Can Modern NLP Systems Reliably Annotate Chest Radiography Exams? A Pre-Purchase Evaluation and Comparative Study of Solutions from AWS, Google, Azure, John Snow Labs, and Open-Source Models on an Independent Pediatric Dataset

TL;DR

This study evaluates the reliability of various commercial and open-source NLP tools in automatically annotating pediatric chest radiograph reports, revealing significant variability and emphasizing the need for validation before clinical deployment.

Contribution

It provides a comprehensive comparison of multiple NLP solutions on a large pediatric dataset, highlighting their strengths and limitations in clinical report annotation.

Findings

Significant differences in entity extraction counts across NLP systems.

Assertion detection accuracy varied, with SparkNLP achieving the highest at 76%.

Commercial NLP tools showed variable performance, underscoring the need for validation.

Abstract

General-purpose clinical natural language processing (NLP) tools are increasingly used for the automatic labeling of clinical reports. However, independent evaluations for specific tasks, such as pediatric chest radiograph (CXR) report labeling, are limited. This study compares four commercial clinical NLP systems - Amazon Comprehend Medical (AWS), Google Healthcare NLP (GC), Azure Clinical NLP (AZ), and SparkNLP (SP) - for entity extraction and assertion detection in pediatric CXR reports. Additionally, CheXpert and CheXbert, two dedicated chest radiograph report labelers, were evaluated on the same task using CheXpert-defined labels. We analyzed 95,008 pediatric CXR reports from a large academic pediatric hospital. Entities and assertion statuses (positive, negative, uncertain) from the findings and impression sections were extracted by the NLP systems, with impression section entities mapped to 12 disease categories and a No Findings category. CheXpert and CheXbert extracted the same 13 categories. Outputs were compared using Fleiss Kappa and accuracy against a consensus pseudo-ground truth. Significant differences were found in the number of extracted entities and assertion distributions across NLP systems. SP extracted 49,688 unique entities, GC 16,477, AZ 31,543, and AWS 27,216. Assertion accuracy across models averaged around 62%, with SP highest (76%) and AWS lowest (50%). CheXpert and CheXbert achieved 56% accuracy. Considerable variability in performance highlights the need for careful validation and review before deploying NLP tools for clinical report labeling.