(702f) A Meal Detection Algorithm Based on Continuous Glucose Measurements for Use in Artificial Pancreas Systems

A Meal Detection Algorithm Based on Continuous Glucose Measurements for use in Artificial Pancreas Systems

Sediqeh Samadi 1, Kamuran Turksoy 2, Jianyuan Feng1,Iman Hajizadeh1, Ali Cinar1, 2

1Department of Chemical and Biological Engineering, Illinois Institute of Technology, 10 West 33rd St., Chicago, IL 60616

2Department of Biomedical Engineering, Illinois Institute of Technology, 3255 S. Dearborn St., Chicago, IL 60616

Analysis of the process data and identification of temporal features caused by various underlying changes in process conditions are essential tasks in engineering problems such as detection, diagnosis and control problems. Historical process data presents information about the effect of the major process changes on observed variables. Modeling the effects of changes either quantitatively or qualitatively is useful for process simulation as well as real time change detection and identification. Once the changes are identified, the controller can make the appropriate remedial actions. In this work, trend analysis and qualitative model development are used to detect the effects of meals for use in artificial pancreas systems (AP). AP systems offer an important alternative for automating the regulation of blood glucose concentration (BGC) of patients with Type 1 Diabetes (T1D). But AP systems may infuse too much insulin that can cause hypoglycemia. Accurate closed-loop control is essential for developing AP systems that adjusts insulin infusion rates of insulin pumps and a multivariable AP developed at IIT was successful in reducing the number of hypoglycemic episodes [1,2]. Various known disturbances such as meals affect the BGC. Traditionally an insulin bolus is computed manually to balance the effects of a meal on BGC and administrated. However, the elimination of manual inputs to AP systems would be very appealing to patients. In AP systems, one of the biggest challenges is automatically handling a meal’s effect that causes deviation of BGC from control target range. An AP control system equipped with a meal detection module can provide accurate predictions of change in BGC for use in control decisions. The current work focuses on the development of a meal detection algorithm that can be incorporated to AP control systems. In the meal prediction system designed, frequently measured subcutaneous glucose concentration from a Continuous Glucose Monitoring (CGM) system provides data both to the controller and the meal detection module. The proposed meal detection algorithm has two main steps. In the first step, the meal’s effect on BGC variation is modeled by some qualitative variables. To do that qualitative trend analysis (QTA) method transforms glucose concentration time series to a sequence of non-overlapped segments. To each segment a member of a small set of qualitative variables is assigned. Each qualitative variable has a different combination of first and second derivative signs. The sequence of segments labeled with such variables called qualitative representation of CGM signal gives valuable information about glucose variation. A single qualitative variable and also a sequence of them may indicate the signs of meal’s effect on glucose change and they are the essence of proposed detection algorithm. In the second step, a multi-stage clustering method separates the false detections from the true ones. The method is tested on real patients’ data. The meals are detected before a remarkable increase in glucose concentration takes place.

The average increase in BGC between the beginning of the meal (known in the experiments conducted) and the time it is detected varies within the range of 10-30 mg/dl for different patients. The largest detection change was observed to be 50 mg/dl, which belongs to a patient whose change in blood glucose is rapid. Less than 10% of test meals (snacks) are not detected, and almost all are snacks. The number of false positive detection alarms is in range of 1-2 for different patients within around 50 hours experimental data.  The impact of the proposed automated meal detection technique will be illustrated by including a module in our AP and comparing the characteristics of BGC profiles under closed-loop control with and without the automated meal detection module.

[1] Turksoy K, Bayrak E.S, Quinn L, Littlejohn E, Cinar A. Multivariable Adaptive Closed-Loop Control of an Artificial Pancreas Without Meal and Activity Announcement. Diabetes Technology & Therapeutics 2013. 15(5), 386-400.

[2] Turksoy K, Quinn L, Littlejohn E, Cinar A. An Integrated Multivariable Artificial pancreas Control System. Journal of Diabetes Science and Technology 2014. 8(3), 498–507.