EXQUAIRO EXECUTIVE FOUNDATION · FOR UNIVERSITY MEDICAL CENTRES LEADERS
Heart Failure Case
You are about to work through the same case our data scientists use — on roughly 900 cardiology patients, with the same models they run in practice. Not a demonstration: the analysis actually runs while you sit there, on your own device.
Twelve steps, about ninety minutes. Nothing to install. Choose how close to the machinery you want to sit — and switch at any moment.
What you will do
- 1
Meet the data
One row per patient: about 900 clinic visits, eleven measurements, and a final column saying whether heart disease was found.
- 2
Data quality
Some values are missing. Others are present but almost certainly wrong — and that is worse.
- 3
How the patients are spread out
Before comparing groups, look at one measurement at a time.
- 4
Which patients turn out to be ill?
Compare one measurement between the patients with heart disease and those without.
- 5
Measurements that move together
When two columns say almost the same thing, adding both adds cost but not knowledge.
- 6
Getting the data ready for a model
Cleaning, translating categories into numbers, adding two new features, and holding patients back for the exam.
- 7
The first model: a decision tree
A flowchart of yes/no questions, learned from the data. You can read this one end to end.
- 8
Many trees together: the random forest
Hundreds of imperfect trees, each shown a slightly different slice of the data, voting on every patient.
- 9
Boosting: XGBoost and LightGBM
Instead of voting, each new tree focuses on the patients the previous trees got wrong.
- 10
Why did the model say that about this patient?
SHAP values split a single prediction into the contribution of each measurement.
- 11
Can we trust this number?
One exam is one exam. Cross-validation repeats it several times on different patients.
- 12
Your position
You have now done the whole cycle: data, quality, exploration, preparation, models, explanation, validation.