PyTrial API & Pipeline

Table of Contents

• PyTrial API & Pipeline

As described in Intro 1: Overview of PyTrial, PyTrial maintains a consistent user interface, including common functions like fit, predict, save_model, and load_model. Each task is defined by its input and output data.

Therefore, the general pipeline is as follows, we take a patient-level outcome prediction task as an example:

1. Prepare the input data for the task.

from pytrial.data.patient_data import TabularPatientBase
from pytrial.utils.tabular_utils import MinMaxScaler
from pytrial.utils.tabular_utils import read_csv_to_df

# Read the data
df = read_csv_to_df('./tabular_patient_outcome_data.csv', index_col=0)
label = df['target_label']
df = df.drop(['target_label'], axis=1)

# Build the dataset for the specific task
# Here, we are working on individual patient level outcome prediction taking tabular inputs.
dataset = TabularPatientBase(df,
    metadata={
        'transformers':
            {'age': MinMaxScaler()}, # specify the data transformation for specific columns
    })

2. Import the models from the corresponding pytrial.tasks module.

from pytrial.tasks.indiv_outcome.tabular import LogisticRegression

model = LogisticRegression()

3. Train the model using the fit function.

model.fit(
    {
    'x': dataset,
    'y': label,
    }
)

4. Make the prediction using the predict function. And save the model using the save_model function.

# make predictions
ypred = model.predict({'x': dataset})

# save the model
model.save_model('./model')

We can see that, except for the first step for data preparation, the rest of the steps are rather straightforward. For the sake of supporting the data preparation, we provide a set of basic dataset classes in the pytrial.data module. We also provide a set of children classes of them for the specific tasks, e.g., pytrial.tasks.trial_patient_match.data.PatientData and pytrial.tasks.trial_patient_match.data.TrialData considering the trial-patient matching task.

We will go through each task with concrete examples in the next chapters.