Improved patient identification and feature extraction through free text query and processing for clinical research

Electronic Health Records (EHRs) provide a wealth of information for secondary uses. Methods are developed to improve usefulness of free text query and text processing and demonstrate advantages to using these methods for clinical research, specifically cohort identification and enhancement. Cohort identification is a critical early step in clinical research. Problems may arise when too few patients are identified, or the cohort consists of a nonrepresentative sample. Methods of improving query formation through query expansion are described. Inclusion of free text search in addition to structured data search is investigated to determine the incremental improvement of adding unstructured text search over structured data search alone. Query expansion using topic- and synonym-based expansion improved information retrieval performance. An ensemble method was not successful. The addition of free text search compared to structured data search alone demonstrated increased cohort size in all cases, with dramatic increases in some. Representation of patients in subpopulations that may have been underrepresented otherwise is also shown. We demonstrate clinical impact by showing that a serious clinical condition, scleroderma renal crisis, can be predicted by adding free text search. A novel information extraction algorithm is developed and evaluated (Regular Expression Discovery for Extraction, or REDEx) for cohort enrichment. The REDEx algorithm is demonstrated to accurately extract information from free text clinical iv narratives. Temporal expressions as well as bodyweight-related measures are extracted. Additional patients and additional measurement occurrences are identified using these extracted values that were not identifiable through structured data alone. The REDEx algorithm transfers the burden of machine learning training from annotators to domain experts. We developed automated query expansion methods that greatly improve performance of keyword-based information retrieval. We also developed NLP methods for unstructured data and demonstrate that cohort size can be greatly increased, a more complete population can be identified, and important clinical conditions can be detected that are often missed otherwise. We found a much more complete representation of patients can be obtained. We also developed a novel machine learning algorithm for information extraction, REDEx, that efficiently extracts clinical values from unstructured clinical text, adding additional information and observations over what is available in structured text alone.