Establishment of a Diabetes-Tailored Data Intelligence Platform Enhances Clinical Care, Enables Risk-Based Monitoring, and Facilitates Population-Health-Based Approaches at a Pediatric Diabetes Network

Abstract

Background:

Patient-generated health data (PGHD) represents an opportunity to customize care, particularly in type 1 diabetes (T1D) care where continuous glucose monitor (CGM) and insulin pump usage continues to rise. Previous solutions to integrating CGM data into the electronic health record (EHR) have been limited in their ability to integrate data from multiple sources, ensure data fidelity, integrate data from multiple data streams, and rapidly adapt to changes in data output from numerous vendors. We developed a novel data infrastructure contained outside of the EHR to provide an alternative approach to PGHD integration, enable diabetes centers to identify and predict risk, and to facilitate research and quality improvement.

Methods:

We identified three key capabilities: ingesting and storing a wide variety of data, refining raw data into actionable insights, and visualizing and reporting to decision makers. To meet these requirements, we built a data intelligence platform we coined the diabetes data dock (D-data dock) in the Microsoft Azure cloud platform.

Results:

The D-data dock houses approximately 100 million CGM measurements, one million clinical events and insulin bolus records, and a near complete EHR record covering approximately 3000 patients per year from 2016 to 2023. We provide case studies detailing how the D-data dock allows timely monitoring of CGM data, enables novel study designs, and powers machine-learning-informed supplemental care interventions.

Conclusions:

The D-data dock is a novel approach to harnessing disparate data streams to improve patient care, enable timely interventions, and drive innovation to improve the lives and care of people with T1D.

Introduction

Significant cultural, economic, and legislative changes over the last two decades have led hospitals and health care providers to increase the adoption of electronic health record (EHR) systems.^1-4 Alongside the move within health care systems to EHR-based documentation, the development and rapid uptake of patient-worn digital health devices has led to a significant increase in patient-generated health data (PGHD).^4-6 While these parallel innovations have each generated large amounts of digital health data, these data streams remain poorly integrated.^7-10

Continuous glucose monitors (CGMs) worn by people living with diabetes are a particularly important source of data providing minute-to-minute data critical to management of this highly prevalent chronic disease. Continuous glucose monitors provide minimally invasive, near real-time glucose data, detect and help prevent hypoglycemia, monitor long-term glycemic trends, guide patient and provider insulin regiment adjustments, and are a key component of automated insulin delivery (AID) systems. ¹¹ Multiple studies have demonstrated that consistent CGM and AID use are each associated with improved glycemic and psychosocial outcomes in pediatric and adult patients with both type 1 diabetes (T1D) and type 2 diabetes (T2D).^12-18

Because of the central importance of glucose data in diabetes care, integrating CGM data in the EHR has been a focus for health care organizations and industry over the past decade (Figure 1). ⁹ ,^19-21 Direct EHR integration has numerous advantages, including reducing documentation burden, minimizing task switching for clinicians, and contextualizing CGM data within the rest of the clinical record.^22-24 The 2022 iCoDE Report: CGM-EHR Integration Standards and Recommendations summarizes these benefits and provides technical and operational guidance to facilitate the process. ²⁵ Despite this progress, CGM-EHR integration is not a panacea, and has its own technical, operational, and financial challenges and limitations. ²⁶ These include the following:

Figure 1.

A conceptual, architectural comparison of the D-data dock approach and the typical cutting edge, EHR-centric approach. The D-data dock acts as middleware for data integration and provides valuable additional capabilities.

• Data sourcing: Organizations may need multiple partners to source data from different devices and manufacturers.

• Data limitations: Most integration efforts focus on CGM summary metrics and PDF reports such as Ambulatory Glucose Profiles (AGP) because EHRs are not well suited to host or display the hundreds of daily data points generated by CGMs.

• Technology agility: The data, visualizations, and features available are dependent on the technical roadmaps of large technology vendors and may not be able to meet the unique and rapidly evolving demands of different care settings.

• Lack of support for population health: Some EHRs have limited capacity to support population health management, and most health systems use other platforms for these workflows.

These EHR and integration challenges make it difficult to benefit from innovations in machine learning and population-level approaches such as those that have helped adults with T2D lower their A1c and increased prescribing rates for complication screening.^27,28

To address these challenges and create a technical foundation capable of helping to innovate diabetes care, we built a data intelligence platform we coined the diabetes data dock (D-data dock). The D-data dock is contained outside of the EHR and makes it possible to draw data from disparate sources to facilitate individual and population-level insights, support research and quality improvement, power machine learning approaches for predictive modeling, and enable other promising population-level approaches.^29-31

Methods

Overview

In 2020, Children’s Mercy Kansas City launched the Rising T1DE Alliance to rapidly scale quality improvement efforts and innovation in T1D care. Doing so required predicting clinically important outcomes, curating and evaluating novel interventional strategies to help youth with diabetes achieve better health and quality of life, and creating a platform to drive rapid-cycle testing of novel interventions. ³² We identified three fundamental capabilities: (1) ingesting and storing a wide variety of health data, (2) refining raw data into actionable insights, and (3) visualizing and reporting data and insights to decision makers. To meet these requirements, we built the D-data dock in the Microsoft Azure cloud platform. A visual overview of the D-data dock architecture is shown in Figure 2.

Figure 2.

The technical architecture of the D-data dock. A variety of data sources (far left) are ingested and stored in an Azure Gen2 data lake. The computing environment Databricks houses the code to ingest, clean, and refine the data. Visualizations are available for clinic and support staff in Power BI. Azure Data Factory is used to schedule and coordinate code execution.

Data Ingestion and Storage

The first function of the D-data dock is to ingest and store a wide range and high volume of data in a central location. The primary categories of data included are EHR data, supplemental survey data, and PGHD. Data ingestion is a critical precursor to the subsequent refinement, analysis, and visualization capabilities. Data are stored in a Gen 2 Azure Data Lake, a Microsoft cloud storage solution. Data are ingested via Application Programming Interface (API) when possible. Otherwise, data are manually downloaded by support staff before being uploaded weekly to a dedicated ingestion storage account. Data are ingested following an Extract-Load-Transform pattern (ELT), a standard process for modern analytics as it is highly flexible and scalable. ³³ The ingestion step makes critical data elements from disparate sources—EHR data, PGHD, and supplemental survey data from Redcap—available in a single data platform.

Data Refinement

After ingestion, the first transformation for PGHD is to map data from the vendor’s data model to an internally created, common data model for all data of that type. We also deduplicate PGHD within and across vendor sources. It is common for our processes to ingest the same records from multiple sources; a given patient’s CGM data may be found in the source vendor’s data feed (eg, Dexcom) as well as an aggregator’s data feed (eg, Glooko). In the case of CGM data, for example, we de-duplicate by selecting the highest volume data source per person-day, then ensuring that measurements are not too frequent—five minutes expected between measurements in most cases. Electronic health record data are mapped to the T1D Exchange specification. ³⁴ The T1D Exchange is a US-based nonprofit dedicated to facilitating research and driving improvements in care and outcomes for people with T1D. Mapping data to this specification and submitting it to the T1D Exchange contributes to collaborative research, enables fuller participation in T1D Exchange initiatives, and could make the D-data dock easier for other sites to implement in the future. After the data have been ingested and mapped to a common model, we refine the data to confirm data quality and generate actionable insights. For example, we created automated processes to corroborate Diabetic Ketoacidosis (DKA) diagnosis codes with relevant lab results and validate diabetes type with a rules-based algorithm. ³⁵

We use Databricks, a cloud-based, data analytics computing environment, to facilitate both the ingestion and refinement processes. Databricks is an industry-standard, HIPAA-compliant platform that simplifies many infrastructure and setup tasks as well as collaboration among data scientists, engineers, and researchers.

Using Databricks has enabled the creation of machine learning models and other refinements that enable the risk-aware, population-health-based approaches of the Rising T1DE Alliance. The D-data dock has enabled the entire lifecycle—data preparation, model training, model validation, deployment, and prediction—of two diabetes-specific machine learning models predicting 90-day change in HbA1c and 180-day DKA risk.^36-38

Data Visualization and Report Generation

After ingestion, storage, and refinement, we display combined EHR and PGHD in a variety of ways to provide actionable insights to decision makers and care providers. Current visualizations include interactive dashboards built using Microsoft Power BI, a data visualization platform, in collaboration with care providers to allow for direct, interpretable data access for clinical care and research coordination. These dashboards fall in two broad categories: individual patient summaries and population health monitors. Dashboards focused on summarizing individual patient data are valuable because they address the persistent EHR shortcomings around streamlined, actionable data presentation. ³⁹ For example, we created a dashboard that visualizes a youth’s engagement with six self-management habits alongside A1c values and CGM metrics. ⁴⁰ This consolidated view of engagement in habits over time alongside key outcomes would otherwise be buried within individual encounter documentation and require significant searching.

Dashboards focused on population health and risk-based screening is the other major type of visualization available in the D-data dock. The Tide dashboard (case study #1 in Results) presents weekly risk categorizations of recent CGM data to allow Certified Diabetes Care and Education Specialists (CDCES) to actively monitor and outreach to youth and their caregiver between their regularly scheduled visits. The DKA and rise in A1c predictions previously described are published in dashboards that quality improvement personnel use weekly to screen, identify, and contact families for clinical interventions to address predicted negative outcomes (case study #3). All dashboards are accessed via a standard web browser either directly through Power BI or via Links embedded in the EHR.

Results

The D-data dock ingests and de-duplicates more than 10 million CGM records per week for approximately 1900 youth across five data sources: Clarity, Glooko, Libreview, TConnect, and Carelink. The D-data dock contains 640 million de-duplicated CGM measurements since the start of 2020 and an additional 120 million records spanning 2014 to 2019. Data volumes by year and type are available in Figure 3.

Figure 3.

This shows the volume of data in the D-data dock by type and year. Approximately 10 000 clinic visits, one million clinical events and insulin boluses, and 100 million CGM measurements.

The following case studies highlight the ability of the D-data dock to support research and care innovation efforts through increased accessibility and usability of PGHD, EHR data, and derived insights for clinicians and researchers.

Case 1: Project TIDE

In 2021, a partner organization developed an open-source algorithm-enabled care model called Timely Interventions for Diabetes Excellence (TIDE).^41,42 A key component of this care model is a patient-data dashboard (the TIDE dashboard), which helps clinic staff identify youth at risk for clinical deterioration based on frequent review of CGM-based data trends outside of clinic visits. This dashboard is monitored by CDCES, who are presented with a prioritized list of youth identified as being at risk for complication based on the previous weeks’ CGM data using CGM-data thresholds, for example, >2% time below 54 mg/dL (3.0 mmol/L), <65% time in range (70-180 mg/dL, 3.9-10.0 mmol/L), and so on. Pilot testing of this intervention at its home institution showed promising reductions in complications in an initial cohort. Our group was asked to validate these findings. ⁴³

The initial development of the TIDE platform—its underlying infrastructure as well as clinical integration and dashboard development—spanned years and included contributions from data scientists, clinicians, and researchers. When our group was approached to implement the TIDE dashboard at our institution, the D-data dock infrastructure allowed our team to rapidly adopt and implement the original dashboard as well as adapt the original design to meet the needs of our clinical team. We launched the D-data dock-based version of the TIDE dashboard (Figure 4) with only two weeks of partial effort from a single data scientist.

Figure 4.

The three main screens of the data dock implementation of the TIDE dashboard allow a user to (a) select the population of interest based on a variety of filters from the CGM data and the EHR, (b) view summary statistics of all members and screen across a variety of risk categories, (c) and see additional details for identified individuals.

The TIDE dashboard implementation demonstrates the advantage of ingesting and working with raw CGM data rather than collecting only summary statistics, as is typically the approach in EHR integrations. Using those raw measurements, the D-data dock infrastructure provides data needed to replicate all AGP visualizations and calculate metrics such as the Glycemia Risk Index (GRI). ⁴⁴

Case 2: Coin2Dose

Coin2Dose is a multicenter pilot randomized controlled trial which seeks to examine the impact of weekly financial incentives on meal-time bolus behavior of youth 11 to 17 years old with T1D. ⁴⁵ Those in the financial incentive arm receive a nominal weekly financial incentive which is paid based on the frequency of pre-meal insulin doses as calculated by the BOLUS score. ⁴⁶ This nation-wide study recruited 173 youth-parent dyads to evaluate the effect of financial incentives on insulin dosing behaviors among adolescents with T1D. All participants use an insulin pump for their T1D management, and bolus data are passed to the D-data dock via Glooko API or Carelink file upload. To automate study data collection and refinement, we were able to use existing D-data dock functionality to handle the data ingestion and mapping processes for Glooko and Carelink data as well as the BOLUS score calculation. ⁴⁶

With the existing infrastructure of the D-data dock, the only custom work required to enable study initiation for Coin2Dose was a weekly BOLUS score dashboard accessible to study team members at each recruitment site. This end-to-end, study-specific workflow was deployed in a matter of weeks and enabled smooth recruitment and timely compensation at the two initial sites and enabled study staff to later widen recruitment to the T1D Exchange research network to meet study recruitment goals.

Case 3: Machine Learning and Remote Patient Monitoring (RPM)

Finally, the D-data dock has enabled the development, training, validation, and deployment of two machine learning models within our clinic—one predicting 90-day change in A1c and another predicting 180-day risk of DKA. ³¹ These models use EHR data ingested and stored by the D-data dock infrastructure as a basis for learning and prediction. The models are parsimonious, containing ~16 features each, and interpretable as each feature is either from or easily derived from the EHR. Model predictions are made each week by the Databricks-based refinement layer and made available to a team of quality improvement coordinators. These coordinators then contact at-risk youth who are predicted to experience either a 90-day A1c rise ≥0.3% or are ranked in the top 20 most at-risk youth for experiencing DKA.

Coordinators invite these youth and caregivers to participate in supplemental care interventions based on their unique needs. These interventions include remote patient monitoring (RPM), mHealth programs, peer support sessions, diabetes education, nutrition studies, and physical activity studies. During the previous four years, these interventions have enrolled more than 800 participants, and risk predictions from our ML models have provided the single largest referral base to these interventions. A pilot group in the RPM intervention showed a trend toward lower 90-day HbA1c increases than a propensity-matched cohort. ⁴⁷ Crucially, the D-Data dock facilitated the creation, implementation, and rapid testing of this intervention and many others like it.

Discussion

Previous solutions to integrating CGM data into the EHR have been limited in their ability to integrate data from multiple sources, ensure data fidelity, integrate data from multiple data streams (eg, insulin pumps and CGMs), and rapidly adapt to changes in data output from numerous vendors. ⁹ With a wide range of external data input streams alongside EHR data, the D-data dock system provides a model for incorporating new data streams and insights alongside existing clinical systems.

As demonstrated in our implementation of the TIDE dashboard, the availability of these data streams outside of the EHR allowed for the implementation of a novel care model with minimal software development effort. The Coin2Dose case study demonstrates the D-data dock’s ability to use incoming data from any data platform and rapidly develop innovative and highly targeted capabilities to support study procedures that would have been difficult or unfeasible otherwise. Finally, the third case study demonstrated the ability of D-data dock enabled machine learning algorithms to prospectively identify high-risk individuals before clinical deterioration, set in motion significant improvements in outcomes for youth, and ensure the efficient use of the limited time of expert clinicians to those most in need. These novel analyses and visualizations allow clinicians to complement individual patient care with risk-informed, population-focused processes that provide timely, supplemental care and support to those at the highest risk for clinical deterioration. Health systems often struggle to adopt and scale health analytics and machine learning efforts because of barriers such as clinical workflow integration challenges and uncertain inter-site generalizability.^48,49 These case studies demonstrate that the D-data dock addresses many foundational challenges of integrating PGHD and EHR data and creates new care and research opportunities.

These challenges and their solutions are not unique to our institution or even to diabetes care—the fundamental approaches to data integration operationalized by the D-data dock provide a blueprint for dissemination of this approach. Those implementing this approach can quickly benefit from the existing processes that ingest, refine, and report PGHD and EHR data and develop interventions and supplemental care processes that suit their needs. To facilitate dissemination, the predictive models are based on the widely adopted T1D Exchange data specification as inputs. As of April 2025, 43 diabetes centers had mapped their data to this specification. In addition, Databricks can be deployed and configured to recreate D-data dock specifications in any of the three major cloud environments: Microsoft Azure, Amazon Web Services, and Google Cloud. Once implemented at other sites, the D-data dock can streamline the process of sharing and disseminating innovations across systems and provide a common environment to allow for collaborative development.

Future development of the D-data dock will include integrating both the data and visuals into the EHR. Such an integration will increase the benefit the D-data dock provides to the clinical workflow by decreasing documentation burden and reducing the need to switch between the EHR and another system.

Because of the early successes and impact of the D-data dock, our future focus is disseminating the platform beyond its current home within a single pediatric hospital system. There are ongoing, current efforts to partner with both pediatric and adult-focused institutions as well as state-level data exchanges to implement the D-data dock. While the underlying data storage and analysis architectures are widely used both in health systems and in broader applications within the health care industry, implementation of the D-data dock architecture and systems requires support and effort from a local technical team to implement and maintain the D-data dock system within clinical and research operations of an organization. This requires both navigation of regulatory barriers and financial investment from institutions to support the ongoing implementation of these systems to support clinical care.

There are three important factors that will contribute to wider adoption of the D-data dock beyond the currently planned efforts:

1. Additional evidence that interventions powered by the platform, such as RPM and TIDE, positively affect clinical outcomes.

2. Hospital willingness and ability to bill for remote data collection and monitoring or value-based care programs that fund the approach.

3. Institutional vision for the long-term research and quality improvement benefit of the comprehensive data set the D-data dock collects.

We also see an opportunity to adapt the D-data dock to conditions beyond its current focus of youth with T1D who are cared for by endocrinologists. Ensuring that the D-data dock meets the needs of adult care and the primary care setting is critical given that primary care provides the majority of outpatient care for persons with diabetes in the United States. ⁵⁰ We believe the framework and philosophy provided by the D-data dock is a critical starting point for improving outcomes across a wide range of chronic diseases.

The D-data dock is a novel approach to harnessing disparate data streams to improve patient care, enable timely interventions, and drive innovation to improve the lives and care of people with T1D. About digital transformation in diabetes care, Kompala and Neinstein wrote, “Each person with diabetes will have a comprehensive, longitudinal diabetes care record that pulls together all of an individual’s relevant data from the electronic health record, home device data, and more.” ⁵¹ The D-data dock can fulfill that prediction and catalyze care transformation.