Enabling Data-Driven Healthcare with Pervasive Technology - PowerPoint PPT Presentation

Enabling Data-Driven Healthcare with Pervasive Technology Temiloluwa Prioleau Assistant Professor of Computer Science Dartmouth College Webinar on Diabetes and Technology Management Center for Digital Health Interventions March 17, 2020

Background Pervasive technologies provide unparalleled opportunities for: 1. Continuous sensing of people in daily living 2. Real-time detection of health abnormalities 3. Data-informed decision and treatment strategies Location Activity Heart Rhythm Blood Glucose Rehabilitation T. Prioleau - Department of Computer Science 2

Diabetes & Technology Diabetes State-of-art technology • Prevalent chronic condition • Continuous glucose monitor • Characterized by impaired • Insulin pump glucose metabolism High blood glucose Low blood glucose https://jdrf.org.uk/information-support/treatments- https://jdrf.org.uk/news/diabetes-tech-whats-on-its-way/ technologies/continuous-glucose-monitors/ T. Prioleau - Department of Computer Science 3

Project: Digital SMD Developing Digital Tools for Improved Self-Management of Diabetes User Studies • Understand clinical and patient needs • Evaluate technology solutions Data Analytics Pervasive Sensing • Investigate behavioral factors • Leverage mobile and wearable that affect management technology for continuous sensing • Develop algorithmic solutions to inform treatment T. Prioleau – Department of Computer Science 4

Digital SMD User Studies Data Mining Reflection Needs This Talk Adherence to for Personal CGM Use Health Data Data Mining Machine Learning Behavioral Algorithms for Other Work Sensing More Factors and Blood Glucose than Physiology Glucose Trends Prediction T. Prioleau – Department of Computer Science 5

Digital SMD User Studies Data Mining Reflection Needs Adherence to for Personal CGM Use Health Data T. Prioleau , A. Sabharwal, M. Vasudevan, “Understanding Reflection Needs for Personal Health • Data in Diabetes”, Int. Conf. on Pervasive Computing Technologies for Healthcare, May 2020. (forthcoming) T. Prioleau – Department of Computer Science 6

Reflection on Personal Health Data Continuous Glucose Monitor (CGM) • Enable real-time monitoring of glycemic trends to inform immediate corrective action • Enable retrospective analysis of glucose trends to inform future management <- Significantly underutilized Limitation • A lack of effective methods for data reporting and interpretation • For example, physicians tend to review primarily the most recent [two-weeks] data to assess management during clinical visits • Potentially missing insights in long-term data that is currently not reviewed/analyzed Fig: Conventional visual is daily overlay plot [https://www.dexcom.com/clarity] T. Prioleau – Department of Computer Science 7

Two-Phase User Study Needs Develop Concept Improve User Study Assessment Visualization Validation Visualization (Phase 2) (Survey) (PixelGrid_v1) (Phase 1) (PixelGrid_v2) Participants: • 20 subjects (10 patients with diabetes, 7 clinicians, 3 care-givers) • 5 subjects in Phase 1 user study • 15 subjects in Phase 2 user study T. Prioleau - Department of Computer Science 8

Needs Assessment Survey Example Question: 1. What do you wish CGM reports included that could help and/or speed up analysis and interpretation of diabetes management? Top responses: 1. Automatically extract trends/patterns, detect changes in management and suggest specific behavior changes 2. Integrate data from insulin pumps, bluetooth insulin pens, and carbohydrate intake 3. Enable reviewing and analysis of multiple days of day T. Prioleau - Department of Computer Science 9

DataViz + Concept Validation PixelGrid : m x n matrix-based plot • Designed to facilitate comparison by date, time of day, and day of week • Recurrent time block chosen to identify underlying temporal patterns 1-Week CGM Assessment Low Wear Time (< 50%) Very High BG (> 250 mg/dL) High BG (180 - 250 mg/dL) Multi-Week Summary In Range BG (70 - 180 mg/dL) q Poor TIR = 26% Low BG (54 - 70 mg/dL) q Moderate TIR = 50% Very Low BG (< 54 mg/dL) q Good TIR = 24% Insufficient Data Poor TIR (< 50%) Moderate TIR (50 - 70%) Good TIR (> 70%) * TIR means Time-in-Range T. Prioleau - Department of Computer Science 10 Sun

User Study – Q&A Example Question: 1. What do you like or NOT like about the PixelGrid? Example responses: If there is no detectable patterns, you [the end- I do like how I can visualize overall patterns in ONE user] will just see random screenshot. colors which can be frustrating to a patient. S18 (physician) S12 (patient) Outcome: Feedback will be used to inform future work on insight- suggestive visualization of long-term diabetes data. 11

Follow-up Work 1. Automated algorithm to identify patterns of adverse glycemic events and support personalized goal setting 2. Integrate data from other mobile devices (e.g. insulin pumps & smartphones) to contextualize adverse glycemic events T. Prioleau - Department of Computer Science 12

Digital SMD - Ongoing Work User Studies Data Mining Reflection Needs Adherence to for Personal CGM Use Health Data • S. Vhaduri, T. Prioleau , “Adherence to Personal Health Devices: A Case Study in Diabetes Management,” Int. Conf. on Pervasive Computing Technologies for Healthcare, May 2020. (forthcoming) 10/10/19 T. Prioleau – Department of Computer Science 13

Adherence to CGM Devices Background CGM Use • CGM use = greatest prediction for lowering hemoglobin A1C [American Diabetes Association] Problem • Adherence to CGM use is a challenge for some patients with diabetes https://hcp.eversensediabetes.com/why-eversense-cgm Research Objective • Investigate whether and to what extent achieving target glycemic goals affect wearing behavior of CGMs in diabetes management T. Prioleau - Department of Computer Science 14

Approach Data-driven analysis of patient data (n=44) 1. Statistical tests to study the significance of data gaps in BG categories 2. Evaluated the duration of data gaps in normal vs. abnormal BG categories T. Prioleau - Department of Computer Science 15

Understanding Data Gaps Data gap is a period of contiguous missing data in CGM recording Assumption for analysis: • Data gaps are associated with non-wearing given that a physiological signal (e.g. blood glucose) always exists when wearing device is worn. • However, other reasons for data gaps include sensor malfunction/battery life Fig: conventional daily overlay plot showing 1-week data Data Gap Characterization: 1. Increase in BG (e.g. C) 2. Decrease in BG (e.g. A) T. Prioleau - Department of Computer Science 16

Some Results 1. Subjects with poorly controlled diabetes (A1C > 7) had on average longer data gaps (i.e. worse adherence to CGMs) than subjects with well- controlled diabetes (A1C < 7) 2. Longer duration of data gaps occurred in suboptimal BG categories and the longest data gaps occurred in the most severe categories (i.e. very low, very high) T. Prioleau - Department of Computer Science 17

Key Takeaway Implement data-driven adherence analysis in embedded algorithms to better understand context associated with CGM nonadherence Improved Increased CGM Glycemic Use Outcomes T. Prioleau - Department of Computer Science 18

Digital SMD User Studies Data Mining Reflection Needs This Talk Adherence to for Personal CGM Use Health Data Data Mining Machine Learning Algorithms for Other Work Behavioral Factors Sensing More than Blood Glucose and Glucose Trends Physiology Prediction • S. Morton, R. Li, S. Dibbo, T. Prioleau , “Data-driven Insights on Behavioral Factors that Affect Diabetes Management”, Under Review, 2020 . • K. Gu, R. Dang, T. Prioleau , "Neural Physiological Encoder: A Simple Module for Blood Glucose Prediction," Under Review, 2020. T. Prioleau – Department of Computer Science 19

Thank You! Questions? Endocrine Team - Dartmouth Hitchcock Medical Center (DHMC) Research Assistants Support https://www.tidepool.org/ Contact: Temiloluwa Prioleau; E. tprioleau@dartmouth.edu; W: www.t-prioleau.com 20