Abstract
Background
Cohort selection for specialized clinical trials is a cardinal pillar of the evidence-based medicine; however, it is the most difficult, complex, time consuming, and expensive step. Determining the efficacy of a new treatment (or intervention) requires finding eligible patients meeting the inclusion and exclusion criteria. In specialized scenarios, the complex criteria may even require researchers to do time consuming manual reviews and analyses of electronic health records (EHRs) to shortlist qualified patients. The major contribution of this research is a set of novel semantic models to build a cohort for clinical research by enabling semantic search over electronic health records represented as Knowledge graphs (KGs).
Methods
We present the design of a novel cohort retrieval system, satisfying inclusion and exclusion criteria of an oncology clinical research study. Knowledge graphs and semantic models: We construct knowledge graphs (KGs) to interconnect different data sources, stored in a data-lake, and develop semantic models that enable semantic search over the processed data. We designed and constructed an oncology knowledge graph that enables semantics driven cohort selection. In addition, we have built a novel cohort retrieval system, satisfying the inclusion and exclusion criteria of an oncology clinical research study, that utilizes a semantics driven dynamic query engine to generate and execute cohort selection queries on heterogeneous EHR data.
Results
We obtained real world oncology data of 21,000 oncology patients and then constructed knowledge graphs for five cancer types - Colon (C18), Lung (C34), Breast (C50), prostate (C61), and Multiple Myeloma (C90). The cohort building scenarios are designed to represent a mix of criterion types and combinations including both inclusion and exclusion criteria, involving conjunctions, disjunctions and numeric ranges. Our cohort builder is evaluated against ten well known key competency scenarios. A team of experts validated the results of our cohort builder obtained against these competency scenarios by directly querying the graph. Our extensive evaluations demonstrate that the cohort builder searches with 100% accuracy the patients that match the criteria specified in all ten competencies. The average time to build the cohort on all graphs for these competencies is less than 10 seconds compared with that of days when patients are manually searched in EHR systems.
Conclusions
Our query engine is not tightly coupled with the architecture of our data-lake; rather its architecture is flexible and can be easily integrated with other enterprise data-lakes or EMR systems. In future, we plan to scale the extent of inclusion and exclusion criteria to provide interoperability with existing clinical trial knowledge. We also aim to empirically evaluate the efficiency of cohort selection queries using a knowledge graph with classical database query approaches.
Key findings
- 1The bottleneck it targets: finding patients who meet a trial's inclusion and exclusion criteria is the most difficult, time-consuming and expensive step of cohort selection, and complex criteria can force manual review of EHR records.
- 2EHR data turned into knowledge graphs: data sources held in a data lake are interconnected as knowledge graphs, with semantic models that allow search by meaning over the processed records, including a purpose-built oncology knowledge graph.
- 3Criteria become queries automatically: a semantics-driven dynamic query engine generates and runs cohort selection queries against heterogeneous EHR data.
- 4Built on real-world oncology data: 21,000 oncology patients, with knowledge graphs for colon (C18), lung (C34), breast (C50) and prostate (C61) cancer and multiple myeloma (C90).
- 5Tested on realistic criteria: ten well-known competency scenarios mixing inclusion and exclusion criteria with conjunctions, disjunctions and numeric ranges, with results validated by a team of experts querying the graph directly.
- 6Every match, in seconds: the cohort builder found the matching patients with 100% accuracy in all ten scenarios, averaging under 10 seconds per cohort against days of manual EHR searching.
Why this matters
Trials stall at screening. A research team that would otherwise read charts patient by patient against each criterion can state the criteria once and get the cohort back in seconds, which turns feasibility questions such as “how many of our patients would qualify?” into something a team can answer before committing to a study.
The results are from ten defined scenarios checked against expert queries of the same graph, not from a live trial screen. The authors designed the query engine so it is not tied to CureMD's own data lake and can be integrated with other enterprise data lakes or EMR systems. Their next steps are wider inclusion and exclusion criteria that work with existing clinical trial knowledge, and a measured comparison of knowledge graph queries with classical database queries.
