Bridging the Semantic Gap Between Research Eligibility Criteria and Clinical Data
Location(s): United States
Averaging about $1,000 per patient1, recruitment remains an expensive bottleneck for human studies. the rapidly increasing adoption of electronic health records (EHR) has made electronic prescreening (E-screening hereafter) a practicable solution to this bottleneck. Our long-term goal is to achieve this "holy grail". Our short- term goal of this competing continuation is to develop an intelligent patient query consultant to improve the accuracy and efficiency of E-screening. One of the difficulties for E-screening is the Semantic Gap between eligibility criteria and clinical data.2 Each eligibility criterion (e.g., hypertension) describes a patient characteristic, which is correlated with multiple data features (e.g., orders of hypertension drugs, elevated blood pressure, and symptoms of hypertension) in EHR. Moreover, each data feature may have multiple Semantic representations (e.g., "SBP", "BP", or "blood pressure") from disparate data sources. For example, elevated systolic blood pressure can be recorded in varying formats in an emergency room, a doctor's office, an ICU, and an in-patient unit, but not all of these readings necessarily indicate chronic hypertension. the use of clinical data to identify patients eligible for clinical research requires specialized knowledge and expert guidance to navigate the vast space of data features and intelligent inferences from data features for eligibility determination. A user must understand the characteristics of available data before using them to search for patients. For example, when only 5% of hypertensive patients have ICD-9 codes for hypertension but 73% of these patients have hypertension drug orders, using drug information to construct a query of hypertensive patients will be more effective than one using ICD-9 codes. Even sophisticated biomedical data query tools such as i2b2, VISAGE, and STRIDE only passively translate user-specified data features into a query statement. They do not guide a researcher in selecting a data feature and its most appropriate Semantic representations or data sources. Little aid is available to inform researchers about data characteristics or to help them conduct exploratory data analyses for optimal data feature selection. Mixed-initiative interaction,3 which allows human and computer to collaboratively contribute to converged problem solutions, can potentially fulfill this need. We hypothesize that by equipping biomedical researchers with a knowledge-based, mixed-initiative dialog system, we can maximize the efficiency and accuracy of E- screening by supporting exploratory analyses of correlated data features for query optimization. Our approach is innovative because it (1) addresses theuser needs for intelligent query interfaces for clinical data, (2) provides a novel data-driven approach to eligibility determination based on correlated data features, and (3) enables efficient query optimization through support of human-computer collaborative problem solving. We will build on the results from our first funding period for Bridgingthe Semantic Gap.4-21 We developed an analysis pipeline called EliXR to construct a Semantic knowledge representation for eligibility criteria 6,9,16,17, which can be used to transform free-text eligibility criteria ito structured narrative.6 We developed methods to dynamically categorize eligibility criteria by data type.8 We accumulated E-screening experience from three NIH-sponsored clinical trials.7,13,21 We developed a method combining PubMed knowledge and EHR data to infer patient phenotype4 and reconciled structured and unstructured clinical data to support E-screening.18 We are prepared with methods and a preliminary understanding of the building blocks necessary to optimally translate eligibility criteria into data features; therefore, our current proposal is the logical next step. Our specific aims are to: 1. Use mixed methods to understand theneeds of biomedical researchers for query clarification and identify common strategies used by query analysts for plan optimization for complex eligibility queries. 2. Develop a knowledge-based, mixed-initiative dialog system to improve human-computer collaboration for query formulation using participatory design methods. 3. Evaluate theefficacy and usability of the mixed-initiative dialog system using a research data warehouse and two use cases: research protocol feasibility testing and trial recruitment prescreening. We will advance the field by contributing knowledge of the needs for query support among biomedical researchers and an effective E-screening method that combines intelligent query recommendation and iterative query by review22 to improve data access for researchers through human-computer collaboration.