Heart Failure Prediction using Explainable AI (Decision Trees)

Author: Jesús Jiménez Serrano
Field: Machine Learning / Health Informatics
Model: Decision Tree Classifier (White-Box)

1. Project Overview

This project addresses Heart Failure (HF) prediction through the lens of Explainable AI (XAI). Unlike traditional "black-box" models, this study utilizes a Decision Tree with limited depth. This ensures that medical professionals can audit and understand the clinical logic behind every diagnosis, fostering trust in AI-assisted systems.

Why this approach?

• Interpretability: We prioritize clinical transparency over pure mathematical optimization.
• Traceability: Decision rules emulate the standard medical triage process.
• Beyond Linear Models: Captures complex, non-linear patterns that traditional models (like the SHFM) often oversimplify.

2. Dataset

The project uses the Heart Failure Prediction Dataset from Kaggle (https://www.kaggle.com/datasets/fedesoriano/heart-failure-prediction), which unifies several clinical cohorts.

• Size: 918 clinical records.
• Key Variables: Age, Cholesterol, Resting Blood Pressure, Exercise ECG (ST_Slope, Oldpeak), among others.

3. Data Preprocessing

• Data Cleaning: Median imputation was applied to the Cholesterol variable (correcting 18.73% of null/zero values) to preserve statistical power.
• Encoding: Label Encoding was applied to binary variables, and One-Hot Encoding to multi-class variables (such as Chest Pain Type) to avoid artificial hierarchical ordering.
• Data Split: 80% Training / 20% Validation.

4. Results

• Final Accuracy: 83.15% on the test set.
• Primary Predictor: The model identified the ST segment slope (ST_Slope) as the root node, confirming its high clinical significance.
• Safety Metrics: Deep analysis of the Confusion Matrix was performed to minimize False Negatives, which is critical in a healthcare environment.

5. Installation and Usage

1. Clone the repository:

   git clone https://github.com/jesusjimenez04/heart-failure-prediction-ai.git
   

2. Install dependencies: pip install -r requirements.txt

3. Run the notebook HF_prediction.ipynb in your Jupyter environment or VS Code.