Predicting Progression to Type 2 Diabetes David Hurwitz
Predicting Progression to Type 2 Diabetes David Hurwitz, MD, FACP Medical Director of Clinical Analytics, Allscripts
Disclosure I disclose the following relevant financial relationship with commercial interests: • I am an Allscripts employee • My presentation does not promote the use of Allscripts or other commercial products. 2 AMIA 2019 Clinical Informatics Conference
Learning Objectives After participating in this session the learner should be better able to: 1. Understand the role of utilizing EHR and non-EHR data to address healthcare challenges 2. Appreciate the value of large clinical data sets for predictive model development 3. Realize the potential to implement a diabetes predictive model within clinician workflow and for population health management 3 AMIA 2019 Clinical Informatics Conference
Prevalence and Cost of Diabetes and Pre Diabetes (2017) American Diabetes Association [1] Ø 30 million Americans with diabetes Ø $327 billion total cost of diagnosed diabetes (health costs, lost productivity) Ø 1 in 7 health care dollars Ø 84 million Americans with pre diabetes Ø Risk for progression to Type 2 Diabetes Ø Nearly 60% progress to Type 2 diabetes over 10 years [2] Ø Lifestyle (moderate exercise, 5 -10% weight loss) and pharmacologic interventions (metformin) can delay or prevent progression to Type 2 Diabetes 4 AMIA 2019 Clinical Informatics Conference
Predictive Model Goal: Accurate prediction of transition from pre-diabetes to Type 2 diabetes within 24 months. Internal collaborative effort between clinicians and data science team 5 AMIA 2019 Clinical Informatics Conference
Methods Data Sources • EHR • Data Warehouse (60 million patient lives) • Non-EHR • U. S. Internal Revenue Service • Food Access Research Atlas (FARA) 6 AMIA 2019 Clinical Informatics Conference
Methods Definitions • Pre-diabetes: HBA 1 C performed • Type 2 Diabetes: International Classification of Disease (ICD) diagnostic codes • Mean adjusted gross income: U. S. Internal Revenue Service [3] • Calculated by dividing the sum of the average of all income in a given state and zip code, by tax bracket, by the sum of the number of tax returns • Healthy food access, Food Access Research Atlas (FARA) [4] • Provides a measure of neighborhood access to healthy, affordable food, focused on low-income and lowsupermarket-access census tracts Predictive Model • Binary classification using a feed forward neural network • 30+ features 7 AMIA 2019 Clinical Informatics Conference
Results Study Population Years Pre-Diabetes 2015 -2016 149, 050 Transition to Diabetes Within 24 Months (%) Geo Distribution (US) 65, 044 (43. 6) 50 States Training and Test Data Transition to Diabetes within 24 Months Train Test Yes 47, 845 20, 673 No 55, 447 24, 325 8 AMIA 2019 Clinical Informatics Conference
Pre Diabetes Patient Distribution by County 9 AMIA 2019 Clinical Informatics Conference
Number of Pre Diabetes Patients by County 10 AMIA 2019 Clinical Informatics Conference
Other Pre Diabetes Demographic Examples 11 AMIA 2019 Clinical Informatics Conference
Predictive Model Final Feature Set • • EHR • Age Non-EHR • Mean adjusted gross income • Gender • Ethnicity • Race • Hypertension diagnosis • Family History of Diabetes • Body Mass Index • Systolic blood pressure (first, last, min, max) • Diastolic blood pressure (first, last, min, max) • Hemoglobin A 1 C 12 AMIA 2019 Clinical Informatics Conference
Predicting Model Performance Sensitivity: 80% Specificity: 89% Positive Predictive Value: 86% Negative Predictive Value: 90% 13 AMIA 2019 Clinical Informatics Conference
Predictive Model Accuracy Accuracy: 0. 86219 14 AMIA 2019 Clinical Informatics Conference
Discussion A large data set of clinical and social determinant data was used to derive an accurate model to predict progression of pre diabetes to Type 2 diabetes within 24 months. This model compares favorably to a previously published model [5] with an AUC of 0. 78 for progression to Type 2 diabetes. The model can potentially help health systems proactively identity individuals in a pre diabetes population at risk for progression to Type 2 Diabetes and implement risk mitigating interventions. 15 AMIA 2019 Clinical Informatics Conference
Discussion Strengths • Data readily accessible within EHR (mean adjusted gross income identified by zip code) • Large data set • • Reduced noise • Important for accuracy Wide geographic distribution • Increases generalizability Limitations • Did not use actual HBA 1 C values, but used result as proxy for pre diabetes (possible misclassification of diabetes as pre-diabetes in some cases) • Certain features not considered Drugs that raise blood sugar- statins, anti-psychotics, other drug classes Comorbid conditions other than hypertension 16 AMIA 2019 Clinical Informatics Conference
Future Directions • Refine model using additional clinical and social determinant features. Create non-binary output, e. g. probability of progression to Type 2 diabetes over time. • Implement predictive model within clinician workflow to inform clinicians and patients of risk. Measure impact on diabetes incidence across population over time. • Create intuitive predictive visualizations of HBA 1 C that are easy for clinicians and patients to understand. 17 AMIA 2019 Clinical Informatics Conference
Implement Model Within Clinician Workflow Visualization Adapted from CDS Hooks, https: //cds-hooks. org/ 18 AMIA 2019 Clinical Informatics Conference
Intuitive Graphics to Enhance Clinical Decision Support 19 AMIA 2019 Clinical Informatics Conference
Population Health Management 20 AMIA 2019 Clinical Informatics Conference
References 1. The Cost of Diabetes, American Diabetes Association, 2019, http: //www. diabetes. org/advocacy/news-events/costof-diabetes. html 2. Cefalu, WT, Diabetes Care 2016 Aug; 39(8): 1472 -1477. https: //doi. org/10. 2337/dc 16 -1143 3. United States Internal Revenue Service, SOI Tax Stats - Individual Statistical Tables by Size of Adjusted Gross Income, https: //www. irs. gov/statistics/soi-tax-stats-individual-statistical-tables-by-size-of-adjusted-gross-income 4. Economic Research Service (ERS), U. S. Department of Agriculture (USDA). Food Access Research Atlas, https: //www. ers. usda. gov/data-products/food-access-research-atlas/ 5. Anderson JP, Parikh JR, Shenfeld DK, Ivanov V, Marks C, Church BW, Laramie JM, Mardekian J, Piper BA, Willke RJ, Rublee DA. Reverse Engineering and Evaluation of Prediction Models for Progression to Type 2 Diabetes: An Application of Machine Learning Using Electronic Health Records. J Diabetes Sci Technol. 2015 Dec 20; 10(1): 6 -18. doi: 10. 1177/1932296815620200. 21 AMIA 2019 Clinical Informatics Conference
Additional Reading 1. National Diabetes Statistics Report, 2017, Centers for Disease Control, 2017, https: //www. cdc. gov/diabetes/pdfs/data/statistics/national-diabetes-statistics-report. pdf 2. Che Ngufor, Holly Van. Houten , Brian S. Caffo, Nilay D. Shah, Rozalina G. Mc. Coy, Mixed effect machine learning: A framework for predicting longitudinal change in hemoglobin A 1 c, Journal of Biomedical Informatics, Volume 89, January 2019, Pages 56 -67, https: //www. sciencedirect. com/science/article/pii/S 1532046418301758? via%3 Dihub 22 AMIA 2019 Clinical Informatics Conference
AMIA is the professional home for more than 5, 600 informatics professionals, representing frontline clinicians, researchers, public health experts and educators who bring meaning to data, manage information and generate new knowledge across the research and healthcare enterprise. AMIA 2019 Clinical Informatics Conference | amia. org @AMIAInformatics @AMIAinformatics Official Group of AMIA @AMIAInformatics #Why. Informatics 23
Thank you! david. hurwitz@Allscripts. com
- Slides: 24