Prof. Dr.-Ing. Bert Arnrich

Analysis of Mealtime Insulin Bolus Behaviors of Users of The Open Artificial Pancreas System (OpenAPS) with Type 1 Diabetes

Master's Thesis

Wiktoria Staszak, Supervisor: Jonas Chromik

In recent years, the increase in commercial availability, as well as improvement in accuracy of diabetes management aids such as insulin pumps and continuous glucose monitors (CGM) has changed many aspects of managing type 1 diabetes [6]. In particular, the introduction of closed-loop systems, i.e. the enablement of communication between the CGM and insulin pumps which aims to partially mimic the function of a pancreas, have been shown to greatly reduce the burden of disease management for the user [8].

However, despite such advances in technology, many tasks must still be done manually by the user. One such task is the administration of mealtime bolus. The administration of correct insulin doses is a crucial factor in maintaining good control over blood glucose levels, i.e. avoiding hypo- and hyperglycemia. Therapy with insulin usually comprises of a basal dose and mealtime bolus. The daily basal dose pattern remains relatively constant from day to day, and in closed loop systems, the task of making slight corrections to the basal rate is automated by creating a feedback loop between CGM and the insulin pump. However, the mealtime bolus requirements vary, depending on the carbohydrate content of the meal, as well as additional factors such as physical exercise, the insulin sensitivity ratio and the amount of insulin which is currently active in the body due to previously administered doses [2]. While closed-loop systems are effective in properly adjusting basal rates, the user is still left with the task of correctly estimating the mealtime bolus [10]. Previous studies have shown that over half of patients do not estimate this dose correctly, and less than 30% of people with diabetes achieve optimal glycemic control [1, 12, 7].

Maintaining good blood glucose control is important for avoiding long term complications of diabetes. Avoiding complications arising from poor glycemic control of people with diabetes does not only help to reduce the premature deaths due to complications such as nephropathy and improve the quality of life of people with diabetes, but can also help to reduce heavy economic burden on the global healthcare system, where the annual 825 billion US dollar costs of managing diabetes are mostly attributed to treating of diabetes complications [6]. This thesis sets out to explore a real world dataset comprised of glucose values from CGM and insulin dose data from insulin pumps from an open source project Open Artificial Pancreas System (OpenAPS), with the aim of gaining better understanding of the behavior of users of closed loop systems in adminis-
tering their mealtime insulin.

Figure 1: A graphical comparison of four models of mealtime bolus administration, available as preset modes in most commercially available insulin pumps [9].

Related Work

The topic of achieving optimal glycemic control through proper dosing of insulin has been extensively researched. While similar analyses of have been performed, most focus on finding a bolus method which works best for a particular meal composition, in a clinical setting. Figure 1 shows four examples of bolus methods examined in literature, summarized by Heinemann [9]. For example, Chase and colleagues [4] examined the glycemic response of nine subjects after a high-carbohydrate, high-fat meal. The four modes used in this study were a single bolus, two boluses of 50% of dose each and delayed bolus over 2 hours, and combination bolus of 70% immediately and the rest over 2 hours. This study concluded that the combination bolus was most successful in controlling post-mealtime blood glucose levels for meals with a high fat and high carbohydrate content.

A similar study has been carried out by Lee and colleagues which aimed to examine the glycemic control of 10 subjects with the following three meal and bolus mode combinations, a high-fat meal with normal bolus delivery, a high-fat meal with combination bolus, a control meal with single bolus delivery, and a control meal with single bolus delivery [11]. This study also reached the conclusion that for high-fat, high-carbohydrate meals, the combination bolus leads to better results than the single bolus. Both studies were carried out in a clinical setting and focused only on meals with a high-fat and high-carbohydrate content.

Several limitations of these studies make it difficult to translate to real-world, outpatient settings. Firstly, both studies were carried out in a clinical setting and focused only on meals with a high-fat and high-carbohydrate content. Secondly, subjects in both experiments had in-range blood glucose values in the hours prior to the meal, which may not always be the case in real life. Furthermore, neither study considered the timing of the bolus administration. As pointed out by Heinemann [9], both studies fails to present the full picture, as in real life, people tend to consume more than one meal a day, and hence may have active insulin in their system, highlighting the need of similar studies to be performed using real-world data.

The only similar analysis which uses was performed on users of closed-loop systems is a 2014 study by Chase and colleagues. In this study, the glycemic reaction of 53 patients was measured to the following four scenarios, standard bolus, standard bolus delivered 15 minutes prior to the meal, over-bolus of 30% and a purposefully omitted bolus. This study focused more on testing whether the artificial pancreas system is able to handle errors in bolus estimation, rather than finding the bolus patterns which lead to best outcomes [5]. One conclusion of the study was the need to perform similar analyses in real-life outpatient scenarios, in order to consider factors such as stress, exercise and different types of meals.

Other novel approaches for improving postprandial glycemic control, while also reducing the disease management burden, include the use of machine learning to predict insulin doses. For example, Cappon and colleagues propose an approach for personalizing bolus calculations using CGM data through the use of neural networks [3].


The goal of this thesis is to expand the current knowledge on mealtime bolus administration for achieving optimal glycemic control by analysing the data donated by users of OpenAPS through the Open Humans Platform. This dataset contains data from CGMs, insulin pumps as well as additional information such as physical exercise and other user-defined events. By analysing various patterns in the behaviour of users in administering their mealtime insulin and comparing them with long and short term health outcomes in terms of achieving glycemic control. Examples of outcomes which can be evaluated based on this data set include postprandial glucose, average blood glucose and its variability as well as time spent in target range. The expected outcome of this thesis is a clustering of users based on their behaviour regarding mealtime insulin and their disease management outcomes for various types of everyday scenarios. Furthermore, this is the first such analysis to use an open source project with data donated by its users. The successful identifications of behaviours which lead to best outcomes can be provided to users of closed-loop systems as a valuable guide for self-management of the condition, as well as play a crucial role in the development of future closed-loop system algorithms.


[1] Aila J. Ahola et al. “Many patients with Type 1 diabetes estimate their prandial insulin need inappropriately”. In: Journal of Diabetes 2.3 (Aug. 2010), pp. 194–202. DOI : 10.1111/j.1753-0407.2010.00086.x. URL: https://doi.org/10.1111/j.1753-0407.2010.00086.x.

[2] Kate Van Brunt et al. “Behaviours, thoughts and perceptions around mealtime insulin usage and wastage among people with type 1 and type 2 diabetes mellitus: A cross-sectional survey study”. In: Diabetes Research and Clinical Practice 126 (Apr. 2017), pp. 30–42. DOI: 10.1016/j.diabres.2016.12.002. URL: https://doi.org/10.1016/j.diabres.2016.12.002.

[3] Giacomo Cappon et al. “A Neural-Network-Based Approach to Personalize Insulin Bolus Calculation Using Continuous Glucose Monitoring”. In: Journal of Diabetes Science and Technology 12.2 (Mar. 2018), pp. 265–272. DOI: 10.1177/1932296818759558. URL: https://doi.org/10.1177/1932296818759558.

[4] H. P. Chase et al. “Post-prandial glucose excursions following four methods of bolus insulin administration in subjects with Type 1 diabetes”. In: Diabetic Medicine 19.4 (Apr. 2002), pp. 317–321. DOI: 10.1046/j.1464-5491.2002.00685.x. URL: https://doi.org/10.1046/j.1464-5491.2002.00685.x.

[5] H. Peter Chase et al. “Multicenter Closed-Loop/Hybrid Meal Bolus Insulin Delivery with Type 1 Diabetes”. In: Diabetes Technology & Therapeutics 16.10 (Oct. 2014), pp. 623–632. DOI: 10.1089/dia.2014.0050. URL: https://doi.org/10.1089/dia.2014.0050.

[6] G. Fagherazzi and P. Ravaud. “Digital diabetes: Perspectives for diabetes prevention, management and research”. In: Diabetes & Metabolism 45.4 (Sept. 2019), pp. 322–329. DOI: 10.1016/j.diabet.2018.08.012. URL: https://doi.org/10.1016/j.diabet.2018.08.012.

[7] Nicole C. Foster et al. “State of Type 1 Diabetes Management and Outcomes from the T1D Exchange in 2016–2018”. In: Diabetes Technology & Therapeutics 21.2 (Feb. 2019), pp. 66–72. DOI: 10.1089/dia.2018.0384. URL: https://doi.org/10.1089/dia.2018.0384.

[8] Satish K Garg and Viral Shah. “Closed-loop insulin delivery systems for patients with
diabetes”. In: The Lancet Digital Health 1.1 (May 2019), e2–e3. DOI : 10 . 1016 / s2589 -
7500(19)30007-x. URL : https://doi.org/10.1016/s2589-7500(19)30007-x.

[9] Lutz Heinemann. “Insulin Pump Therapy: What is the Evidence for Using Different Types of Boluses for Coverage of Prandial Insulin Requirements?” In: Journal of Diabetes Science and Technology 3.6 (Nov. 2009), pp. 1490–1500. DOI: 10.1177/193229680900300631. URL: https://doi.org/10.1177/193229680900300631.

[10] Rayhan A Lal et al. “Realizing a Closed-Loop (Artificial Pancreas) System for the Treatment of Type 1 Diabetes”. In: Endocrine Reviews 40.6 (July 2019), pp. 1521–1546. DOI: 10.1210/er.2018-00174. URL: https://doi.org/10.1210/er.2018-00174.

[11] S.W. Lee et al. “The dual-wave bolus feature in continuous subcutaneous insulin infusion pumps controls prolonged post-prandial hyperglycaemia better than standard bolus in Type 1 diabetes”. In: Diabetes, Nutrition Metabolism 17.4 (Aug. 2004), pp. 211–216. URL: https://pubmed.ncbi.nlm.nih.gov/15575341/.

[12] Allen Sussman et al. “Performance of a Glucose Meter with a Built-in Automated Bolus Calculator versus Manual Bolus Calculation in Insulin-Using Subjects”. In: Journal of Diabetes Science and Technology 6.2 (Mar. 2012), pp. 339–344. DOI: 10.1177/193229681200600218. URL: https://doi.org/10.1177/193229681200600218.