Objective
Controller design optimization phase
In year 1, we will optimize the Tidepool Loop controller parameters to minimize the increased risk of hyperglycemia at the onset of meal intake and hypoglycemia after meal intake with no meal announcement. The Tidepool Loop controller has several internal parameters that control how the algorithm responds to deviations in blood glucose. These include, but are not restricted to, the weight assigned to glucose momentum around a meal occurrence, the carb-to-insulin ratio defining how much carbohydrate one unit of insulin can cover, and the speed of the decay of the expected retrospective correction effect. The initial goal will be to identify the combination of these parameters that enable Loop to rapidly respond to increases in blood glucose while safely avoiding hypoglycemia. As necessary, we will leverage our high-throughput simulation platform to evaluate other modifications to improve how the Loop algorithm detects and responds to unannounced meals. At the end of this phase, we will have a candidate version of Loop designed to quickly and safely respond to unannounced meals.
Conduct in-silico evaluation
In year 1, following the optimization phase, we will design and conduct an in-silico evaluation to assess the performance of the optimized Loop algorithm for unannounced meals. We will use our simulation platform to evaluate how optimizing the Tidepool Loop control parameters affects how well the algorithm compensates for unannounced meals of different sizes across a range of different metabolic profiles (virtual patients) and insulin action profiles in real-life scenarios. Additionally, we will conduct a sensitivity analysis of the Tidepool Loop parameters to precisely determine their contribution in achieving safely and effectively control for glucose deviations caused by unannounced meals, in preparation for the clinical study.
Deliver Tidepool Loop v3.0 to pivotal trial to support regulatory submission for FDA clearance and CE marking with no meal announcement requirements
Tidepool, makers of the FDA cleared interoperable automated glycemic controller Tidepool Loop, will implement the project team’s algorithm changes into the Tidepool Loop application, including an updated version of the Tidepool Loop app to be used in the clinical study.
Stanford will lead the clinical trial. Tidepool will begin the clinical trial component of the project in year 2. The clinical trial will be a multicenter US clinical trial that will use the modified Tidepool Loop algorithm optimized for no meal announcement. Participants will eat without meal announcement and some will separately record approximate carbohydrate intake.
The real-world, at-home pivotal study of the Tidepool Loop v3.0 system will be designed to support regulatory submission for FDA clearance, including labeling for use without meal announcements, and CE marking.
Background Rationale
Even with the most recent advances in diabetes treatment and technology, i.e., AID systems, most people with T1D do not meet recommended glycemic targets and, strikingly, continue to have a dramatically increased risk of developing premature cardiovascular disease. To date, four pump manufacturers have cleared/approved AID systems for commercial use in the US: Medtronic, Tandem, Insulet, and Beta Bionics. Real-world data for the OmniPod 5 from 36,634 adults shows that only 46% of users meet all glycemic targets, even with the entry of carbohydrates. Tidepool received clearance in the US for the Tidepool Loop iAGC. Despite their sophistication, these systems still require manual entry for insulin doses at mealtime. AID without meal announcements could alleviate these struggles.
Additional Background on Real-World Datasets Available to Support Enhancing Tidepool Loop for Unannounced Meals
We will leverage a unique combination of real-world datasets to advance in-silico testing and controller design optimization for the Enhancing Tidepool Loop for Unannounced Meals project.
First, the Tidepool Big Data Donation Project provides a rich source of anonymized patient data, including glucose, insulin, carbohydrate intake, etc. This extensive dataset, contributed by over 50,000 PwDs, will be used to refine simulation models. This data includes a large breadth of diabetes device data, including basal, bolus, glucose and carbohydrates.
Second, a Tidepool-sponsored observational study of DIY Loop users evaluating carbohydrate absorption models in H12025 in collaboration with Replica Health will capture detailed information about meal patterns, physical activity, and physiological responses in 30+ participants using DIY Loop over a period of 30 days. This data will be used to evaluate the performance of models of carbohydrate absorption and postprandial glucose response used in a system comparable to Tidepool Loop v2.0. By analyzing data from a diverse population, these data will be used to inform future algorithm parameter optimizations in v3.0.
The data set from each participant will include: insulin, carbs, exercise, heart rate, location, steps and calories burned. The dataset will include participants diverse in age, gender, diabetes type and date since diagnosis and will also include participants that use medications that impact gastric emptying (e.g. GLP-1s) and will not exclude individuals with comorbidities that may impact carb absorption (e.g. gastroparesis). This data set will allow us to:
- Analyze in-silico performance of the algorithm parameter adjustments with and without the presence of selected external variables, such as physical activity
- Classify the performance of various meal-detection algorithms
- Perform subanalyses on different types of meal-time and exercise behaviors, where we will compare the range of postprandial glucose effects by cluster based on different meal-time behavior
Third, the data collected by Jaeb and Tidepool during the 2019 Observational Study of DIY Loop will be used as an additional data source contributing longitudinal datasets of more than 800 participants using the DIY Loop app for at least 6 months.
Finally, publicly available datasets, such as T1-DEXI and T1-DEXIP Data Sets may complement our analysis by providing insights into the impact of exercise on glucose control. T1-DEXI represents a unique dataset to investigate the factors around the activity as it provides multiple observations, including both structured and unstructured exercise sessions, as well as unplanned activities, over a long period of time for 497 adult participants with T1D using MDI or insulin pump.
With these data sources, we will create a more comprehensive understanding of the factors influencing glucose levels and develop more robust algorithms to address the challenges of unannounced meals. The data present a remarkable opportunity to analyze, investigate, and model the body’s response to meals.
Description of Project
Community-derived, open-source automated insulin dosing systems were developed by individuals with a personal interest in insulin delivery for type 1 diabetes (T1D). One such system called “Loop” was developed by Nate Racklyeft, Pete Schwamb, and others. Loop is an open-source application for automated insulin dosing using an iPhone, compatible insulin pump and continuous glucose monitor (CGM). The Loop Observational Study, conducted by the Jaeb Center for Health Research (JCHR), evaluated the safety and effectiveness of the Loop system in 558 children and adults with T1D. The data collected in the study supported FDA clearance of Tidepool Loop as an automated insulin dosing application for management of individuals with T1D.
In January 2023, Tidepool received FDA clearance for Tidepool Loop v1.0. The Loop algorithm is a model predictive controller and functions optimally with announcements of both carbohydrate count, time, and carbohydrate absorption rate. While the algorithm makes use of several forms of short term adaptation that can foster insulin delivery for unannounced meals, they are not yet optimized for this use case.
A review of opportunities and challenges (Wilson, 2022; Aiello, 2021) and discontinuers of similar AID systems (Messer, 2020) highlights therapy burden due to requirements to announce meals, manage alarms, and optimize and fine-tune therapy setting parameters. Approaches to automated meal detection and estimation have been described in literature (Samadi, 2017; Mosquera-Lopez, 2023; Gingras, 2018).
This proposal supports the algorithm optimization, feature development, and clinical validation required to support unannounced meals, which will then be released within Tidepool Loop v3.0. This enhancement specifically aims to reduce significant postprandial hyperglycemic excursions and late postprandial hypoglycemia, leading to improvement in postprandial glycemic outcomes and reduction of therapy burden. Importantly, this project is intended to alleviate cognitive and emotional burden associated with meal glucose regulation and simplify day-to-day management of T1D.
Tidepool and our collaborators will use in-silico methodologies leveraging the existing Tidepool Loop simulator to extend, optimize and validate the Tidepool Loop algorithm to enhance postprandial glucose control in response to unannounced meals. Tidepool will validate enhancements in a prospective clinical trial intended to support FDA clearance in the US and CE marking in the EU, and commercialization of Tidepool Loop v3.0, labeled for use with no meal announcements.
This project will be conducted by Tidepool, a Palo Alto, CA based non-profit technology company building software for people with diabetes in collaboration with Stanford University, University of California, Santa Barbara, University of Pavia, Italy, Oregon Health & Science University (OHSU), and Replica Health, a company focused on developing AI and ML technology to support personalized metabolic health care.
References:
Aiello, E. M. et al. Review of automated insulin delivery systems for individuals with type 1 diabetes: tailored solutions for subpopulations. Curr. Opin. Biomed. Eng. 19, 100312 (2021).
Messer LH, et al. Real world hybrid closed-loop discontinuation: Predictors and perceptions of youth discontinuing the 670G system in the first 6 months. Pediatr Diabetes. 2020 Mar;21(2):319-327. doi: 10.1111/pedi.12971. Epub 2020 Jan 3. PMID: 31885123; PMCID: PMC7204392.
Mosquera-Lopez, C., Wilson, L.M., El Youssef, J. et al. Enabling fully automated insulin delivery through meal detection and size estimation using Artificial Intelligence. npj Digit. Med. 6, 39 (2023).
Samadi, S. et al. Meal detection and carbohydrate estimation using continuous glucose sensor data. IEEE J. Biomed. Health Inform. 21, 619–627 (2017).
Wilson, L. M., Jacobs, P. G., Riddell, M. C., Zaharieva, D. P. & Castle, J. R. Opportunities and challenges in closed-loop systems in type 1 diabetes. Lancet Diabetes Endocrinol. 10, 6–8 (2022).
Gingras, V., et al. (2018). The challenges of achieving postprandial glucose control using closed‐loop systems in patients with type 1 diabetes. Diabetes, Obesity and Metabolism, 20(2), 245-256.
Anticipated Outcome
Stanford will lead the clinical protocol development and clinical trial. The clinical trial will be a multicenter, prospective interventional clinical trial conducted with institutional review board (IRB) oversight and under investigational device exemption (IDE) from FDA.
Using the modified Tidepool Loop algorithm optimized for no meal announcement, participants will eat without meal announcement and will separately record approximate meal intake in a real-world, at-home pivotal clinical study of the Tidepool Loop v3.0 system, designed to support regulatory submission for FDA clearance and CE marking.
In keeping with Tidepool’s mission to create advanced tools for the diabetes community, the algorithms developed will be made readily usable by people living with diabetes via the open-source Tidepool Loop project and all discoveries, inventions, code, designs, and clinical materials will be made publicly available and, as applicable, assigned a highly permissive open source license such as BSD2 or MIT.
Exploratory Outcomes
Outside the scope of this grant proposal, subsequent to the clinical trial at the end of year 2, Tidepool and its collaborators will use data collected during the course of the trial to evaluate personalized, context-aware meal detection and estimation technology to dynamically adjust insulin dosing in response to meals based on additional factors, such as location, user activity, time of day, and day-of-week. Based on the new data collected from our clinical trials, we will first perform data analysis to evaluate how various factors, such as location, time of day, and physical activity (e.g., steps), affect glucose changes during and after meals without meal announcements. Next, we will incorporate these factors by designing a statistical model to provide a probabilistic prior to help dynamically refine the optimized controller parameters in Tidepool Loop 3.0, and thus achieve a dynamic and personalized glycemic control without meal announcements.
As a result of the clinical study, extensive data on eating behaviors will become available. By leveraging these data, we will be able to identify a meal model that not only captures the patterns observed in the data, but also provides a tool that mimics a personalized sequence of meal events consistent with the eating habits [1]. The meal model will also play a role in the design of meal detection algorithms. It can provide the probability of meal time and the size of the meal before the information can be derived from the measured glucose level, providing a prior probability of a meal event, which is useful to develop a stochastic meal detector. The availability of a realistic model of the prior probability is expected to reduce the uncertainty inherent in glucose data-driven meal detection, leading to a more advanced control algorithm that provides a dynamic and personalized approach to insulin control to help individuals with T1D achieve better glycemic control before, during and after meals.
In addition to a focus on meal-time enhancements specifically, Tidepool and its collaborators may use this data to explore additional modifications to further enhance the glucose prediction capabilities of the Loop algorithm for future versions of the system. This exploration would involve evaluating ML-based blood glucose prediction methods within the optimized simulator. Possible alternative approaches include replacing glucose momentum and retrospective correction terms in Loop’s model with a single Machine Learning (ML) based approach or alternatively evaluating a novel ML based blood glucose prediction model.
[1] Aiello, E. M., Toffanin, C., Magni, L., & De Nicolao, G. (2022). Model-based identification of eating behavioral patterns in populations with type 1 diabetes. Control Engineering Practice, 123, 105128
Relevance to T1D
The primary objective of this work is to substantially reduce the cognitive and therapeutic burden of type 1 diabetes management. By refining short-term adaptation parameters, the Tidepool Loop v3.0 algorithm is intended to be a closed-loop system capable of mitigating postprandial excursions without meal announcements. This project intends to demonstrate that users can achieve consensus Time in Range goals despite unannounced or late meal boluses.
Furthermore, leveraging data from the T1-DEXI study, this project will evaluate "Activity Presets"— a novel control strategy designed to simplify exercise management. Exercise and activity are associated with many benefits for individuals with type 1 diabetes including decrease in the incidence/severity of diabetes-related complications. Unfortunately, these activities, and aerobic exercise in particular, also increase the risk of hypoglycemia. (Riddell et al, 2017) Activity-specific pre-configured presets represent a fundamentally new, unique, and powerful strategy to help inform physical activity guidelines and optimize glycemia, carb intake and intensity/duration of physical activity during exercise.
Current automated insulin dosing (AID) systems typically manage this via temporary suspensions or elevated glucose targets. In contrast, Tidepool Loop Activity Presets utilize a fundamentally distinct mechanism: the simultaneous modulation of basal rates, insulin sensitivity factors (ISF), and carbohydrate-to-insulin ratios (CIR). This approach adjusts the underlying insulin kinetics rather than merely shifting the glycemic target, allowing for granular control tailored to specific modalities (e.g., walking, biking, jogging, strength training). Through our planned clinical trial extension/exercise phase, structured exercise sessions will be conducted to compare the performance of these parameter-based adjustments against traditional target-based modifications (Tidepool Loop v1.0).
Together, these innovations will lead up to a fundamental change in the management of the meals and physical activity for individuals with T1D and aim to decouple glycemic control from rigorous user input. Our research plan operationalizes this objective by measuring TIR alongside patient-reported outcomes regarding burden reduction. By eliminating the requirement for announcing meals, Tidepool Loop v3.0 has the potential to reduce health disparities for pediatric and adult populations facing barriers related to numeracy, health literacy, or caregiving support. Demonstrating safe, effective automated insulin delivery without meal announcements will inform future algorithm development, regulatory pathways, and deployment strategies, with the potential to improve equity, usability, and long-term outcomes for people living with type 1 diabetes. We believe supporting this work will play a pivotal role in advancing the next generation of automated insulin delivery.
Finally, Tidepool amplifies the impact of this work through a commitment to open science. In keeping with Tidepool’s mission to create advanced tools for the diabetes community, the algorithms developed will be made available via the open-source Tidepool Loop project for industry, academic researchers, citizen scientists, and people with diabetes alike, and all discoveries, inventions, code, designs, and clinical materials will be made publicly available and, as applicable, assigned a highly permissive open source license such as BSD2 or MIT.
(Please keep the remainder of this section confidential) Tidepool currently has two signed deployment partners for Tidepool Loop, Sequel Med Tech and [redacted], and is in active discussion with two additional insulin pump manufacturers. Breakthrough T1D’s funding will be augmented by Tidepool’s deployment partners. Breakthrough T1D’s funding will accelerate the clinical trial and ensure that intellectual property developed is owned by Tidepool, rather than our commercial partners. Tidepool will make the intellectual property openly available to be usable by any individual or entity in the diabetes space.