Progress 06/15/12 to 05/31/13
Outputs Target Audience: Health disparity populations living in rural regions of the Tennessee and Mississippi delta who have diabetes. Changes/Problems:
Nothing Reported
What opportunities for training and professional development has the project provided?
Nothing Reported
How have the results been disseminated to communities of interest?
Nothing Reported
What do you plan to do during the next reporting period to accomplish the goals?
Nothing Reported
Impacts What was accomplished under these goals?
Phase I SPECIFIC AIMS - Accomplishments Aim 1: Network Hubble's metadata platform interface (MPI) with the Delta Health Alliance's EHR to determine the feasibility of a population-based screenign program. Milestone: generate a minimum of 800 clinical metadata analyses Accomplishments: Under a University of Tennessee IRB-approved protocol, we collected and anonymized historical EHR retinal image and metadata sets comprising over 800 patients in collaboration with the Delta Health Alliance (DHA) and the UTHSC Office of Biomedical Informatics (BMI). These data were from patients who had previously been examined for DR through Hubble Telemedical’s TRIAD ocular telehealth network. The metadata fields selected were based upon the diabetic and ophthalmic literature on DR epidemiology, co-morbidity, and risk, and comprised the following:Gender, age, and ethnicity (risk factors);BMI and other obesity measures (risk factors);All medications (e.g., insulin dose, oral hypoglycemic, antihypertensives/duration of use);Duration and type of diabetes, age at onset (risk factor for DR severity);Hemoglobin A1C values and historical trends (i.e., long term control of blood sugars);Systolic, diastolic, mean arterial blood pressure, age at onset and control (co-morbidity);Cholesterol, triglyceride and lipoprotein levels (LDL, HDL), (co-morbidity);BUN/Creatinine and urinalysis (measures of renal function, microangiopathy);C-reactive protein and other inflammatory markers in laboratory datasets (co-morbidity);Prior laser or surgery for DR (by history or evidence of retinal features and severity);Diabetic complications (renal disease, coronary artery disease, or stroke) (co-morbidity), and;Social history - smoking history, current pregnancy (co-morbidity). Using novel analytical methods developed through this project, our goal was to identify and validate specific longitudinal meta-data profiles that were highly correlated with the presence and severity of DR and also its absence. We utilized multiple strategies for the analyses, described in detail below. The performance of the diagnostic algorithms was statistically validated using retinal image datasets ground-truthed by a retina expert at the UT Hamilton Eye Institute (Dr. Chaum, CMO of Hubble Telemedical). Aim 2:Determine the required network interface features. Milestone:automated metadata field-populating, facility of data acquisition, accessible EHR data mining parameters, HIPAA compliance, and auditing functions. Accomplishments: We employed various methods for performing fusion between disparate image and non-image data sets. These included a machine learning-based system for a similar application to predict the occurrence of DR, nonlinear classifications such as neural networks, and multiple-classifier approaches, where a set of classifiers are trained based on individual characteristics or sub-set of characteristics and a fusion method is used. Utilizing similar methods in a different cohort, we recently showed an important correlation between DR incidence and African American race, even among age-matched controls. Although the increased prevalence of DR among Latino groups is known, the association of African American race to increased disease incidence is a new and important finding. Metadata indicates some correlation between patient age (and duration of disease) and higher classes of DR. However, these results do not suggest a causal relationship between DR and the patient’s age since Class 1 (No DR) and Class 2 (Mild DR-CSME) span somewhat evenly across the entire age range. This will be further explored by looking at more complex correlations during the Phase II project. For example, the data show a significant correlation between both lower and higher levels of DR severity and the prevalence of high systolic blood pressure (BP), coronary artery disease (CAD), myo-cardial infarction (MI), and stroke, whereas the prevalence of renal disease is closely correlated only with more severe DR. Under the Phase I project, we made substantial strides in developing protocols for meaningful data extraction from EHRs to use for medical informatics analysis. To date, we have focused on the development of data-extraction, transform and load techniques, and harmonization methods required to overcome the proprietary software and database structures of commercial EHR vendors. These standards and functions are essential if a truly interoperable system is to be developed. The GUI interface developed under the Phase I project included the following features:Interoperability between various EHR platforms;Automatic data acquisition, field population, and report generation;Visual Analytics and Graphical User Interface; andHIPAA-compliance. This application involved the secure migration of select clinical metadata fields from the AllScripts EHR used by the DHA, and their integration with DICOM images acquired by Hubble Telemedical acquired on the TRIAD platform. Due to the implementation of universal standards such as the ICD and CPT codes, core data such as diagnoses, procedures, demographics, co-morbidities etc. are essentially identical within EHR systems; thus, once one knows the underlying architecture of a system, it is straightforward to create a harmonization protocol using a unified data dictionary. The BMI and Hubble teams identified which fields could be migrated as-is into Slim-Prim and which fields required further analysis and harmonization in order to provide input to address key research questions. Although most demographic and diagnostic fields were standard, some Hubble questions were not answered directly from DHA data but needed to be inferred. For example, “Type of Diabetes” was not always available in the DHA database; however, other metadata (e.g., age at diagnosis and diet control) could be used to make an informed inference about the disease status. Once the migration data dictionary was complete, the clinical metadata from the DHA cohort was imported using secure FTP into a secure ‘staging zone’ inside the Slim-Prim application servers. This staging zone was the initial environment for harmonization of disparate EHR database outputs onto a common database structure. The dataset was uploaded as an encrypted comma separated value (.csv) file, and harmonized via extraction-transform-load protocols (ETL) with the Hubble module in the Slim-Prim database. Where possible, tasks were automated with PHP scripts and when necessary SQL scripts were used to convert data types to the Hubble module data structure. Post-harmonization, the transformed dataset was integrated with pre-existing records in the Hubble module, thus creating a longitudinal patient record. After successful harmonization, the original CSV files were deleted from the server to minimize risk of HIPAA violations. Finally, the metadata in the Hubble module were linked via key identifiers to DICOM retinal image files for each patient so that search algorithms could be run through the project report interface. In this manner Slim-Prim served as the central repository linking EHR data with image files in an environment for the algorithms to be tested and validated. In addition to designable report queries, an integrated Pivot Viewer tool houses the image repository containing retinal images, together with associated patient metadata. This feature allows users to quickly visualize and test the statistical significance of metadata correlations, trends, and outliers in both simple and complex queries. Users can save reports for future data retrieval and analysis.
Publications
|