This project implements the MedCAT NLP application as a service behind a REST API. The general idea is to be able send the text to MedCAT NLP service and receive back the annotations. The REST API is built using Flask.
Git Branches:
- devel: development branch, latest updates and features, might be unstable.
- master: stable releases
- legacy: branch for old MedCAT version (pre v1.0, new models wont work, only v0.x models will)
Feel free to ask questions on the github issue tracker or on our discourse website which is frequently used by our development team!
The API definition follows the one defined in CogStack GATE NLP Service. Currently, there are 3 endpoints defined, that consume and return data in JSON format:
-
GET
/api/info- displays general information about the MedCAT application, -
POST
/api/process- processes the provided documents and returns back the annotations, -
POST
/api/process_bulk- processes the provided list of documents and returns back the annotations.
The full specification is available is OpenAPI specification.
The application can be run either as a standalone Python application or as running inside the Docker container (recommended).
Please note that prior running the application a number of requirements need to installed (see: requirements.txt).
There are two scripts provided implementing starting the application:
-
start-service-dev.sh- starts the application in the development mode and usingwerkzeugserver for serving Flask applications, -
start-service-prod.sh- starts the application in 'production' mode and usinggunicornserver.
The recommended way to run the application is to use the provided Docker image. The Docker image can be either downloaded from the Docker Hub (cogstacksystems/medcat-service:latest) or build manually using the provided Dockerfile.
Please note that by default the built docker image will run the Flask application in 'production' mode running start-service-prod.sh script.
To build the Docker image manually:
docker build -t medcat-service .
To run the container using the built image:
docker run -it -p 5000:5000 \
--env-file=envs/env_app --env-file=envs/env_medcat \
-v <models-local-dir>:/cat/models:ro \
cogstacksystems/medcat-service:latest
By default the MedCAT service will be running on port 5000. MedCAT models will be mounted from local directory <models-local-dir> into the container at /cat/models.
If you have a gpu and wish to use it, please change the docker/docker-compose.yml file, use the cogstacksystems/medcat-service-gpu:latest image or change the build: directive to build ../Dockerfile_gpu.
If you wish to run this docker service manually, use the docker/docker-compose.yml file, execute docker-compose up -d whilst in the docker folder.
Alternatively, an example script ./docker/run_example_medmen.sh was provided to run the Docker container with MedCAT service. The script will download an example model (using the ./models/download_medmen.sh script),it will use an example environment configuration, then it will build and start the service using the provided Docker Compose file, the service WONT WORK without the model being present.
All models should be mounted from the models/ folder.
1. cd ./models/
2. bash ./download_medmen.sh
3. cd ../docker/
4. docker-compose up -d
DONE!
Or, if you wish to use the above mentioned script ( the sample model is downloaded via script, you don't need to do anything):
1. cd ./docker/
2. bash ./run_example_medmen.sh
DONE!
Assuming that the application is running on the localhost with the API exposed on port 5000, one can run:
curl -XPOST http://localhost:5000/api/process \
-H 'Content-Type: application/json' \
-d '{"content":{"text":"The patient was diagnosed with leukemia."}}'
and the received result:
{
"result": {"text": "The patient was diagnosed with leukemia.",
"annotations": {"entities": {"0": {"pretty_name": "Patients", "cui": "C0030705", "type_ids": ["T101"], "types": ["Patient or Disabled Group"], "source_value": "patient", "detected_name": "patient", "acc": 0.99, "context_similarity": 0.99, "start": 4, "end": 11, "icd10": [], "ontologies": [], "snomed": [], "id": 0, "meta_anns": {"Status": {"value": "Affirmed", "confidence": 0.9999303817749023, "name": "Status"}}}, "1": {"pretty_name": "Diagnosis", "cui": "C0011900", "type_ids": ["T060"], "types": ["Diagnostic Procedure"], "source_value": "diagnosed", "detected_name": "diagnosed", "acc": 0.6657139492748229, "context_similarity": 0.6657139492748229, "start": 16, "end": 25, "icd10": [], "ontologies": [], "snomed": [], "id": 1, "meta_anns": {"Status": {"value": "Affirmed", "confidence": 0.9999918341636658, "name": "Status"}}}, "2": {"pretty_name": "leukemia", "cui": "C0023418", "type_ids": ["T191"], "types": ["Neoplastic Process"], "source_value": "leukemia", "detected_name": "leukemia", "acc": 0.2572544372951888, "context_similarity": 0.2572544372951888, "start": 31, "end": 39, "icd10": [], "ontologies": [], "snomed": [], "id": 2, "meta_anns": {"Status": {"value": "Affirmed", "confidence": 0.9999804496765137, "name": "Status"}}}}, "tokens": []},
"success": true,
"timestamp": "2021-11-11T11:54:28.856+00:00"
},
"medcat_info": {"service_app_name": "MedCAT", "service_language": "en", "service_version": "1.2.5", "service_model": "MedMen"}
}
process_bulk example :
curl -XPOST http://localhost:5000/api/process_bulk \
-H 'Content-Type: application/json' \
-d '{"content": [{"text":"The patient was diagnosed with leukemia."}, {"text": "The patient was diagnosed with cancer."}] }'
example bulk result :
{
"result": [
{
"text": "The patient was diagnosed with leukemia.",
"annotations": {
"0": {
"pretty_name": "Patients",
"cui": "C0030705",
"type_ids": [
"T101"
],
"types": [
"Patient or Disabled Group"
],
"source_value": "patient",
"detected_name": "patient",
"acc": 0.99,
"context_similarity": 0.99,
"start": 4,
"end": 11,
"id": 0,
"meta_anns": {
"Status": {
"value": "Affirmed",
"confidence": 0.9999303817749023,
"name": "Status"
}
}
},
"1": {
"pretty_name": "Diagnosis",
"cui": "C0011900",
"type_ids": [
"T060"
],
"types": [
"Diagnostic Procedure"
],
"source_value": "diagnosed",
"detected_name": "diagnosed",
"acc": 0.6657139492748229,
"context_similarity": 0.6657139492748229,
"start": 16,
"end": 25,
"id": 1,
"meta_anns": {
"Status": {
"value": "Affirmed",
"confidence": 0.9999918341636658,
"name": "Status"
}
}
},
"2": {
"pretty_name": "leukemia",
"cui": "C0023418",
"type_ids": [
"T191"
],
"types": [
"Neoplastic Process"
],
"source_value": "leukemia",
"detected_name": "leukemia",
"acc": 0.2572544372951888,
"context_similarity": 0.2572544372951888,
"start": 31,
"end": 39,
"id": 2,
"meta_anns": {
"Status": {
"value": "Affirmed",
"confidence": 0.9999804496765137,
"name": "Status"
}
}
}
},
"success": true,
"timestamp": "2021-12-08T18:49:55.255+00:00"
},
{
"text": "The patient was diagnosed with cancer.",
"annotations": {
"0": {
"pretty_name": "Patients",
"cui": "C0030705",
"type_ids": [
"T101"
],
"types": [
"Patient or Disabled Group"
],
"source_value": "patient",
"detected_name": "patient",
"acc": 0.99,
"context_similarity": 0.99,
"start": 4,
"end": 11,
"id": 0,
"meta_anns": {
"Status": {
"value": "Affirmed",
"confidence": 0.9999236464500427,
"name": "Status"
}
}
},
"2": {
"pretty_name": "cancer diagnosis",
"cui": "C0920688",
"type_ids": [
"T060"
],
"types": [
"Diagnostic Procedure"
],
"source_value": "diagnosed with cancer",
"detected_name": "diagnosed~with~cancer",
"acc": 1,
"context_similarity": 1,
"start": 16,
"end": 37,
"id": 2,
"meta_anns": {
"Status": {
"value": "Affirmed",
"confidence": 0.9999957084655762,
"name": "Status"
}
}
}
},
"success": true,
"timestamp": "2021-12-08T18:49:55.255+00:00"
}
],
"medcat_info": {
"service_app_name": "MedCAT",
"service_language": "en",
"service_version": "1.2.6",
"service_model": "MedMen"
}
}
IMPORTANT info regarding annotation output style
As the changes from MedCAT intoduced dictionary annotation/entity output.
The mode in which annotation entities should be outputted in the JSON response, by default this was outputted as a "list" of dicts in older versions, so the output would be :
{"annotations": [{"id": "0", "cui" : "C1X..", ..}, {"id":"1", "cui": "...."}]}
newer versions of MedCAT (1.2+) output entities as a dict, where the id of the entity is a key and the rest of the data is a value, so for "dict", the output is
{"annotations": [{"0": {"cui": "C0027361", "id": 0,.....}, "1": {"cui": "C001111", "id": 1......}]}
This setting can be configured in the ./envs/env_medcat file, using the ANNOTATIONS_ENTITY_OUTPUT_MODE variable.
By default, the output of these entities is set to respect the output of the MedCAT package, hence the latter will be used. Please change the above mentioned env variable and make sure your CogStack-Nifi annotation script is adapted accordingly.
Please note that the returned NLP annotations will depend on the underlying model used. For evaluation, we can only provide a very basic model trained on MedMentions. Models utilising SNOMED CT or UMLS may require applying for licenses from the copyright holders.
In the current implementation, configuration for both MedCAT Service application and MedCAT NLP library is based on environment variables. These will be provided usually in two files in env directory:
-
env_app- configuration of MedCAT Service app, -
env_medcat- configuration of MedCAT library.
Both files allow tailoring MedCAT for specific use-cases. When running MedCAT Service, these variables need to be loaded into the current working environment.
When using MedCAT for a different language than English, it can be useful to use a different spaCy model. A spaCy model can be included in the MedCAT model pack, but when not using this functionality, it can be useful to install models in the Docker image. This can be done by setting a build-time variable. See the SPACY_MODELS variable in Dockerfile for default value and usage.
MedCAT Service application are defined in envs/env_app file.
The following environment variables are available for tailoring the MedCAT Service gunicorn server:
-
SERVER_HOST- specifies the host address (default:0.0.0.0), -
SERVER_PORT- the port number used (default:5000), -
SERVER_WORKERS- the number of workers serving the Flask app working in parallel (default:1; only used in production server). -
SERVER_WORKER_TIMEOUT- the max timeout (in sec) for receiving response from worker (default:300; only used with production server).
The following environment variables are available for tailoring the MedCAT Service wrapper:
-
APP_MODEL_NAME- an informative name of the model used by MedCAT (optional), -
APP_MODEL_CDB_PATH- the path to the model's concept database, -
APP_MODEL_VOCAB_PATH- the path to the model's vocabulary, -
APP_MODEL_META_PATH_LIST- the list of paths to meta-annotation models, each separated by:character (optional), -
APP_BULK_NPROC- the number of threads used in bulk processing (default:8), -
APP_TRAINING_MODE- whether to run the application with MedCAT in training mode (default:False). -
APP_MEDCAT_MODEL_PACK- MedCAT Model Pack path, if this parameter has a value IT WILL BE LOADED FIRST OVER EVERYTHING ELSE (CDB, Vocab, MetaCATs, etc.) declared above.
Theres a range of factors that might impact the performance of this service, the most obvious being the size of the processed documents (amount of text per document) as well as the resources of the machine on which the service operates.
The main settings that can be used to improve the performance when querying large amounts of documents are : SERVER_WORKERS (number of flask web workers that chan handle parallel requests) and APP_BULK_NPROC (threads for annotation processing).
MedCAT parameters are defined in selected envs/env_medcat* file.
For details on available MedCAT parameters please refer to the official GitHub repository.
