Submission

We propose a secondary data analysis project utilizing the Type 1 Diabetes EXercise Initiative (T1-DEXI) dataset to develop predictive models for hypoglycemic events in adults with type 1 diabetes. By employing a novel algorithmic framework called Trajectory Flow Matching on wearable device data that also accounts for conditions such as meal composition and insulin administration, we will create a digital twin platform that will integrate this information to predict glycemic responses to exercise. This approach will provide a novel wearable-based monitoring system that can predict and suggest personalized strategies to prevent hypoglycemic events.  

This project has the potential to impact the scientific understanding of how high-throughput wearable sensor data can predict clinically relevant events. This project also may impact human health by personalizing treatment plans, improving patient-physician communication, and enhancing management strategies using wearable devices.

Aims:

Develop a predictive model using TFM to simulate individual glycemic responses to exercise in T1D patients by integrating on dynamic physiological data captured by wearable devices, psychological questionnaire data, and other dietary, lifestyle, and medication-related variables.

Identify key factors (e.g., heart rate variability, stress levels, meal composition) that significantly influence hypoglycemic events post-exercise.

Enhance understanding of how psychological states interact with physiological parameters to affect glycemic control during exercise.

Study Hypotheses

H1: Incorporating psychological stress levels with other dietary, medication-related, and lifestyle variables into TFM models improves the prediction accuracy of hypoglycemic events in T1D patients during exercise.

H2: Dynamic measurements generated by wearable devices, when analyzed in conjunction with conditional variables like meals and insulin administration, can effectively predict hypoglycemic episodes.

Data Sources

We will utilize the T1-DEXI dataset (https://vivli.org/t1-dexi-analytics/), which is available through Vivli, a GREI-participating generalist repository. The dataset includes:

• Continuous glucose monitoring (CGM) data.
• Insulin administration records.
• Wearable device data (Polar H10 sensor and Verily Study Watch) capturing heart rate, steps, and activity levels.
• Self-reported stress levels and psychological assessments.
• Meal timing and composition data (Remote Food Photography Method).
• Demographic and clinical characteristics.

Methodology

Data Preparation

• Data Pre-Processing: Address noise and artifacts in wearable data through filtering techniques, conversion to other formats that are clinically meaningful (e.g. heart rate variability), and evaluation across different time-scales.
• Multimodal Data Fusion: Integrate dynamic wearable data along with conditional variables (meals, insulin administration) and psychological assessments.

Trajectory Flow Matching (TFM)

• Apply TFM to model the continuous trajectories of physiological parameters as measured by the wearable device.
• Condition TFM models on meal composition, lifestyle choices and insulin dosing to account for baseline variations.
• Incorporate psychological state as a modifying variable in the TFM framework.

Statistical Analysis

• Predictive Modeling: Use machine learning algorithms (e.g., random forests, gradient boosting) using features derived from TFM models to predict hypoglycemic events.
• Validation: Employ cross-validation techniques to assess model performance.
• Interpretation: Identify significant predictors and their interactions affecting glycemic responses.

Ethical Considerations

• Ensure de-identification and confidentiality of patient data.
• Utilize data in an ethical, non-discriminatory manner, adhering to all applicable guidelines.

In our research project we expect to identify a machine learning model that will predict the risk for future hypoglycemic events for patients with Type 1 Diabetes. We expect to establish a platform for digital twins integrating wearable device measurements with other variables that can be useful for predicting clinical outcomes. 

Findability: we will directly attribute the specific data used to train the predictive model to the appropriate variables defined in the T1-DEXI dataset hosted by Vivli.

Accessibility: we will make the meta-data accessible according to the pre-specified protocol by Vivli. 

Interoperability: We will design the platform to be able to integrate with other sources of multi-modal data, including other wearable devices. 

Re-Use: we will release the data used for the project according to community standards and the protocol for data re-use as defined by Vivli. 

Outputs and Dissemination Plan

Outputs

• Predictive Models: Algorithms capable of forecasting hypoglycemic events based on individual patient data.
• Digital Twin Framework: A prototype for creating personalized digital representations of patients.

Dissemination

• Publication: We will submit manuscripts to open-access journals and journals with the option of publishing open-access articles.
• Conference Presentations: We will present our findings at relevant scientific meetings across different disciplines (e.g. Neural Information Processing Systems, Conference for Computer Human Interaction, International Conference for Machine Learning, American Diabetes Association Scientific Session, ENDO) 

For replicability and reproducibility of our results, we will deposit code on a publicly available repository on GitHub with clear documentation and licensing for reuse. We will also make the code behind the platform available for other groups to use with their own computing resources. 

Impact and Scientific Significance

Contributions to Science

• This is the first unified algorithmic framework to use wearable data in creating a dynamic risk prediction for hypoglycemic episodes in patients with Type 1 Diabetes.
• This project provides the potential to better understand the potential of high-throughput wearable data in predicting glycemic control.
• This project creates a platform for other similar clinical issues that may benefit from integrating wearable device data into digital twins to guide personalized treatment plans. 

Impact on Health

• Improved Patient Outcomes: Potentially reduce hypoglycemic events through better prediction and management.
• Enhanced Patient Engagement: Utilize digital twins to facilitate patient understanding and adherence using wearable devices in their treatment plans.
• Support for Clinicians: Provide a platform for physicians to make data-driven suggestions for patients and improve patient-physician touchpoints.

We are an inter-institutional, interdisciplinary team with expertise in clinical informatics (DLS), generative artificial intelligence (YP, NZ, AT, SK), algorithmic development (AT), and wearable data analytics (GA). We (DLS, AT, YP, and NZ) collaborated together to develop a novel generative artificial intelligence framework for dynamic trajectory modeling for clinical data called Trajectory Flow Matching, which was accepted as a Spotlight Presentation (Top 3%) at Neural Information Processing Systems 2024. 

• Appropriate preprocessing procedures for the wearable data into a format that is computationally informative
•  Adaptations to the algorithmic framework to provide signal for clinical events
• Definition of clinically meaningful outcomes and endpoints that are computationally tractable