# Named Entity Recognition in Unstructured Medical Text Documents

arXiv: https://arxiv.org/abs/2110.15732 · HTML: https://mlsanity.com/paper/2110.15732/

## Key facts

- Cited by 3 papers in the map (data through 14 Oct 2026).
- Builds on 2 of its 34 references that are in the map.
- First posted on arXiv 15 Oct 2021.

## Builds on (most cited first, 2 of 2)

- [De-identification of Patient Notes with Recurrent Neural Networks](https://mlsanity.com/paper/1606.03475/) (2016)
- [Named Entity Recognition in Electronic Health Records Using Transfer Learning Bootstrapped Neural Networks](https://mlsanity.com/paper/1901.01592/) (2019)

## Built on by (most cited first, 3 of 3)

- [Mitigating Presentation Attack using DCGAN and Deep CNN](https://mlsanity.com/paper/2207.00161/) (2022)
- [Korean Bio-Medical Corpus (KBMC) for Medical Named Entity Recognition](https://mlsanity.com/paper/2403.16158/) (2024)
- [A Comparative Study of Light-weight Language Models for PII Masking and their Deployment for Real Conversational Texts](https://mlsanity.com/paper/2512.18608/) (2025)

## Abstract (arXiv)

Physicians provide expert opinion to legal courts on the medical state of patients, including determining if a patient is likely to have permanent or non-permanent injuries or ailments. An independent medical examination (IME) report summarizes a physicians medical opinion about a patients health status based on the physicians expertise. IME reports contain private and sensitive information (Personally Identifiable Information or PII) that needs to be removed or randomly encoded before further research work can be conducted. In our study the IME is an orthopedic surgeon from a private practice in the United States. The goal of this research is to perform named entity recognition (NER) to identify and subsequently remove/encode PII information from IME reports prepared by the physician. We apply the NER toolkits of OpenNLP and spaCy, two freely available natural language processing platforms, and compare their precision, recall, and f-measure performance at identifying five categories of PII across trials of randomly selected IME reports using each models common default parameters. We find that both platforms achieve high performance (f-measure > 0.9) at de-identification and that a spaCy model trained with a 70-30 train-test data split is most performant.