and improving controller performance without the need for extensive model identification processes.

Randomized Crossover Comparison of Personalized MPC and PID Control Algorithms for the Artificial Pancreas.

OBJECTIVE: To evaluate two widely used control algorithms for an artificial pancreas (AP) under nonideal but comparable clinical conditions. RESEARCH DESIGN AND METHODS: After a pilot safety and feasibility study (n = 10), closed-loop control (CLC) was evaluated in a randomized, crossover trial of 20 additional adults with type 1 diabetes. Personalized model predictive control (MPC) and proportional integral derivative (PID) algorithms were compared in supervised 27.5-h CLC sessions. Challenges included overnight control after a 65-g dinner, response to a 50-g breakfast, and response to an unannounced 65-g lunch. Boluses of announced dinner and breakfast meals were given at mealtime. The primary outcome was time in glucose range 70-180 mg/dL. RESULTS: Mean time in range 70-180 mg/dL was greater for MPC than for PID (74.4 vs. 63.7%, P = 0.020). Mean glucose was also lower for MPC than PID during the entire trial duration (138 vs. 160 mg/dL, P = 0.012) and 5 h after the unannounced 65-g meal (181 vs. 220 mg/dL, P = 0.019). There was no significant difference in time with glucose <70 mg/dL throughout the trial period. CONCLUSIONS: This first comprehensive study to compare MPC and PID control for the AP indicates that MPC performed particularly well, achieving nearly 75% time in the target range, including the unannounced meal. Although both forms of CLC provided safe and effective glucose management, MPC performed as well or better than PID in all metrics.

Early Detection of Physical Activity for People With Type 1 Diabetes Mellitus.

BACKGROUND: Early detection of exercise in individuals with type 1 diabetes mellitus (T1DM) may allow changes in therapy to prevent hypoglycemia. Currently there is limited experience with automated methods that detect the onset and end of exercise in this population. We sought to develop a novel method to quickly and reliably detect the onset and end of exercise in these individuals before significant changes in blood glucose (BG) occur. METHODS: Sixteen adults with T1DM were studied as outpatients using a diary, accelerometer, heart rate monitor, and continuous glucose monitor for 2 days. These data were used to develop a principal component analysis based exercise detection method. Subjects also performed 60 and 30 minute exercise sessions at 30% and 50% predicted heart rate reserve (HRR), respectively. The detection method was applied to the exercise sessions to determine how quickly the detection of start and end of exercise occurred relative to change in BG. RESULTS: Mild 30% HRR and moderate 50% HRR exercise onset was identified in 6 ± 3 and 5 ± 2 (mean ± SD) minutes, while completion was detected in 3 ± 8 and 6 ± 5 minutes, respectively. BG change from start of exercise to detection time was 1 ± 6 and -1 ± 3 mg/dL, and, from the end of exercise to detection time was 6 ± 4 and -17 ± 13 mg/dL, respectively, for the 2 exercise sessions. False positive and negative ratios were 4 ± 2% and 21 ± 22%. CONCLUSIONS: The novel method for exercise detection identified the onset and end of exercise in approximately 5 minutes, with an average BG change of only -6 mg/dL.

Empirical Dynamic Model Identification for Blood-Glucose Dynamics in Response to Physical Activity.

In this paper, the dynamic response of blood glucose concentration in response to physical activity of people with Type 1 Diabetes Mellitus (T1DM) is captured by subspace identification methods. Activity (input) and subcutaneous blood glucose measurements (output) are employed to construct a personalized prediction model through semi-definite programming. The model is calibrated and subsequently validated with non-overlapping data sets from 15 T1DM subjects. This preliminary clinical evaluation reveals the underlying linear dynamics between blood glucose concentration and physical activity. These types of models can enhance our capabilities of achieving tighter blood glucose control and early detection of hypoglycemia for people with T1DM.

Design of the Glucose Rate Increase Detector: A Meal Detection Module for the Health Monitoring System.

The Glucose Rate Increase Detector (GRID), a module of the Health Monitoring System (HMS), has been designed to operate in parallel to the glucose controller to detect meal events and safely trigger a meal bolus. The GRID algorithm was tuned on clinical data with 40-70 g CHO meals and tested on simulation data with 50-100 g CHO meals. Active closed- and open-loop protocols were executed in silico with various treatments, including automatic boluses based on a 75 g CHO meal and boluses based on simulated user input of meal size. An optional function was used to reduce the recommended bolus using recent insulin and glucose history. For closed-loop control of a 3-meal scenario (50, 75, and 100 g CHO), the GRID improved median time in the 80-180 mg/dL range by 17% and in the >180 range by 14% over unannounced meals, using an automatic bolus for a 75 g CHO meal at detection. Under open-loop control of a 75 g CHO meal, the GRID shifted the median glucose peak down by 73 mg/dL and earlier by 120 min and reduced the time >180 mg/dL by 57% over a missed-meal bolus scenario, using a full meal bolus at detection. The GRID improved closed-loop control in the presence of large meals, without increasing late postprandial hypoglycemia. Users of basal-bolus therapy could also benefit from GRID as a safety alert for missed meal corrections.

Clinical evaluation of an automated artificial pancreas using zone-model predictive control and health monitoring system.

BACKGROUND: This study was performed to evaluate the safety and efficacy of a fully automated artificial pancreas using zone-model predictive control (zone-MPC) with the health monitoring system (HMS) during unannounced meals and overnight and exercise periods. SUBJECTS AND METHODS: A fully automated closed-loop artificial pancreas was evaluated in 12 subjects (eight women, four men) with type 1 diabetes (mean±SD age, 49.4±10.4 years; diabetes duration, 32.7±16.0 years; glycosylated hemoglobin, 7.3±1.2%). The zone-MPC controller used an a priori model that was initialized using the subject's total daily insulin. The controller was designed to keep glucose levels between 80 and 140 mg/dL. A hypoglycemia prediction algorithm, a module of the HMS, was used in conjunction with the zone controller to alert the user to consume carbohydrates if the glucose level was predicted to fall below 70 mg/dL in the next 15 min. RESULTS: The average time spent in the 70-180 mg/dL range, measured by the YSI glucose and lactate analyzer (Yellow Springs Instruments, Yellow Springs, OH), was 80% for the entire session, 92% overnight from 12 a.m. to 7 a.m., and 69% and 61% for the 5-h period after dinner and breakfast, respectively. The time spent < 60 mg/dL for the entire session by YSI was 0%, with no safety events. The HMS sent appropriate warnings to prevent hypoglycemia via short and multimedia message services, at an average of 3.8 treatments per subject. CONCLUSIONS: The combination of the zone-MPC controller and the HMS hypoglycemia prevention algorithm was able to safely regulate glucose in a tight range with no adverse events despite the challenges of unannounced meals and moderate exercise.

Adaptive blood glucose control for intensive care applications.

Control of blood glucose concentration for patients in intensive care units (ICUs) has been demonstrated to be beneficial in reducing mortality and the incidence of serious complications, for both diabetic and non-diabetic patients. However, the high degree of variability and uncertainty characterizing the physiological conditions of critically ill subjects makes automated glucose control quite difficult; consequently, traditional, nurse-implemented protocols are widely employed. These protocols are based on infrequent glucose measurements, look-up tables to determine the appropriate insulin infusion rates, and bedside insulin administration. In this paper, a novel automatic adaptive control strategy based on frequent glucose measurements and a self-tuning control technique is validated based on a simulation study for 200 virtual patients. The adaptive control strategy is shown to be highly effective in controlling blood glucose concentration despite the large degree of variability in the blood glucose response exhibited by the 200 simulated patients.

Predicting subcutaneous glucose concentration using a latent-variable-based statistical method for type 1 diabetes mellitus.

BACKGROUND: Accurate prediction of future glucose concentration for type 1 diabetes mellitus (T1DM) is needed to improve glycemic control and to facilitate proactive management before glucose concentrations reach undesirable concentrations. The availability of frequent glucose measurements, insulin infusion rates, and meal carbohydrate estimates can be used to good advantage to capture important information concerning glucose dynamics. METHODS: This article evaluates the feasibility of using a latent variable (LV)-based statistical method to model glucose dynamics and to forecast future glucose concentrations for T1DM applications. The prediction models are developed using a proposed LV-based approach and are evaluated for retrospective clinical data from seven individuals with T1DM and for In silico simulations using the Food and Drug Administration-accepted University of Virginia/University of Padova metabolic simulator. This article provides comparisons of the prediction accuracy of the LV-based method with that of a standard modeling alternative. The influence of key design parameters on the performance of the LV-based method is also illustrated. RESULTS: In general, the LV-based method provided improved prediction accuracy in comparison with conventional autoregressive (AR) models and autoregressive with exogenous input (ARX) models. For larger prediction horizons (≥30 min), the LV-based model with exogenous inputs achieved the best prediction performance based on a paired t-test (α = 0.05). CONCLUSIONS: The LV-based method resulted in models whose glucose prediction accuracy was as least as good as the accuracies of standard AR/ARX models and a simple model-free approach. Furthermore, the new approach is less sensitive to changing conditions and the effect of key design parameters.

Clinically relevant hypoglycemia prediction metrics for event mitigation.

BACKGROUND: The purpose of this study was to develop a method to compare hypoglycemia prediction algorithms and choose parameter settings for different applications, such as triggering insulin pump suspension or alerting for rescue carbohydrate treatment. MATERIALS AND METHODS: Hypoglycemia prediction algorithms with different parameter settings were implemented on an ambulatory dataset containing 490 days from 30 subjects with type 1 diabetes mellitus using the Dexcom™ (San Diego, CA) SEVEN™ continuous glucose monitoring system. The performance was evaluated using a proposed set of metrics representing the true-positive ratio, false-positive rate, and distribution of warning times. A prospective, in silico study was performed to show the effect of using different parameter settings to prevent or rescue from hypoglycemia. RESULTS: The retrospective study results suggest the parameter settings for different methods of hypoglycemia mitigation. When rescue carbohydrates are used, a high true-positive ratio, a minimal false-positive rate, and alarms with short warning time are desired. These objectives were met with a 30-min prediction horizon and two successive flags required to alarm: 78% of events were detected with 3.0 false alarms/day and 66% probability of alarms occurring within 30 min of the event. This parameter setting selection was confirmed in silico: treating with rescue carbohydrates reduced the duration of hypoglycemia from 14.9% to 0.5%. However, for a different method, such as pump suspension, this parameter setting only reduced hypoglycemia to 8.7%, as can be expected by the low probability of alarming more than 30 min ahead. CONCLUSIONS: The proposed metrics allow direct comparison of hypoglycemia prediction algorithms and selection of parameter settings for different types of hypoglycemia mitigation, as shown in the prospective in silico study in which hypoglycemia was alerted or treated with rescue carbohydrates.

Design of the health monitoring system for the artificial pancreas: low glucose prediction module.

BACKGROUND: The purpose of this study was to design and evaluate a safety system for the artificial pancreas device system (APDS). Safe operation of the APDS is a critical task, where the safety system is engaged only as needed to ensure reliable operation without positive feedback to the controller. METHODS: The Health Monitoring System (HMS) was designed as a modular system to ensure the safety of the APDS and the user. It was designed using a large set of ambulatory data and evaluated in silico by inducing hypoglycemia with a missed meal [bolus for a 65 g carbohydrate (CHO) meal] and administering rescue CHOs per HMS alerting. The HMS was validated in-clinic with a real-life challenge of a subject who overdosed insulin prior to admission. RESULTS: The HMS was evaluated for clinical use with a 15 min prediction horizon. Retrospectively, 93.5% of episodes were detected with 2.9 false alarms per day. During in silico evaluation, the HMS reduced the time spent <70 mg/dl from 15% to 3%. When the HMS was first tested in-clinic, the subject overdosed ~3 U of insulin prior to her arrival to a closed-loop session (against protocol). The controller reduced insulin delivery, and the HMS gave four alerts that were successfully received via clinical software and text and multimedia messages. Even with insulin reduction and CHO supplements, hypoglycemia was unavoidable but manageable due to the HMS, confirming that a safety system to detect adverse events is an essential part of the APDS. CONCLUSIONS: The ability of the HMS to be an effective alert system that provides a safety layer to the APDS controller has been demonstrated in a clinical setting.

Control-relevant models for glucose control using a priori patient characteristics.

One of the difficulties in the development of a reliable artificial pancreas for people with type 1 diabetes mellitus (T1DM) is the lack of accurate models of an individual's response to insulin. Most control algorithms proposed to control the glucose level in subjects with T1DM are model-based. Avoiding postprandial hypoglycemia ( 60 mg/dl) while minimizing prandial hyperglycemia ( > 180 mg/dl) has shown to be difficult in a closed-loop setting due to the patient-model mismatch. In this paper, control-relevant models are developed for T1DM, as opposed to models that minimize a prediction error. The parameters of these models are chosen conservatively to minimize the likelihood of hypoglycemia events. To limit the conservatism due to large intersubject variability, the models are personalized using a priori patient characteristics. The models are implemented in a zone model predictive control algorithm. The robustness of these controllers is evaluated in silico, where hypoglycemia is completely avoided even after large meal disturbances. The proposed control approach is simple and the controller can be set up by a physician without the need for control expertise.

Automatic Detection of Stress States in Type 1 Diabetes Subjects in Ambulatory Conditions.

Two levels of control are crucial to the robustness of an artificial ß-cell, a medical device that would automatically regulate blood glucose levels in patients with type 1 diabetes. A low-level component would attempt to regulate blood glucose continuously, while a supervisory-level, or monitoring, component would detect underlying changes in the subject's glucose-insulin dynamics and take corrective actions accordingly. These underlying changes, or "faults," can include changes in insulin sensitivity, sensor problems, and insulin delivery problems, to name a few. A multivariate statistical monitoring technique, principal component analysis (PCA), has been applied to both simulated and experimental type 1 diabetes data. The objective of this study was to determine if PCA could be used to distinguish between normal patient data, and data for abnormal conditions that included a variety of "faults." The PCA results showed a high degree of accuracy; for data from nine type 1 diabetes subjects in ambulatory conditions, 33 of 37 total test days (89%), including fault days and normal days, were classified correctly. Thus, the proposed monitoring technique shows considerable promise for incorporation into an artificial ß-cell.

Quest for the artificial pancreas: combining technology with treatment.

The various components of the artificial pancreas puzzle are being put into place. Features such as communication, control, modeling, and learning are being realized presently. Steps have been set in motion to carry the conceptual design through simulation to clinical implementation. The challenging pieces still to be addressed include stress and exercise; as integral parts of the ultimate goal, effort has begun to shift toward overcoming the remaining hurdles to the full artificial pancreas. The artificial pancreas is close to becoming a reality, driven by technology, and the expectation that lives will be improved.

Experimental evaluation of a recursive model identification technique for type 1 diabetes.

BACKGROUND: A model-based controller for an artificial beta cell requires an accurate model of the glucose-insulin dynamics in type 1 diabetes subjects. To ensure the robustness of the controller for changing conditions (e.g., changes in insulin sensitivity due to illnesses, changes in exercise habits, or changes in stress levels), the model should be able to adapt to the new conditions by means of a recursive parameter estimation technique. Such an adaptive strategy will ensure that the most accurate model is used for the current conditions, and thus the most accurate model predictions are used in model-based control calculations. METHODS: In a retrospective analysis, empirical dynamic autoregressive exogenous input (ARX) models were identified from glucose-insulin data for nine type 1 diabetes subjects in ambulatory conditions. Data sets consisted of continuous (5-minute) glucose concentration measurements obtained from a continuous glucose monitor, basal insulin infusion rates and times and amounts of insulin boluses obtained from the subjects' insulin pumps, and subject-reported estimates of the times and carbohydrate content of meals. Two identification techniques were investigated: nonrecursive, or batch methods, and recursive methods. Batch models were identified from a set of training data, whereas recursively identified models were updated at each sampling instant. Both types of models were used to make predictions of new test data. For the purpose of comparison, model predictions were compared to zero-order hold (ZOH) predictions, which were made by simply holding the current glucose value constant for p steps into the future, where p is the prediction horizon. Thus, the ZOH predictions are model free and provide a base case for the prediction metrics used to quantify the accuracy of the model predictions. In theory, recursive identification techniques are needed only when there are changing conditions in the subject that require model adaptation. Thus, the identification and validation techniques were performed with both "normal" data and data collected during conditions of reduced insulin sensitivity. The latter were achieved by having the subjects self-administer a medication, prednisone, for 3 consecutive days. The recursive models were allowed to adapt to this condition of reduced insulin sensitivity, while the batch models were only identified from normal data. RESULTS: Data from nine type 1 diabetes subjects in ambulatory conditions were analyzed; six of these subjects also participated in the prednisone portion of the study. For normal test data, the batch ARX models produced 30-, 45-, and 60-minute-ahead predictions that had average root mean square error (RMSE) values of 26, 34, and 40 mg/dl, respectively. For test data characterized by reduced insulin sensitivity, the batch ARX models produced 30-, 60-, and 90-minute-ahead predictions with average RMSE values of 27, 46, and 59 mg/dl, respectively; the recursive ARX models demonstrated similar performance with corresponding values of 27, 45, and 61 mg/dl, respectively. The identified ARX models (batch and recursive) produced more accurate predictions than the model-free ZOH predictions, but only marginally. For test data characterized by reduced insulin sensitivity, RMSE values for the predictions of the batch ARX models were 9, 5, and 5% more accurate than the ZOH predictions for prediction horizons of 30, 60, and 90 minutes, respectively. In terms of RMSE values, the 30-, 60-, and 90-minute predictions of the recursive models were more accurate than the ZOH predictions, by 10, 5, and 2%, respectively. CONCLUSION: In this experimental study, the recursively identified ARX models resulted in predictions of test data that were similar, but not superior, to the batch models. Even for the test data characteristic of reduced insulin sensitivity, the batch and recursive models demonstrated similar prediction accuracy. The predictions of the identified ARX models were only marginally more accurate than the model-free ZOH predictions. Given the simplicity of the ARX models and the computational ease with which they are identified, however, even modest improvements may justify the use of these models in a model-based controller for an artificial beta cell.

An improved PID switching control strategy for type 1 diabetes.

In order for an "artificial pancreas" to become a reality for ambulatory use, a practical closed-loop control strategy must be developed and validated. In this paper, an improved PID control strategy for blood glucose control is proposed and critically evaluated in silico using a physiologic model of Hovorka et al. [1]. The key features of the proposed control strategy are: 1) a switching strategy for initiating PID control after a meal and insulin bolus; 2) a novel time-varying setpoint trajectory; 3) noise and derivative filters to reduce sensitivity to sensor noise; and 4) a practical controller tuning strategy. Simulation results demonstrate that proposed control strategy compares favorably to alternatives for realistic conditions that include meal challenges, incorrect carbohydrate meal estimates, changes in insulin sensitivity, and measurement noise.

A Feedforward-Feedback Glucose Control Strategy for Type 1 Diabetes Mellitus.

As the "artificial pancreas" becomes closer to reality, automated insulin delivery based on real-time glucose measurements becomes feasible for people with diabetes. This paper is concerned with the development of novel feedforward-feedback control strategies for real-time glucose control and type 1 diabetes. Improved post-meal responses can be achieved by a pre-prandial snack or bolus, or by reducing the glucose setpoint prior to the meal. Several feedforward-feedback control strategies provide attractive alternatives to the standard meal insulin bolus and are evaluated in simulations using a physiological model.

Use of continuous glucose monitoring to estimate insulin requirements in patients with type 1 diabetes mellitus during a short course of prednisone.

BACKGROUND: Insulin requirements to maintain normoglycemia during glucocorticoid therapy and stress are often difficult to estimate. To simulate insulin resistance during stress, adults with type 1 diabetes mellitus (T1DM) were given a three-day course of prednisone. METHODS: Ten patients (7 women, 3 men) using continuous subcutaneous insulin infusion pumps wore the Medtronic Minimed CGMS (Northridge, CA) device. Mean (standard deviation) age was 43.1 (14.9) years, body mass index 23.9 (4.7) kg/m(2), hemoglobin A1c 6.8% (1.2%), and duration of diabetes 18.7 (10.8) years. Each patient wore the CGMS for one baseline day (day 1), followed by three days of self-administered prednisone (60 mg/dl; days 2-4), and one post-prednisone day (day 5). RESULTS: Analysis using Wilcoxon signed rank test (values are median [25th percentile, 75th percentile]) indicated a significant difference between day 1 and the mean of days on prednisone (days 2-4) for average glucose level (110.0 [81.0, 158.0] mg/dl vs 149.2 [137.7, 168.0] mg/dl; p = .022), area under the glucose curve and above the upper limit of 180 mg/dl per day (0.5 [0, 8.0] mg/dl.d vs 14.0 [7.7, 24.7] mg/dl.d; p = .002), and total daily insulin dose (TDI) , (0.5 [0.4, 0.6] U/kg.d vs 0.9 [0.8, 1.0] U/kg.d; p = .002). In addition, the TDI was significantly different for day 1 vs day 5 (0.5 [0.4, 0.6] U/kg.d vs 0.6 [0.5, 0.8] U/kg.d; p = .002). Basal rates and insulin boluses were increased by an average of 69% (range: 30-100%) six hours after the first prednisone dose and returned to baseline amounts on the evening of day 4. CONCLUSIONS: For adults with T1DM, insulin requirements during prednisone induced insulin resistance may need to be increased by 70% or more to normalize blood glucose levels.

Practical issues in the identification of empirical models from simulated type 1 diabetes data.

BACKGROUND: A model-based controller for an artificial beta-cell automatically regulates blood glucose levels based on available glucose measurements, insulin infusion and meal information, and model predictions of future glucose trends. Thus, the identification of simple, accurate models plays an important role in the development of an artificial beta-cell. METHODS: Glucose data simulated from a nonlinear physiological model of type 1 diabetes are used to identify linear dynamic models of two types: autoregressive exogenous input (ARX) and output-error (OE) models. The model inputs are meal carbohydrates and exogenous insulin, which in practice are often administered simultaneously and in the same ratio, i.e., the insulin-to-carbohydrate ratio. The effect of modeling these inputs as impulses versus time-smoothed profiles ("transformed inputs") is explored in depth. The models are evaluated based on their ability to describe the data from which they were identified (i.e., calibration data) as well as independent data (i.e., validation data). RESULTS: In general, the best models described their calibration data more accurately using transformed inputs (R(Cal) (2) = 71% for the ARX models and R (Cal) (2) = 78% for the OE models) than using impulse inputs (R (Cal) (2) = 14% for the ARX models and R (Cal) (2) = 70% for the OE models). The only model/input combination that resulted in consistently accurate validation fits was the ARX models using transformed inputs (39%

Fault detection and diagnosis in an industrial fed-batch cell culture process.

A flexible process monitoring method was applied to industrial pilot plant cell culture data for the purpose of fault detection and diagnosis. Data from 23 batches, 20 normal operating conditions (NOC) and three abnormal, were available. A principal component analysis (PCA) model was constructed from 19 NOC batches, and the remaining NOC batch was used for model validation. Subsequently, the model was used to successfully detect (both offline and online) abnormal process conditions and to diagnose the root causes. This research demonstrates that data from a relatively small number of batches (approximately 20) can still be used to monitor for a wide range of process faults.