Revolutionizing Diabetes Management with AI-Powered Predictive Analytics for a Diabetes MedTech Company

11 min readFeb 5, 2024

TL;DR: This is Part 2 of the TISE Trans-disciplinary field project (See Part 1 here) and focuses on the solution design of AI-powered predictive analytics for diabetes management for an Austrian MedTech company. Key highlights include:

Context, Opportunities, and Challenges
Business Impact: Quantifies the potential cost savings and revenue generation through the adoption of the AI/ ML solution
Solution Design and Proof of Concept: Describes the process of developing and validating predictive analytics models using cloud-native services, demonstrating the feasibility and effectiveness of the solution.

In summary, this blog serves as a comprehensive case study, illustrating the transformative impact of AI-powered predictive analytics on patient care and the MedTech industry as a whole. While business-specific data are abstracted away to maintain confidentiality, the insights shared aim to showcase the potential of AI in revolutionizing healthcare delivery and driving measurable growth.

Authors/ Consultants: Issaliyeva Kamilya, Miozzo Chiara, Pham Ngan, Ruiz Diana, Schranz Franziska.

Supervisors: Professor Emeritus Roland W Scholz (ETH Zurich, DUK) , Dr. Liliya Satalkina, Christiane Ulbrich (DUK)

Company Overview

The chosen MedTech company specializes in developing innovative solutions for diabetes management. Their flagship product is an automated insulin delivery system designed to improve the quality of life for individuals with Type 1 diabetes. This system integrates an on-body insulin pump, a controller app, and a continuous glucose monitor (CGM) to automate insulin delivery based on real-time glucose level readings.

The system consists of three main components:

On-Body Insulin Pump: The insulin pump is a wearable device that delivers insulin subcutaneously, mimicking the function of a healthy pancreas.
Continuous Glucose Monitor (CGM): CGM is a sensor-based device that continuously measures glucose levels in the interstitial fluid, providing users with real-time data on their blood sugar levels.
Controller App: A controller app allows users to view their glucose level readings, track trends over time, set personalized alerts, and adjust insulin delivery settings remotely.

The case for Big data and AI

Working with the company, we analyzed and defined opportunities for them to leverage Big Data and AI to further enhance the capabilities of their automated insulin delivery system and improve diabetes care outcomes. In particular:

Predictive Analytics: Employ predictive analytics models to forecast future glucose trends and anticipate potential fluctuations, allowing users to take proactive measures to prevent adverse events. By analyzing historical data and identifying patterns, the system can provide users with early warnings and actionable insights to prevent complications.
Remote Monitoring and Support: Implement remote monitoring capabilities to enable healthcare practitioners to remotely access and review patient data in real time. AI-driven algorithms can flag abnormal glucose patterns or medication adherence issues, allowing healthcare providers to intervene promptly and adjust treatment plans as needed.
Personalize Treatment Plans: Utilize AI algorithms to analyze individual patient data and develop personalized treatment plans tailored to each user’s unique needs, preferences, and health goals. This personalized approach ensures optimal glycemic control and reduces the risk of hypo- or hyperglycemia.
Continuous Improvement: Leverage Big Data analytics to continuously refine and improve the performance of the automated insulin delivery system, incorporating feedback from patients and healthcare practitioners to optimize treatment algorithms and user experience.

Business Impacts

Revenue Generation:

Insulin Pump Sales: Assuming the company captures 20% of the Austrian market initially, serving approximately 114,600 to 129,000 individuals, and achieves a conservative 10% increase in device adoption due to Big Data and AI enhancements, the company could generate significant revenue from insulin pump sales. With an average selling price of €2500 per annual subscription, the company could potentially generate additional revenue of €28.65 million to €32.25 million from the Austrian market alone.
CGM Sales: Similarly, if the company sells its CGM within the range of €900 to €1500, with an average selling price of €1200, and achieves a 10% increase in device adoption, the company could generate additional revenue of €120 per device sold. Assuming an average patient goes through 3 CGMs per year, that means the company could generate additional revenue of €41 million to €68.4 million from CGM sales.

Taking these together, the total estimated added revenue from the Austrian Market alone, the company could earn an additional €70 to €100 million per year.

Market Expansion:

Expanded Market Share: By establishing its market leader position in Austria, the company could potentially double its share in insulin pumps and the CGM market in Austria.
Overseas Expansion: With a successful track record in the Austrian market, the company could replicate its success and capture a significant share of the European market for diabetes management solutions.

Cost Savings:

Reduced Complications: By improving glycemic control and reducing the risk of diabetes-related complications, the system can lead to substantial cost savings in terms of avoided hospitalizations, emergency room visits, and medication expenses. With a conservative estimate of €1000 per avoided complication, the potential cost savings for individuals using the company products could range from €11.46 million to €12.9 million annually in Austria alone.

Clinical Outcomes:

Improved Glycemic Control: By providing patients and healthcare practitioners with timely access to glucose level data and actionable insights, the data-driven diabetes management system can lead to improved glycemic control among Type 1 diabetes patients. Studies have shown that improved glycemic control can reduce the risk of long-term complications and improve overall health outcomes for patients.
Enhanced Patient Satisfaction: Patients with Type 1 diabetes value convenience and accessibility when it comes to managing their condition. By offering near-real-time glucose level reports and actionable insights, the data-driven diabetes management system can enhance patient satisfaction and engagement with their treatment plans.

After convincing the company of comprehensive and measurable needs and impacts, our team started working on building out a custom solution for them, adopting Agile and a Lean management approach.

Solution Design

The solution design for the company’s data-driven diabetes management system encompasses the following components:

Data Collection and Integration: Integrate data streams from Continuous Glucose Monitoring (CGM) sensors, insulin pumps, and controller apps into a centralized repository. This ensures seamless data collection and interoperability across devices, facilitating a comprehensive view of patient’s glucose levels and insulin supply.

Real-time Monitoring: Employ advanced algorithms to generate reports in a standardized format compliant with industry regulations. These reports provide daily patterns and weekly summaries of key metrics like average glucose, glucose variability, and time-in-range. Customizable time frames empower clinicians to tailor monitoring to individual patient needs.

Data Storage and Management: Implement a robust data storage and management infrastructure capable of securely handling large volumes of patient data while adhering to strict data privacy regulations such as GDPR. This ensures data integrity, confidentiality, and accessibility for authorized personnel.

Predictive Analytics: Develop AI-driven predictive analytics algorithms to extract actionable insights from complex datasets. Leveraging ML techniques, these algorithms provide personalized recommendations for insulin dosage adjustments and lifestyle modifications based on individual patient data and historical trends.

User Interface and Experience:

Controller App: Offer a user-friendly interface within the controller app for patients to access their glucose level readings, track trends over time, and receive personalized recommendations. This empowers users to actively engage in their diabetes management and make informed decisions about their health.
Healthcare Practitioner Portal: Provide a web-based portal for healthcare practitioners to access their patients’ glucose level data, review trends and insights, and communicate with patients remotely. Integration with Electronic Health Records (EHR) enables seamless access to patient information, facilitating informed decision-making and streamlined patient management.

Solution Building Process

During the design and build phase, the collaboration involves the following steps:

Requirements Gathering: Conduct collaborative workshops and stakeholder interviews to define the functional requirements, user workflows, and success criteria for the diabetes management system.
Solution Architecture: Develop a comprehensive solution architecture that outlines the technical components, data flow, and integration points required to implement the diabetes management platform.
Prototype Development: Build a prototype of the data-driven diabetes management and proof-of-concept ML algorithms with simulated data, focusing on key features such as data analysis and predictive analytics.
Proof of Concept Testing: Conduct testing and validation of the ML algorithms, involving review with simulated data and medical experts to assess clinical efficacy and reliability.
Iterative Development and Rollout: Iterate on the prototype based on feedback from stakeholders and validation results, the company would incorporate refinements to develop a production-ready system in a phased approach, starting with a pilot deployment with core monitoring capabilities and scaling up to further enhancements over time.

System Architecture

The AI-powered diabetes management system comprises several interconnected components designed to collect, analyze, and visualize glucose level data and provide actionable insights to healthcare practitioners and patients.

The system architecture includes:

Data Acquisition Layer:

This layer handles data ingestion from insulin pumps and Continuous Glucose Monitoring (CGM) devices. These devices continuously collect glucose level readings and transmit the data to a cloud-based data storage platform. The goal is seamless integration and interoperability with the rest of the system.

Data Processing and Analytics Layer:

In this layer, incoming glucose level data undergoes processing, aggregation, and analysis. Advanced analytics and machine learning algorithms are applied to identify patterns, trends, and anomalies. The system leverages Big Data technologies to generate actionable insights in near-real time.

Insights Generation and Visualization Layer:

A personalized dashboard and reporting interface are accessible to both healthcare practitioners and patients via the controller app. The dashboard presents:

Glucose Level Trends: Visualizations of how glucose levels change over time.
Medication Adherence Metrics: Insights into patient compliance with prescribed medications.
Personalized Recommendations: Tailored suggestions for managing glucose levels effectively.

Integration with Electronic Health Records (EHRs):

Seamless integration with existing EHR systems used by healthcare practitioners ensures the automatic transfer of patient data and insights. This streamlines workflow processes and maintains continuity of care across different healthcare settings.

Security and Compliance Measures:

Security protocols, encryption standards, and data privacy measures are woven throughout the system. These safeguards protect patient confidentiality and ensure compliance with regulatory requirements (such as GDPR and HIPAA). Regular audits maintain system integrity and security.

To enable accessibility across different devices, we designed a sample simple scalable cloud-native AI-powered predictive analytics system using Azure services:

System Design

Data Collection: Collect CGM data and physical activity data from individuals using Azure IoT Hub.
Data Preprocessing: Preprocess the data to handle missing values, outliers, and noise.
Model Training: Train the machine learning model using Azure Databricks and evaluate its performance using metrics like root mean square error (RMSE) and correlation coefficient (rho).
Real-time Analytics: Use Azure Stream Analytics to parse real-time stream data and transform it into actionable insights.
Storage: Use Azure Data Lake for long-term storage of data.
Reporting: Use Azure Synapse for dimensional models’ creation and reporting for non-real-time data, which can be enriched with other data from production.
Security and Monitoring: Use Azure AD, Azure Key Vault, Azure Logic Apps, Azure Functions, Azure DevOps, and Azure Monitor to ensure security, orchestration, and monitoring of Azure services utilization.
Model Deployment: Deploy the trained model to Azure Container Instances (ACI) or Azure Kubernetes Service (AKS) for production use.

Diagram: Sample cloud-native ML Predictive analytics system design using Azure services.

App Development

API Integration: Create an API using Azure API Management to expose the predictive analytics functionality to the controller app.
The app should be available on bothiOS and Android.

Building a Proof of Concept ML for Predictive Analytics

To test the efficacy of predictive analytics of the diabetes management system, we focused on developing and validating ML models for glucose prediction using continuous glucose monitoring (CGM) data.

Drawing inspiration from the successful research study “Machine learning-based glucose prediction with use of continuous glucose and physical activity monitoring data: The Maastricht Study” by William P. T. M. van Doorn (2020), we aimed to replicate and adapt their approach to our PoC. In particular, the study found Machine learning-based models, particularly recurrent neural networks (RNNs) and long short-term memory (LSTM), can accurately and safely predict glucose values at 15- and 60-minute intervals based on CGM data only.

Table: Baseline statistical and machine learning model comparison for predicting glucose values. (Foreman et al., 2021)

The steps for building the PoC are as follows:

1. Data Collection and Preparation:

Acquire a diverse dataset of CGM data from individuals with Type 1 diabetes from the company, capturing glucose readings at regular intervals along with corresponding physical activity data.
Clean and preprocess the dataset, handling missing values, and outliers, and ensuring data consistency and quality.

2. Feature engineering: Extract relevant features from the data, such as glucose levels, time intervals, physical activity, and meal times. Feature engineering is crucial for improving the accuracy of the predictive models.

3. Model Development:

Choose a suitable machine learning algorithm for predicting glucose levels. Based on the available literature, long short-term memory (LSTM) networks are chosen as it has shown promising results in predicting glucose levels for type 1 diabetes for both 15 and 60 mins intervals.

4. Model Training and Evaluation:

Split the dataset into training, validation, and test sets to train and evaluate the predictive models.
Train the models on historical CGM and activity data, optimizing hyperparameters and model architectures to maximize prediction accuracy.
Evaluate model performance using standard metrics such as root mean square error (RMSE) and correlation coefficient (rho), comparing predictions against actual glucose values at 15 and 60-minute intervals.

5. Clinical Safety Assessment:

Assess the clinical safety of the predictive models using error grid analysis, ensuring that model predictions fall within clinically acceptable ranges and do not pose risks to patient safety.
The error grid results indicated that model predictions were clinically safe (15 min: >99%, 60 min: >91%).

6. Stakeholder Feedback and Iteration:

Gather feedback from healthcare practitioners, and other stakeholders on the effectiveness of the predictive analytics features.

7. Looking forward — Getting Production Ready:

After stakeholders assess and approve the successful proof of concept (PoC), the company can build upon this PoC to deploy for real users. The plan would involve transitioning from the PoC phase to full-scale implementation and deployment of the ML within its diabetes management system, including:

Integrate the predictive analytics solution seamlessly with existing GlucoTech diabetes management systems, including the controller app, healthcare practitioner portal, and data storage infrastructure.
Ensure interoperability and compatibility with other healthcare IT systems and electronic health record (EHR) platforms.
Ensure compliance with regulatory standards and requirements, such as obtaining necessary approvals from relevant regulatory bodies (e.g., FDA in the United States, EMA in Europe).
Continuously monitor model performance, gather feedback from users, and iterate on the solution to improve effectiveness and address emerging challenges.

Reference

Foreman, Y. D., Schaper, N. C., M. Savelberg, H. C., Koster, A., Wesselius, A., Schram, M. T., A. Henry, R. M., Dagnelie, P. C., Bekers, O., A. Stehouwer, C. D., R. Meex, S. J., & J. Brouwers, C. G. (2021). Machine learning-based glucose prediction with use of continuous glucose and physical activity monitoring data: The Maastricht Study. PLoS ONE, 16(6). https://doi.org/10.1371/journal.pone.0253125