Pediatric Cardiac Disease Prediction

Machine learning pipeline for predicting cardiac abnormalities in children, built on a real-world clinical dataset (UCMF) from a Portuguese pediatric cardiology unit.

Built as a final project for the Fundamentals of Data Science course at Universidade de Aveiro (2024/2025).

---

Overview

The goal of this project is to predict whether a child presents a normal or abnormal cardiac condition based on clinical measurements. The full pipeline covers everything from raw data exploration to predictive modelling, producing 50+ outputs including statistical reports, visualisations, and model evaluations.

---

Dataset

The UCMF dataset contains anonymised pediatric cardiology records with the following features:

Feature	Description
IDADE	Patient age (0-19 years)
SEXO	Sex
IMC	Body Mass Index
PA SISTOLICA	Systolic blood pressure
PA DIASTOLICA	Diastolic blood pressure
FC	Heart rate
PULSOS	Pulse assessment
SOPRO	Heart murmur
B2	Second heart sound
NORMAL X ANORMAL	Target variable (Normal / Abnormal)

---

Pipeline

1. Raw Data Analysis — raw_data_analysis.py

Exploratory analysis of the raw dataset before any preprocessing.

Key outputs: variable classification, missing value analysis, physiologically impossible values, descriptive statistics, histograms, boxplots, KDE density curves, categorical distributions.

2. Data Preprocessing — pre_process.py

Cleaning and preparation for downstream analysis.

Transformations applied:

• Duplicate removal (ID column)
• Age validation (0-19 years)
• Weight, height and BMI cleaning
• Systolic/diastolic blood pressure and heart rate validation
• Categorical standardisation (SEXO, TARGET, PULSOS, SOPRO, B2)
• Irrelevant column removal

Output: UCMF_CLEAN.xlsx

3. Exploratory Data Analysis — eda_basic.py

Detailed descriptive analysis of the cleaned dataset: statistics summary, histograms, boxplots, dispersion measures, normality tests (Shapiro-Wilk, Kolmogorov-Smirnov).

4. Bivariate Analysis — bivariate_analysis.py

Analysis of pairwise variable relationships: Pearson, Spearman and Kendall correlations, simple linear regressions, t-tests comparing Normal vs Abnormal groups, ANOVA for categorical variables, partial correlations controlling for age.

5. Multivariate Analysis — multivariate_analysis.py

Complex multi-variable relationships: logistic regression with Odds Ratios, Mutual Information scores, PCA (scree plot, biplot, loadings), K-Means clustering with Silhouette and Davies-Bouldin evaluation.

6. Feature Selection — feature_selection.py

Identification of the most relevant predictive variables using: F-test, Mutual Information, Random Forest importance, Gradient Boosting importance, Recursive Feature Elimination (RFE), cross-validation performance curve.

7. Predictive Modelling — predictive_models.py

Training and evaluation of 5 classification models:

Model	Description
Decision Tree	Interpretable tree-based classifier
SVM	Support Vector Machine with RBF kernel
Naive Bayes	Gaussian probabilistic classifier
Random Forest	Ensemble of decision trees
Neural Network	Multi-layer Perceptron (MLP)

Evaluation metrics: Accuracy, Precision, Recall, F1-score, ROC-AUC, confusion matrices, 5-fold cross-validation.

---

How to run

Prerequisites

pip install pandas numpy matplotlib seaborn scikit-learn openpyxl xlrd

Run the full pipeline in order

python src/raw_data_analysis.py
python src/pre_process.py
python src/eda_basic.py
python src/bivariate_analysis.py
python src/multivariate_analysis.py
python src/feature_selection.py
python src/predictive_models.py

Each script reads from the previous step's output. Start with raw_data_analysis.py and run sequentially.

---

Tech stack