https://github.com/plandes/medacy_bertcrf_model_clinical_notes

Clinical Notes Model for medaCy (BERT)
https://github.com/plandes/medacy_bertcrf_model_clinical_notes

bert-model machine-learning medical natural-language-processing

Last synced: 18 days ago
JSON representation

Clinical Notes Model for medaCy (BERT)

• Host: GitHub
• URL: https://github.com/plandes/medacy_bertcrf_model_clinical_notes
• Owner: plandes
• License: gpl-3.0
• Created: 2021-03-27T00:37:29.000Z (almost 4 years ago)
• Default Branch: master
• Last Pushed: 2021-03-30T22:36:41.000Z (almost 4 years ago)
• Last Synced: 2024-11-08T21:27:28.672Z (2 months ago)
• Topics: bert-model, machine-learning, medical, natural-language-processing
• Language: Python
• Homepage:
• Size: 20.5 KB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
[![spaCy](https://img.shields.io/badge/built%20with-spaCy-09a3d5.svg)](https://spacy.io)

# medaCy

:hospital: Clinical Notes Model for medaCy (BERT+CRF) :hospital:

This repository contains a versioned, medaCy compatible Model for information extraction from clinical notes.

![alt text](https://nlp.cs.vcu.edu/images/Edit_NanomedicineDatabase.png "Nanoinformatics")

# Description

This is the light-weight version (no metamap) of medaCy's model for extracting 9 unique entities from clinical notes:

`Drug, Strength, Duration, Route, Form, ADE, Dosage, Reason, Frequency`

This model was trained using the

[ClinicalBERT](https://huggingface.co/emilyalsentzer/Bio_ClinicalBERT)

pre-trained embeddings with CRF.

# Results

Model generalization ability is evaluated over 202 patient clinical note files not seen during training. *Strict* indicates exact matches of spans, *Lenient* indicates a fuzzy matching of spans (model predictions are off by single characters).

| Entity (Count)    |   Precision |   Recall |    F1 |   F1_Min |   F1_Max |

|-------------------|-------------|----------|-------|----------|----------|

| ADE (1584)        |       0.49  |    0.327 | 0.387 |    0.342 |    0.466 |

| Dosage (6902)     |       0.941 |    0.951 | 0.946 |    0.936 |    0.961 |

| Drug (26800)      |       0.904 |    0.891 | 0.898 |    0.883 |    0.905 |

| Duration (970)    |       0.836 |    0.8   | 0.816 |    0.768 |    0.861 |

| Form (11010)      |       0.937 |    0.939 | 0.938 |    0.931 |    0.954 |

| Frequency (10293) |       0.878 |    0.952 | 0.914 |    0.9   |    0.926 |

| Reason (6400)     |       0.653 |    0.513 | 0.571 |    0.554 |    0.598 |

| Route (8989)      |       0.929 |    0.932 | 0.93  |    0.925 |    0.937 |

| Strength (10921)  |       0.956 |    0.955 | 0.955 |    0.95  |    0.961 |

| system (83869)    |       0.894 |    0.896 | 0.893 |    0.889 |    0.902 |

# Training Data

N2C2 2018 Shared Task

The data used to induce this model is protected by HIPAA privacy regulations and thus cannot be published.

Authors

=======

Andriy Mulyar and Bridget McInnes

Acknowledgments

===============

- [VCU Natural Language Processing Lab](https://nlp.cs.vcu.edu/)     ![alt text](https://nlp.cs.vcu.edu/images/vcu_head_logo "VCU")

- [Nanoinformatics Vertically Integrated Projects](https://rampages.us/nanoinformatics/)