An Extensible Evaluation Framework Applied to Clinical Text Deidentification Natural Language Processing Tools: Multisystem and Multicorpus Study

Abstract

Background Clinical natural language processing (NLP) researchers need access to directly comparable evaluation results for applications such as text deidentification across a range of corpus types and the means to easily test new systems or corpora within the same framework. Current systems, reported metrics, and the personally identifiable information (PII) categories evaluated are not easily comparable. Objective This study presents an open-source and extensible end-to-end framework for comparing clinical NLP system performance across corpora even when the annotation categories do not align. Methods As a use case for this framework, we use 6 off-the-shelf text deidentification systems (ie, CliniDeID, deid from PhysioNet, MITRE Identity Scrubber Toolkit [MIST], NeuroNER, National Library of Medicine [NLM] Scrubber, and Philter) across 3 standard clinical text corpora for the task (2 of which are publicly available) and 1 private corpus (all in English), with annotation categories that are not directly analogous. The framework is built on shell scripts that can be extended to include new systems, corpora, and performance metrics. We present this open tool, multiple means for aligning PII categories during evaluation, and our initial timing and performance metric findings. Code for running this framework with all settings needed to run all pairs are available via Codeberg and GitHub. Results From this case study, we found large differences in processing speed between systems. The fastest system (ie, MIST) processed an average of 24.57 (SD 26.23) notes per second, while the slowest (ie, CliniDeID) processed an average of 1.00 notes per second. No system uniformly outperformed the others at identifying PII across corpora and categories. Instead, a rich tapestry of performance trade-offs emerged for PII categories. CliniDeID and Philter prioritize recall over precision (with an average recall 6.9 and 11.2 points higher, respectively, for partially matching spans of text matching any PII category), while the other 4 systems consistently have higher precision (with MIST’s precision scoring 20.2 points higher, NLM Scrubber scoring 4.4 points higher, NeuroNER scoring 7.2 points higher, and deid scoring 17.1 points higher). The macroaverage recall across corpora for identifying names, one of the more sensitive PII categories, included deid (48.8%) and MIST (66.9%) at the low end and NeuroNER (84.1%), NLM Scrubber (88.1%), and CliniDeID (95.9%) at the high end. A variety of metrics across categories and corpora are reported with a wider variety (eg, F2-score) available via the tool. Conclusions NLP systems in general and deidentification systems and corpora in our use case tend to be evaluated in stand-alone research articles that only include a limited set of comparators. We hold that a single evaluation pipeline across multiple systems and corpora allows for more nuanced comparisons. Our open pipeline should reduce barriers to evaluation and system advancement.