Machine learning natural language processing for identifying venous thromboembolism: systematic review and...

Area of Science:

• Medical Informatics
• Computational Linguistics
• Health Services Research

Background:

• Venous thromboembolism (VTE) is a significant cause of preventable hospital mortality.
• Manual review of medical records and diagnosis codes hinders efficient VTE case monitoring.
• Natural Language Processing (NLP) offers automation potential, with Machine Learning (ML)-NLP showing promise.

Purpose of the Study:

• To systematically review and meta-analyze studies using ML-NLP for VTE diagnosis identification in electronic health records (EHRs).
• To evaluate the pooled performance of ML-NLP models in detecting VTE diagnoses.

Main Methods:

• Systematic review of studies published before May 2023, excluding rule-based NLP methods.
• Meta-analysis of the best-performing ML-NLP models for pulmonary embolism and/or deep vein thrombosis detection.
• Calculation of pooled sensitivity, specificity, PPV, and NPV using a random-effects model.
• Assessment of study quality using an adapted TRIPOD tool.

Main Results:

• Thirteen studies were included; 8 provided data for meta-analysis.
• Pooled sensitivity: 0.931, specificity: 0.984, PPV: 0.910, NPV: 0.985.
• Top models utilized vectorization and deep learning (e.g., CNNs).
• Studies demonstrated fair quality, meeting most TRIPOD criteria, but heterogeneity and limited external validation were noted.

Conclusions:

• ML-NLP models demonstrate high accuracy for identifying VTE diagnoses in EHRs.
• Further standardization and external validation are needed for real-world implementation of ML-NLP in VTE monitoring.