In this example we use Veritable to predict which patients have heart disease. The dataset for this example is standard, part of the UC Irvine machine learning repository (http://archive.ics.uci.edu/ml/datasets/Heart+Disease).

This is a classification problem. We predict the target column, a measure of the presence of heart disease in the patient. This column has five possible values, and it's unclear from the data dictionary whether they should be understood as ordinal (measures of the severity of heart disease) or as categorical (different forms of heart disease). All the published studies have focused on the binary classification problem of distinguishing absence (value 0) from presence of heart disease (values 1, 2, 3, and 4). This script runs both the binary and multiway classification problems.

• heart_disease/run.py: A Python script that divides the data into train and test sets, runs an analysis on the training set, and makes predictions for the target column on the test set.
• heart_disease/data.json: A json file containing the preprocessed data from the hungarian.data, long-beach-va.data, and switzerland.data files.
• heart_disease/schema.json: A json file containing a schema with column types for the data.
• original_data/: A directory containing the original data files; note that the cleveland.data file is excluded, as this file is reportedly corrupt.
• original_data/preprocess.py: A Python script that converts the original data file format into Veritable-ready data and schema files; also produces a csv data file for convenience.
• original_data/column_info.py: A Python file containing the column names; used by preprocess.py to create data.json and schema.json.

The combined.data file contains the concatenation of the three original dataset files that are included in the original_data directory.

• 617 rows
• 75 columns
• some missing values
• all column types: boolean, categorical, real, and count
• target column is target: whether the patient developed heart disease

$ python heart_disease/run.py

multinomial dataset, raw predictions: 41.9% test error
multinomial dataset, binary transform: 21.0% test error
binary dataset, raw predictions: 17.7% test error