Using our medaCy Trained Model

Once a medaCy model has been trained - it can be saved, distributed, and used for prediction.
To predict with a trained learning model, it is required that documents be run through a pipeline identical to that which the model was trained with. Fortunately, medaCy encapsulates this functionality via the Model class.

Loading a trained model for prediction

Once a CRF model has been trained and saved to disk, it can be loaded again for use by configuring a Model object with the pipeline used to train the CRF model. The below example shows how to configure the medaCy clinical model.

from medacy.pipelines import ClinicalPipeline
from medacy.model.model import Model

pipeline = ClinicalPipeline(metamap=None, entities=['Drug'])
model = Model(pipeline)
model.load('/home/medacy/trained_model.pickle')

annotation = model.predict("The patient took 5 mg of aspirin.")  # Returns an Annotations object

Model prediction over a string returns a medaCy Annotations object.
Useful functionalities are provided in the Annotation class such as the ability to see a diff between
two annotations for empirical analysis.

Trained CRF models are pickled (serialized) binary files.

medaCy Model Management

One of medaCy's most powerful features is the ability to maintain, version and distribute medaCy compatible models with ease. The idea is simple - all the set-up code for a Model including a trained machine learning model is abstracted into an outside installable python package. This allows one to maintain the model with a version history just like any piece of software.

Once a model has been packaged and installed it can be used as follows:

from medacy.model.model import Model

model = Model.load_external('medacy_model_clinical_notes')
annotations = model.predict("The patient took 5 mg of aspirin.")

See Packaging a medaCy Model for information on how to distribute your own trained models either internally or to the world.