python电子病历交接班系统_Med7：临床电子病历可迁移自然语言处理模型

Med7

This repository dedicated to the first release of Med7: a transferable clinical natural language processing model for electronic health records, compatible with spaCy, for clinical named-entity recognition (NER) tasks. The en_core_med7_lg model is trained on MIMIC-III free-text electronic health records and is able to recognise 7 categories:

The trained model comprises three components in its pipeline:

tagger

parser

clinical NER with seven categories.

Self-supervised pre-training has shown its efficiency in achieving good results even with a small number of gold-annotated training data. We have experimented with the spacy pretrain approach and trained a number of weights for model initialisation for various parameters of the width and depth of convolutional layers. Following the notations of spaCy pretrain with --width, --depth, --embed-rows flags for width, depth and the number of embedding rows respectively:

--width

--depth

--embed-rows

model size (MB)

epochs

URL

96

4

10000

3.8

350

128

8

10000

18.3

596

256

8

10000

47.6

450

256

16

10000

66.1

332

300

8

20000

89.6

338

The models were pre-trained on the entire MIMIC-III data, comprising a collection of 2,083,054 documents with the total of 3,129,334,419 words. Models' losses (logarithmically scaled) are presented below:

The model achieved a lenient (strict) micro-averaged F1 score of 0.957 (0.893) across all seven categories.

Installation

It is recommended to create a dedicated virtual environment and install all recent required packages in there. The trained model was tested with spaCy version 2.2.3 and Python 3.7. For example, if the anaconda distribution of Python is already installed:

create a new virtual environment:

(base) conda create -n med7 python=3.7

activate and install spaCy:

(base) conda activate med7

(med7) pip install -U spacy

once all went through smoothly, install the Med7 model:

(med) pip install https://med7.s3.eu-west-2.amazonaws.com/en_core_med7_lg.tar.gz

Usage

import spacy

med7 = spacy.load("en_core_med7_lg")

# create distinct colours for labels

col_dict = {}

seven_colours = ['#e6194B', '#3cb44b', '#ffe119', '#ffd8b1', '#f58231', '#f032e6', '#42d4f4']

for label, colour in zip(med7.pipe_labels['ner'], seven_colours):

col_dict[label] = colour

options = {'ents': med7.pipe_labels['ner'], 'colors':col_dict}

text = 'A patient was prescribed Magnesium hydroxide 400mg/5ml suspension PO of total 30ml bid for the next 5 days.'

doc = med7(text)

spacy.displacy.render(doc, style='ent', jupyter=True, options=options)

[(ent.text, ent.label_) for ent in doc.ents]

The Med7 model identifies correctly all seven entities in the following example and highlights them in different colours for better visualisation:

and the resulting output:

[('Magnesium hydroxide', 'DRUG'),

('400mg/5ml', 'STRENGTH'),

('suspension', 'FORM'),

('PO', 'ROUTE'),

('30ml', 'DOSAGE'),

('bid', 'FREQUENCY'),

('for the next 5 days', 'DURATION')]

It is straightforward to extract relations between the entities, since Med7 has both parser and tagger pipelines, similar to this example.

The code in above can also be run in Colab

Citing

This model is the very first step in our programme on clinical NLP for electronic health records (cNLPEHR). We are committed to developing FAIR - Findable, Accessible, Interoperable and Reusable tools which will benefit the wider community.

If you found this model useful, please acknowledge by citing as:

@article{kormilitzin2020med7,

title={Med7: a transferable clinical natural language processing model for electronic health records},

author={Kormilitzin, Andrey and Vaci, Nemanja and Liu, Qiang and Nevado-Holgado, Alejo},

journal={arXiv preprint arXiv:2003.01271},

year={2020}

}