Stroke prediction

This repository contains an imbalanced supervised binary classification task: predicting patients that will have a stroke given sociological and biological factors, using the Kaggle dataset present in "https://www.kaggle.com/datasets/fedesoriano/stroke-prediction-dataset".

Tools used

The main code is scattered across three notebooks:

1 - Exploratory data analysis

In the notebook 1-eda.ipynb, I perform exploratory data analysis on the data:

• Check for missing values and data types
• Draw the blueprint for the data pipeline
• Perform univariate and bivariate data analysis

2 - Model selection and tuning

This notebook, 2-model.ipynb, selects a model across many different classifiers and tunes the best selected classifiers using cross-validation.

The following approach is used:

• Creating a data pipeline
• Selecting the best models using cross-validation
• Performing cross-validaition hyperparameter tuning on the best models using the optuna package
• Saving the best model pipelines for later evaluation

3 - Model evaluation

Notebook 3-eval.ipynb evaluates the tuned models from the previous notebook and benchmarks them across various different metrics on the test set.

The evaluation consists of the following steps:

• Accuracy, ROC AUC and F1 score
• Confusion matrix
• ROC curve
• Precision Recall curve
• True vs predicted distributions
• Threshold tuning using F1-score