(344a) Medically Inspired Benchmarks for Hypoglycemic Event Prediction and Alarming

Type 1 diabetes mellitus (T1DM) is a metabolic disease characterized by the destruction of the insulin-producing beta cells in the pancreas. As endogenous insulin is either lacking or completely absent in people with T1DM, exogenous insulin administration is required to regulate the concentration of glucose in the blood. The goal of diabetes management is to maintain blood glucose (BG) near normal levels and thus to avoid both immediate and long-term complications. In healthy individuals, insulin exists in a feedback loop with glucose to maintain a tight range of BG concentration. In those with T1DM, exogenous insulin must be administered by the patient; this is a procedure that often results in poor control. Severe adverse symptoms can be experienced if BG is not well controlled. Hyperglycemia, or high blood glucose, can result in long term complications such as blindness and kidney, vascular, and nerve diseases. To avoid long term problems, intensive therapy is recommended, which decreases the occurrence of hyperglycemia but increases the risk of hypoglycemia, a fast-acting, life-threatening condition (The Diabetes Control and Complication Trial Research Group). Hypoglycemia can cause dizziness and confusion, in the short term, to coma and death if severe and prolonged (Widmaier, Raff and Strang). Treatment of T1DM requires either multiple daily insulin injections or continuous subcutaneous insulin infusion delivered via an insulin infusion pump. Treatment necessitates frequent BG measurements (8 ? 10 times/day, including fasting, pre- and post-prandial, before bedtime and in the middle of the night) to determine the daily insulin requirements for maintaining near normal BG levels. This is a tedious process and is fraught with errors and problems: errors in sensor readings, incorrect insulin dosing, and unexpected meals and exercise, for instance, can cause fluctuations outside of the normal range (Jovanovič). Many researchers are devoting their efforts for controlling T1DM to the artificial beta cell (ABC) project, a closed-loop system that will allow for better control with less patient input. In recent years, the advent of continuous glucose monitors (CGMs), which report interstitial glucose concentrations approximately every minute, and the development of hardware and software to communicate with and control insulin pumps, has allowed ABC development to occur quickly. In individuals without diabetes, blood glucose is controlled by various neural and hormonal inputs from the brain, gut, liver, and pancreas that respond to various situations such as meals, stress, and exercise. Likewise, a closed-loop system for T1DM must be responsive to all daily challenges in life and be able to accurately predict blood glucose levels in advance. The ABC will include a combination of features necessary to bring a person with diabetes as close as possible to normoglycemia using subcutaneous insulin therapy (Harvey et al.). The center of the ABC is the control algorithm designed to bring the BG into a normal pattern. As in any process, the controller may not always perform perfectly, and abnormal situations may arise. Therefore, monitoring and safety becomes an essential part of the system. In the case of T1DM, violating emergency limits, such as the hypoglycemia threshold, can be life-threatening. A safety algorithm independent of the controller is required to accurately predict hypoglycemia and warn the patient in time to prevent the event or treat the event. The design of a hypoglycemia safety algorithm can be taken from two points of view: process design and clinical design. Using classic process design methods that are constrained by physiologic limits allows optimization for this specific problem. The Numerical Logic Algorithm (NLA) is a hypoglycemia algorithm designed with clinical factors in mind (Dassau et al.; Buckingham et al.). Currently, the only input is a vector of data from CGMs, which frequently exhibit spikes and high frequency noise. To cope with this, frequent assessment of the signal is performed with constraints designed to maximize the usefulness of the CGM data. Five major constraints are used to focus on hypoglycemia events: 1) maximum BG limit for algorithm operation; 2) rate of change limit; 3) prediction horizon; 4) prediction threshold; and 5) successive violation requirement. The maximum BG limit acts to reduce unnecessary computation and the occurrence of extremely early alarms: if the last point is above this limit, the algorithm does not execute. This is because the BG will not decrease fast enough to cause alarm if starting from above this limit. If this test is passed, the rate of change is estimated using a three-point backward difference calculation. A limit on rate of change (between -3 and 0 mg/dL/min) is imposed, as a drop faster than 3 mg/dL/min is considered non-physiologic for an adult. This constraint helps to filter out noise spikes. The prediction horizon and threshold are used as parameters to alarm: if the prediction crosses the threshold within the horizon, it is considered a violation. Two successive violations are required to flag an alarm; this also decreases the effect of sensor noise. NLA has been evaluated using a historical data record to validate the ability of the algorithm to alert patients to impending hypoglycemia. This record contained data from 30 patients in ambulatory conditions with a CGM with a five minute sampling period. Several metrics have been developed during this study to accurately report the performance of the algorithm. Design goals for an algorithm include a small footprint for implementation into the ABC and production of alarms with a high true positive rate, low false positive rate, and a warning time that allows corrective action to be taken. Metrics have been developed to address these goals, and include determining the best method of calculating true and false positive rates, and a developing a consistent way of reporting warning times. Preliminary results from the a historical study show that NLA produces less than 1 false positive alarm per day using a 15 minute prediction horizon and 60 mg/dL prediction threshold (all following results are under these conditions). Of all true positive alarms, 66% occurred within 15 minutes and 91% occurred within 30 minutes of the event. The time of the first alarm for each event was determined. When including alarms up to 120 minutes before an event, the first alarms occurred within 15 minutes 61% of the time, within 30 minutes 78% of the time, and within 60 minutes 88% of the time. The rate of first alarms (observation rate) was calculated within bins of 15, 30, 60, 90, and 120 minutes before an event. When normalized by the maximum observation rate, the distribution among these bins is readily seen. In the figure below, the relative rate of observations (# observations/minute) is shown for NLA with a prediction threshold of 60mg/dL and prediction horizon of 15 minutes. The diameter of the rings corresponds to the amount of time from the event (written in white), while the color corresponds to the relative rate. The maximum rate occurs within 15 minutes, while rates decrease rapidly to 27, 7, 6, and 4 % of the maximum in the outer windows. This result demonstrates the effect of tuning the algorithm with a short prediction horizon. The goal is to alarm the patient within a short window of the event to ensure maximum response and lower the risk of false positive alarms. The overall goal of the ABC project is to maintain normoglycemia using a controller with input from a continuous glucose monitor and output to an insulin delivery device in a closed-loop fashion. This is complicated by disturbances such as meals and exercise, as well as time lags and sensor error. A safety layer to prevent or alert the user to potentially dangerous events is essential. In contrast to classic process control problems, this safety measure relies on a single sensor that is prone to error. NLA utilizes physiologic constraints to reduce the effect of erroneous sensor reading on the prediction. This type of robust algorithm that can alert the patient to real impending danger will provide a much-needed safety layer to the ABC. References Buckingham, Bruce, et al. "Prevention of Nocturnal Hypoglycemia Using Predictive Alarm Algorithms and Insulin Pump Suspension." Diabetes Care 33.5 (2010): 1013-17. Dassau, E., F. Cameron, H. Lee, B. W. Bequette, H. Zisser, L. Jovanovič, H. P. Chase, D. M. Willson, B. A. Buckingham, and F. J. Doyle III. "Real-Time Hypoglycemia Prediction Using Continuous Glucose Monitoring (Cgm), a Safety Net to the Artificial Pancreas." Diabetes Care (2010). Harvey, R. A., Wang Youqing, B. Grosman, M. W. Percival, W. Bevier, D. A. Finan, H. Zisser, D. E. Seborg, L. Jovanovič, F. J. Doyle III, and E. Dassau. "Quest for the Artificial Pancreas: Combining Technology with Treatment." Engineering in Medicine and Biology Magazine, IEEE 29.2 (2010): 53-62. Jovanovič, L. "Rationale for Prevention and Treatment of Postprandial Glucose-Mediated Toxicity." The Endocrinologist 9 (1999): 87-92. The Diabetes Control and Complication Trial Research Group. "The Effect of Intensive Treatment of Diabetes on the Development and Progression of Long-Term Complications in Insulin-Dependent Diabetes Mellitus." N Engl J Med 329 (1993): 977-86. Widmaier, Eric P., Hershel Raff, and Kevin T. Strang. Vander's Human Physiology: The Mechanisms of Body Function. 11th ed. Boston: McGraw-Hill Higher Education, 2008.