A natural language processing algorithm to define a venous thromboembolism phenotype.

Area of Science:

• Medical Informatics
• Clinical Research
• Health Data Science

Background:

• Deep venous thrombosis (DVT) and pulmonary embolism (PE) contribute significantly to patient morbidity and mortality.
• Identified risk factors account for only a small proportion of venous thromboembolic disease (VTE) cases.
• Unidentified genetic factors are presumed to play a role in the remaining VTE cases.

Purpose of the Study:

• To develop and validate a general-purpose Natural Language Processing (NLP) algorithm for retrospective VTE case identification.
• To assess the algorithm's effectiveness in capturing both acute and historical VTE cases from de-identified electronic health records (EHRs).

Main Methods:

• A Natural Language Processing (NLP) algorithm was designed to extract VTE cases from EHRs.
• The algorithm was applied to training and separate evaluation datasets to determine its performance metrics.
• Performance was evaluated using positive predictive value (PPV), sensitivity, and F-measure.
• Specific strategies, including analysis of problem lists without VTE ICD-9 codes, were tested for capturing historical cases.

Main Results:

• The NLP algorithm demonstrated high performance in the evaluation set with a positive predictive value (PPV) of 84.7% and sensitivity of 95.3% (F-measure 0.897).
• Performance on the training set was comparable, with an F-measure of 0.925.
• Utilizing the algorithm on problem lists for patients without VTE ICD-9 codes proved most effective for identifying historical VTE cases, achieving a PPV of 83%.

Conclusions:

• The developed NLP algorithm offers an effective approach for capturing both acute and historical cases of venous thromboembolic disease (VTE).
• Combining NLP analysis of VTE ICD-9 codes with non-ICD-9 problem lists enhances the comprehensive identification of VTE.
• This methodology holds promise for improving VTE surveillance and research using EHR data.