Temporal Characterization of Clinical Trial Descriptions

Objectives/Goals: Issues with recruiting the targeted number of participants in a timely manner often results in underpowered studies, with more than 60% of clinical studies failing to complete or requiring extensions due to enrollment issues. The objective of this study is to understand temporal complexities within clinical trial descriptions in order to develop tractable computational approaches for automatic matching of trial descriptions with patient records.

Methods/Study Population: In order to understand the temporal complexity present within clinical trial descriptions, we did the following:

1. We randomly selected 100 cardiac study descriptions from Clinicaltrials.gov (as of September 2017), and had two experts in clinical vocabularies independently categorize the trial descriptions into those having and lacking temporality. The experts discussed discrepancies in their categorizations and came up with a consensus.
2. We automatically mapped all the trial description available from ClinicalTrials.gov (as of September 2017) to Unified Medical Language System (UMLS) concepts using Metamap version 2016v2. We then extracts concepts of the semantic type ‘Temporal Concept (TMCO)’ and analyzed their description across the trial descriptions. In addition, we reconstructed the UMLS hierarchy for the concepts available within the trial descriptions in order to group them into semantic groups using the UMLS version 2017AB. We also analyzed these UMLS concepts with existing time related taxonomies such the Time ontology and Allen’s Interval Algebra.

Results/Anticipated Results:

1. The three experts categorized 86 and 90 percent of the clinical trial descriptions as having temporalities within them. Upon discussion, the experts arrived at a consensus result of 89% of the trial descriptions having temporality within them.
2. 90% of the trial descriptions from ClinicalTrials.gov contained one or more UMLS concepts of the ‘Temporal Concept’ semantic type. We found 1019 distinct UMLS concepts of the within the trial descriptions from ClinicalTrials.gov. The top 20 concepts were present in at least 10% of the trial descriptions. 206 of these 1019 concepts were equivalent to or contained the thirteen elementary relations between time periods as described the Time ontology. The existing relationships between the extracted temporal concepts resulted in 46 trees and 427 unconnected concepts. Informed by Allen’s Interval Algebra and the Time ontology along with existing UMLS tree structures within these concepts, we manually assigned the concepts into 5 categories:
   • Stage/Duration/Interval
   • Time point/Instance/Event
   • Time Frequency/Pattern
   • Time Unit/Measure
   • Temporal Relation

We also found that many of the concepts belonged to more than one of these categories forming a complex network of temporal concepts and their use in trial descriptions was context dependent.

Discussion/Significance of Impact:

Our results from above two methods show that most clinical trial descriptions include a temporal dimension. The semantics of the temporality within these descriptions is also very complex and is context dependent. The temporality in clinical trial description could be unbounded; dense; and represented as instants or intervals including non-convex intervals, branching and circular times, and relative as well as absolute times. In addition, there might be differences in temporal granularities, and indeterminacy or veracity of time (incomplete knowledge of when considered event or fact happened), uncertainty regarding how long event or fact was true.

Time is an essential component of most entries in a patient record. Key domains in this data such conditions, procedures, medications and laboratory as all documented as events in relation to patient visits and actual ages of patients. In addition, our analysis of notes within the MIMIC-III data using similar approaches as described above show the that patient notes have similar distributions of temporal concepts. 

Automated matching of clinical trial descriptions with patient records would there need to accommodate the temporal nature of both these document types.  Our future work will include the development of temporal reasoning methods to optimize patient recruitment based on our current analysis findings. These methods will utilize fuzzy time-interval pattern mining over the electronic health record based patient trajectories to predict the most adequate requirement schedules based on the current and potential patient states.