Supplementary Appendix

Transcription

1 Supplementary Appendix This appendix has been provided by the authors to give readers additional information about their work. Supplement to: Phillip M, Battelino T, Atlas E, et al. Nocturnal glucose control with an artificial pancreas at a diabetes camp. N Engl J Med 2013;368: DOI: /NEJMoa

2 1 APPENDIX TABLE OF CONTENTS Title Page 1. LIST OF INVESTIGATORS 2 2. METHODS 2.1. The MD-Logic Artificial Pancreas 2 system's alerts module 2.2. Secondary endpoints list Sensor Calibration protocol Estimation of hypoglycemia events 3 primary endpoint after sensor retrofitting and stochasting transformation 2.5. Power simulation protocol 4 3. SUPPLEMENTARY TABLES 6 4. SUPPLEMENTARY FIGURE REFERENCES 11

3 2 LIST OF INVESTIGATORS Members of the DREAM Consortium are as followed (asterisks indicate principal investigators): 1. The Jesse Z and Sara Lea Shafer Institute for Endocrinology and Diabetes, National Center for Childhood Diabetes, Schneider Children s Medical Center of Israel, Petah-Tikva, Israel: M. Phillip *, E. Atlas, R. Nimri, S. Miller, I. Muller, A. Farfel, S. Demol, T. Oron, G. Shiovitch-Mantzuri, M. Geva, G. Fayman, A. Hamou, O. Hermon, A. Liberman and R. Naveh 2. Department of Pediatric Endocrinology, Diabetes and Metabolism, University Medical Centre-University Children s Hospital, and Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia: T. Battelino *, N. Bratina, M. Avbelj Stefanija, A. Gianini, N. Godina and B. Mali 3. Diabetes Centre for Children and Adolescents, AUF DER BULT, Kinder- und Jugendkrankenhaus, Hannover, Germany: T. Danne *, O. Kordonouri *, T. Biester, Sarah Bläsig, Kerstin Remus, Bärbel Aschemeier and Christiana Tsioli METHODS The MD-Logic Artificial Pancreas System with an Alerts Module The main goal of an artificial pancreas is to regulate insulin delivery automatically, based on continuous glucose measurements (made by a glucose sensor) in order to achieve tight and safe glucose control. Thus, we have recently developed and introduced an alerts model into our artificial pancreas system. The alerts module uses different proprietary algorithms that are utilizing the individual patient s history of past glucose levels and insulin delivery plus models of insulin pharmacodynamics. We believe this provides this device with an advantage over the commercial glucose sensors alarms (that use glucose data alone), data from which were reported by Hughes CS et al 1. In addition to slowing the rate of insulin infusion if a risk for hypoglycemia is present, a decision programmed into the artificial pancreas controller, the alerts module offers real-time alarms that alert the patient about impending hypoglycemia (imminent hypoglycemia that cannot be prevented by slowing the rate of insulin infusion alone), longstanding hyperglycemia and technical errors (e.g., comparing sensor glucose reading to the glucomerer results to alarm sensor inaccuracies, or comparing instructed insulin

4 3 injections to subsequent glucose level alterations to alarm possible pump dysfunction or blockage). In many instances the alerts module alarms may result in a response before there is any clinical indication for intervention. Secondary Endpoints Secondary individual endpoints in the trial included the percentage of time sensor glucose level was within the range of mg/dl; the percentage of time sensor glucose level was within the range of mg/dl; the percentage of time glucose was within a tightly-defined normal range, defined as sensor glucose level within mg/dl; the average and median of blood glucose levels; the percentage of time glucose was below 63mg/dl; the percentage of time glucose was below 70mg/dl; the percentage of time glucose spent above 140 mg/dl; the percentage of time glucose spent above 180 mg/dl; the percentage of time glucose spent above 250mg/dl; the area under the curve for glucose values <70; the area under the curve for glucose values <63; the area under the curve for glucose values >140; the area under the curve for glucose values >180; the area under the curve for glucose values >250mg/dl (each calculated using the trapezoidal rule and then averaged over the overnight length of time, i.e. 480 minutes); the glucose variability as measured by the standard deviation of the glucose level (the glucose SD calculated for each night, i.e., 8 hours of sensor data which equals to 96 glucose values), and Kovatchev indexes; 2,3 the control variability grid analysis (CVGA) 4 ; the number of alerts during the night. Sensor Calibration Protocol During the study, the sensor calibration regiment was the same for the MD-Logic and control arms. There were no acceptance criteria prior to MD-logic nights while patients were instructed to calibrate their sensor at wake-up time, half an hour before dinner, and at bed time. One patient from group-a was excluded from ITT analyses due to a sensor crash, which disabled the allocated intervention and the delivery of any personalized sensor data (as reported in the manuscript). During the nights (of both arms), an automated algorithm was used to detect sensor errors incidences based on capillary glucose measurements. Capillary measurements were automatically transmitted to the system, where a specific algorithm was automatically activated comparing between capillary and sensor readings. Detection of sensor inaccuracy resulted in an alarm that inform the patient and staff that "sensor accuracy is in doubt.". This was followed by a calibration action, which was equally done in both groups.

5 4 Analyzing the number of patients who experienced sensor inaccuracy alarms and sensor recalibration during the night revealed no difference between the control nights (29/54 patients) and the MD-logic nights (26/54 patients). Estimation of the Hypoglycemia Events Primary Endpoint after Sensor Retrofitting and Stochastic Transformation In order to increase the confidence in the results established with the glucose, we reassessed the two hypoglycemia primary endpoints. Events below 63mg/dl We used the capillary values were taken at three-hour intervals and alarm episodes to reevaluate the number of events below 63mg/dl by defining an event using one of the following criteria: A single glucometer measurement of blood glucose level below 63 mg/dl. An accurate sensor-based event, validated by proximate before or after capillary measurements (according to availability of the data). This was rendered feasible because we have proximate before or after recorded capillary measurements for some sensor values below 63 mg/dl,. In such cases, the sensor specific values used to count an event were considered accurate (and were validated by capillary measurement readings) only if the relevant glucometer-sensor pairs (glucometer-sensor values before or after the event) fell in the A zone of the Clarke Error Grid analysis 5. Time Spent below Glucose Level of 60mg/dl We used stochastic transformation on the sensor glucose levels as proposed by Hovorka et al. 6 Stochastic continuous glucose monitoring transformation acknowledges that a continuous glucose monitoring reading measures true glucose with an error, which we assumef to be normally distributed with mean of the measured sensor value and a standard deviation which is dependent on the sensors coefficient of variation (sensor dependent value). The Enlite sensor (Medtronic Co.) coefficient of variation was set as 13.9%. 7 Power Simulation Protocol We conducted power simulations using a database consisting of 146 patients under standard treatment at home. The patients in the database were aged years, had a history of type 1 diabetes for at least one year and had used insulin pump therapy for at least three months.

6 5 They had glycated hemoglobin levels (A1c) of 7-10% and Body Mass Index for age below the 97 th percentile. The current study data set contains sensor values for several nights for each patient. Since the present study was a crossover study, and since we used non-parametric tests to compare groups, we used the within patient variability observed in the database to estimate the number of patients needed to achieve power for each primary outcome. The final sample size was that for the primary outcome with the largest required sample. The following steps describe the simulation method we used: 1. We defined a significance level of α= We calculated the power for a study with n patients; while n is changing from 20 to 100.For example, n 1 =20, n 2 =21,..., n 81 = For each n i we conducted 1000 iterations of the following procedure: a. We randomly chose n i patients from the database (for example, for n 1 =20, we randomly chose 20 patients from the 146 patient database) b. For each chosen patient, we selected 2 nights randomly: one was marked as the control night and the other as the MD-Logic night. c. We calculated the value of the primary endpoint from the reference data d. We estimated the impact of the MD-Logic system on the primary outcomes from the MD-Logic night data, as follows: i. For number of events below 63 mg/dl we reduced the number of events by 50% ii. For time below 60 mg/dl we reduced the observed time by 60% iii. For mean overnight glucose levels we reduced the mean value by 13mg/dl. iv. We collected all pairs and calculate the P-value from comparing control nights to estimate MD-Logic nights using the two-sided, paired nonparametric Wilcoxon sign rank test. 4. For each n i, we thus obtained 1000 simulated P values, defined as P_ValueSet. We calculated the power for each n i as follows: Power ni =[ P_ValueSet<0.05]/ We plotted the Power ni against n i and chose n i todeliver power of 90% In this way our power calculation reflects the actual level of variability in outcomes according to the study design and the statistical test that was chosen.

7 6 SUPPLEMENTARY TABLES Table S1: Insulin Treatment analysis ITT patients (N=54) Variable Artificial Pancreas Nights Ŧ Control Nights Total insulin dose [Units] 7.8 (4.9 to 11.0)* 5.4 (4.1 to 7.3)* Total insulin basal dose [Units] 6.1 (4.3 to 7.8) 5.4 (3.5 to 7.1) Pre-dinner + snack insulin bolus dose [Units] 5.1 (2.6 to 9.8) 5.5 (3.5 to 10.1) Ŧ --Values with brackets are median (interquartile range) * --Significant differences of P<0.05 between MD-Logic and control night after controlling for multiple testing Ŧ

8 7 Table S2: Carbohydrate Interventions Variable Artificial Pancreas Nights Control Nights Total carbohydrate interventions Carbohydrate interventions per site: Site Site Ŧ Site Interventions after an alarm that occurred while capillary glucose level was 3/26 5/25 <60mg/dl Interventions not related to alarms that occurred while capillary glucose level was 1/26 4/25 <60mg/dl Amount of carbohydrates per treatment (Mean ± SD) 24.4 ± ± 9.4 Ŧ There were additional 5 cases where the physician decided to suspend insulin delivery instead of treating with carbohydrates. Was determined retrospectively according to the capillary glucose value at the time of the intervention. In most cases the sensor reading was above the hypoglycemic range.

9 8 Table S3: Glucose values at the time of carbohydrates interventions Variable Artificial Pancreas nights Ŧ Control nights Ŧ Capillary glucose level [mg/dl] 80.8 ± ± 14.5 Glucose sensor level [mg/dl] 77.7 ± ± 12.8 Ŧ Values are mean ± SD Table S4: Hypoglycemia Events below 63mg/dl after Manual Retrofitting of the Sensor Variable Artificial Pancreas Nights Control Nights An event based on a single capillary measurement (First criteria) 6 9 A sensor-based event, validated by proximate before or after capillary 0 10 measurements (Second criteria) Total 6* 19* * P=0.0123, Wilcoxon sign-rank test

10 9 Table S5: Time In Range Endpoints ITT Analysis in the 54 Patients Variable Artificial Pancreas Ŧ Control Ŧ Primary Endpoint Time glucose level spent below 60 mg/dl [min] Secondary Endpoints Glucose Control 1.2 (0-5.1)* 5.5 (0-40.6)* Time within mg/dl [h] 5.2 ( )* 3.6 ( )* Time within mg/dl [h] 5.4 (4-6.4)* 3.8 ( )* Time within mg/dl [h] 3 ( )* 2 ( )* Hypoglycemia Time < 70mg/dl [min] 7.6 ( ) 16.4 ( ) Time < 63mg/dl [min] 2.3 ( )* 8.1 ( )* Hyperglycemia Time > 140mg/dl [min] ( ) ( ) Time > 180mg/dl [min] 18.7 ( ) 66.9 ( ) Time > 250mg/dl [min] 0 (0-2.1)* 0.1 (0-46.6)* Ŧ-- Values with brackets are median (interquartile range) *-- Significant differences of P<0.05 between artificial pancreas and control night after controlling for multiple testing performed separately from primary and secondary endpoints.

11 11 SUPPLEMENTARY FIGURE Assessed for eligibility (n=56) Randomized (n=56) Group A (n= 28), Allocated to closed-loop night and then control night Group B (n= 28), Allocated to control night and then closed-loop night Did not receive the allocated intervention (CGM collapsed early on) (n= 1) Received allocated intervention (n= 27) Received allocated intervention (n=28) Discontinued intervention (withdrawal) (n=1) ITT analyses (n= 27). The following detail the number of patients in this group per site: Site 1 8 patients Site 2 10 patients Site 3 9 patients ITT analyses (n= 27). The following detail the number of patients in this group per site: Site 1 8 patients Site 2 10 patients Site 3 9 patients Figure S1: Participant Flow Diagram: Outlined for each group the numbers of participants who were randomly assigned, received intended treatment, and were analyzed for the primary and secondary outcomes (ITT analyses). Exclusions and discontinued intervention after randomization are indicated with the reasons for each group.

12 11 REFERENCES 1. Hughes CS, Patek SD, Breton MD, Kovatchev BP. Hypoglycemia prevention via pump attenuation and red-yellow-green "traffic" lights using continuous glucose monitoring and insulin pump data. J Diabetes Sci Technol 2010;4: Kovatchev B, Otto E, Cox D, Gonder-Frederick L, Clarke W. Evaluation of a new measure of blood glucose variability in diabetes. Diabetes Care 2006;29: McCall A, Cox D, Crean J, Gloster M, Kovatchev B. A novel analytical method for assessing glucose variability: using CGMS in type 1 diabetes mellitus. Diabetes Technol Ther 2006;8: Magni L, Raimondo D, Dalla Man C, et al. Evaluating the Efficacy of Closed-Loop Glucose Regulation via Control-Variability Grid Analysis. J Diabetes Sci Technol 2008;2: Kovatchev BP, Gonder-Frederick LA, Cox DJ, Clarke WL. Evaluating the accuracy of continuous glucose-monitoring sensors: continuous glucose-error grid analysis illustrated by TheraSense Freestyle Navigator data. Diabetes Care 2004;27: Hovorka R, Nodale M, Haidar A, Wilinska ME. Assessing Performance of Closed-Loop Insulin Delivery Systems by Continuous Glucose Monitoring: Drawbacks and Way Forward. Diabetes Technol Ther Keenan DB, Mastrototaro JJ, Zisser H, et al. Accuracy of the Enlite 6-day glucose sensor with guardian and Veo calibration algorithms. Diabetes Technol Ther 2012;14: