Tidepool Project
There were no major team changes from what list of contributors outlined in phase 1.
Diabetes technologies include basic exercise features but these approaches lack the fine-grained control to effectively manage physical activity. Tidepool partnered with Stanford, UCSB, and U Pavia to develop activity-specific presets for walking, biking, jogging, and strength training. Using the T1-DEXI dataset, we derived optimal activity presets and developed a Physical Activity Metabolism Model, which was used to validate the presets via an exhaustive simulation. The presets are integrated into Loop 2.0 with education materials. The analysis showed that a mathematical model can predict the impact of exercise on glucose due to both insulin and non-insulin mediated effects, and that derived presets consistently improved Time in Range and Time Below Range. This work provides the first systematic, data-driven quantification of optimal insulin-reduction strategies across common activities and demonstrates the value of large-scale in silico evaluation to inform diabetes technology.
The T1-DEXI dataset (DOI: 10.25934/PR00008428) accessed through GREI/Vivli was used to identify activity presets that improve glucose levels during physical activity by reducing insulin delivery before and during exercise. Presets adjust the insulin sensitivity factor, basal rate, and carb-ratio, which evidence suggests mimic metabolic changes in insulin needs from physical activity. The T1-DEXI dataset included data from 512 adults with T1D over a 4-week observation period with CGM data, insulin dosing records, heart rate measurements, self-reported activity logs and contextual data (i.e. meals). Using a data-driven approach we found optimal presets for four of the most common physical activities: walking, jogging, biking, and strength training.
Each exercise session in the T1-DEXI data set was replayed with different preset values to determine the reduction in insulin delivery that minimized the Magni risk score (a glucose measure that balances the risk of hypoglycemia and hyperglycemia). Once the optimal preset was found for each exercise session, a final preset value was determined by taking the median optimal preset per activity type, yielding presets for walking, biking, jogging, and strength training. Preset values align with the known physiology for these types of activities.
Glucose uptake during exercise is modulated by insulin dependent and insulin independent effects. We created a new model in silico for these effects to further validate the data-derived presets in simulation. To model these effects, we first used exercise sessions with netIOB < 0 IU to fit a first order model, Gii, that estimated the relationship between heart rate and insulin independent glucose uptake by skeletal muscles. This model for insulin-independent glucose uptake was then added to the model for insulin-dependent glucose uptake, Gid as a weighted sum: Wii * Gii + Wid * Gid
These weights were learned using the T1-DEXI data with netIOB > 0 IU.
The identified Physical Activity (PA) Metabolism model was integrated into the Tidepool Simulator alongside existing carbohydrate and insulin models. Four activity types were simulated for a cohort of 72 adult virtual patients with varying ISFs, carb-to-insulin ratios (CIRs) and basal rate profiles. For each virtual patient, we simulated 30 and 60-minute sessions for all four activities. The T1-DEXI sessions used in identifying the model had a median netIOB of 2.05 U one hour before the activity. To align with this condition, we administered a bolus dose of 2U one hour before each simulated activity. We then compared the drop in glucose (ΔG) resulting from the simulated activities to that observed in the T1-DEXI dataset. To account for the varying activity durations in the T1-DEXI dataset, we normalized the ΔG for each session by its duration in minutes and then scaled it to 30 or 60 minutes to allow direct comparison with the corresponding simulation durations. Validation of the Physical Activity Metabolism Model is covered in greater length in Diabetes: Identifying Insulin and Non–Insulin-Mediated Mechanisms during Physical Activity from Real-World T1D Data, listed here and in Section 3.
We performed a large-scale grid search (960k) on 10 Tidepool virtual patients, varying patient parameters and introducing noise in the range of ±0-25% to the controller settings to simulate controller-metabolism mismatch. We calculated safety metrics (percent time in/above/below range) and risk metrics (LBGI, HBGI, BGRI, Magni) over various time intervals, starting one hour pre-activity, at activity start, activity end, and 1, 2, or 3 hours post-activity. To better reflect real-world conditions, simulations were weighted based on the log-normal distribution of starting glucose values in the T1-DEXI data. We observed that presets that temporarily reduced overall insulin needs plus a raised target (150 - 170 mg/dL) consistently improved %TIR by 10 to 22% and reduced Time Below Range by 12 to 19%, depending on activity type, as compared to taking no action during exercise. We also observed activity presets plus a raised target consistently improved %TIR by 6 - 7% and reduced Time Below Range by 5 to 8%, as compared to raised target only, the standard functionality available in commercial systems.
Conclusions
Tidepool and its collaborators have developed and validated novel, activity-specific presets that are intended to reduce hypoglycemia risk during exercise for people with type 1 diabetes. Our work represents the first systematic, data-driven derivation of presets for common activities, validated through large-scale simulation. This work has been integrated into the Tidepool Loop V2.0, representing a direct translation of scientific advancement into a powerful tool for the T1D community.
The central outcome of this project is that we advanced the long-standing clinical challenge of safe, confident exercise in type 1 diabetes from abstract guidance to an actionable and comprehensive patient-facing solution. We expect activity presets will be available in the twiistTM automated insulin dosing system powered by Tidepool in 2026.
By design, the outputs bridge research and practice:
Research Methods and Metadata Considerations
For each session, events were harmonized to a common schema: activity type, reported intensity, duration, time-stamped start and stop, meal/bolus context, insulin on board, heart-rate traces (minute cadence), and CGM values (minute cadence, mg/dL). Simulation datasets were logged with structured inputs/outputs and stored in columnar format with dictionaries for re-use. Derived variables were mapped to CDISC SDTM/ADaM, enabling re-sharing through GREI/Vivli.
Standards, Resources, and Tools
We aligned the project with FAIR (Findable, Accessible, Interoperable, Reusable) and CARE (Collective Benefit, Authority to Control, Responsibility, Ethics) principles and implemented CDISC harmonization to make re-share straightforward. In practice, this translated to the following activities:
Replicability and Reproducibility
Reproducibility was treated as a requirement per Tidepool engineering standards:
Conclusion
The outcomes of this research redefine how diabetes technology supports physical activity. Tidepool Loop v2.0 now offers evidence-based starting points tailored to activity type, in-app guardrails that reflect type 1 physiology, and training to support patients to adapt presets to individual needs. For the broader community, we are contributing a validated, open-source framework of methods and models that will accelerate the next wave of innovation, including pediatric applications and fully dynamic, adaptive activity presets that support personalization.
Establishing a New Paradigm for Exercise in Automated Insulin Dosing Systems
From our analysis of the T1-DEXI dataset to the design of four validated activity presets, this research project has produced contributions across multiple disciplines. We have advanced our scientific understanding of exercise physiology, developed validated large-scale, data-driven research methods for algorithm advancement and established a new precedent for real-world exercise tools to make physical activity safer and easier for people with type 1 diabetes. Beyond diabetes, this work advances the field of data science in healthcare by demonstrating how large global datasets can be transformed into clinically useful tools through modeling and large-scale simulation. The outcome is a reproducible framework that others can adapt to advance innovations for the T1D community or other chronic conditions. This project also has real-world applications. In September, Tidepool submitted a Pre-Sub to the FDA in advance of a 510(k) submission for Tidepool Loop V2.0, supporting the goal of releasing activity presets in 2026 as part of the twiistTM AID system.
In the short-term, activity presets prevent acute complications by lowering the likelihood of severe hypoglycemia during or after exercise. In the long-term, presets make participation in physical activity safer and easier, reducing cardiovascular risk, improving insulin sensitivity, lowering blood pressure, and enhancing mental health.
This project advances the community’s understanding of exercise physiology in type 1 diabetes, validates new modeling approaches, and demonstrates the value of large-scale in silico evaluation. Clinically, it delivers differentiated, validated presets that reduce hypoglycemia, improve Time in Range, and empower patients to exercise safely. Preventively, it makes sustained physical activity more achievable long-term, improving health outcomes. For the broader health ecosystem, it delivers on a new model for how open source data collection and secondary data analysis can rapidly produce impactful interventions. This project sets a new standard of care for supporting physical activity for people with type 1 diabetes.
The project was completed within the award period. There were minor revisions to scope from Phase 1. For the validation approach, we proposed a performance analysis of the Tidepool Loop 1.0 and 2.0 features in a clinical trial featuring a structured randomized cross-over exercise intervention. Instead, we plan to use an in silico validation strategy to fast track clearance. We also proposed the design of dynamic activity-specific presets, but to date, we focused on optimizing the static presets and laying the groundwork for future adaptive approaches.
We successfully navigated data constraints to realize this research. The T1-DEXI dataset included adults with T1D, limiting pediatric inferences. In silico work is in progress to extend the use of activity presets to younger populations. There was expected subjectivity in self-reported logs. We addressed this by harmonizing activity categories into four major types and applying consistent thresholds for inclusion of data (no bolus insulin within 3 hours before and during exercise, no carbohydrate intake within 3 hours before exercise, based on self-reported meal intake and the starting blood glucose is between 70 and 200 mg/dL). Finally, contextual information (meal timing, sleep) was not always complete. To mitigate this, we simulated a series of controller-metabolism mismatch scenarios by introducing ±0-25% levels of noise to the controller’s parameter settings.
We applied multiple layers of validation to ensure quality and completeness. At the data ingestion stage, we performed quality control checks, including verification of CGM cadence and validation of heart rate signals against expected physiological ranges. Activity logs were cross-referenced with device time stamps, and sessions with inconsistent or incomplete timing data were excluded. We also standardized activities by mapping self-reported logs into four types, walking, biking, jogging, and strength training, and created rules for handling borderline cases (e.g., “hiking” is walking equivalent). At the analysis stage, we validated the data by comparing observed glucose trajectories with published literature. Finally, during our large-scale simulation, we explored a range of parameter adjustments to evaluate the impact of different control strategies (i.e. activity duration, starting glucose, etc) and a subset of simulations were re-run to confirm reproducibility.
T1-DEXI posed expected challenges: self-reported activity types and missing contextual fields. We addressed these challenges, building a data dictionary and harmonization schema aligned with CDISC SDTM/ADaM standards, collapsing free-text into four activity categories and applying conservative inclusion thresholds. On the analytic side, the primary challenge was balancing the complexity of exercise physiology with the need for generalizability. Glycemic responses to exercise vary by intensity, timing, iOB, and individual physiology, making it difficult to generalize from real-world data. To overcome this, we developed a feedforward metabolism model parameterized by T1-DEXI data to disentangle insulin-independent and insulin-dependent effects. Finally, to ensure usability, we focused on clinically meaningful safety and risk metrics to align results (i.e. Time Below/Above Range). Together, these steps turned noisy real-world data into reproducible insights and actionable clinical tools.
