Automated natural language processing to identify venous thromboembolism from diagnostic imaging reports

Area of Science:

• Hematology
• Medical Informatics
• Computational Biology

Background:

• Allogeneic hematopoietic stem cell transplant (HSCT) is associated with a significantly increased risk of venous thromboembolism (VTE), potentially up to 50-fold.
• Accurate VTE incidence data and identification of associated clinical risk factors in HSCT patients traditionally rely on time-consuming manual chart reviews.
• Natural Language Processing (NLP) offers a promising automated solution for VTE detection within electronic medical record (EMR) systems.

Purpose of the Study:

• To develop and evaluate an institutional NLP algorithm for detecting acute VTE in patients undergoing HSCT.
• To compare the performance of the NLP algorithm against manual chart review for VTE identification within 100 days post-HSCT.
• To assess the accuracy metrics (sensitivity, specificity, PPV, NPV) of the NLP tool in this specific patient population.

Main Methods:

• Retrospective analysis of adult patient records from 2016 to 2020.
• Development of an NLP algorithm to identify acute VTE within 100 days of HSCT from EMR data.
• Comparison of NLP-based VTE detection with traditional manual chart review for performance evaluation.

Main Results:

• A total of 1300 electronic health records were analyzed.
• The 100-day VTE incidence was 10.3% by manual review and 8.8% by NLP.
• The NLP algorithm demonstrated high performance with specificity, sensitivity, PPV, and NPV exceeding 0.85, with minor discrepancies attributed to overlooked report types.

Conclusions:

• The developed NLP algorithm shows excellent performance in identifying VTE in HSCT patients.
• NLP significantly enhances the efficiency of VTE detection compared to manual chart review in this high-risk cohort.
• Future refinements and integration of NLP with other methods could further improve VTE detection in HSCT and other at-risk populations.