Cardiovascular Disease Prediction

Cleaning noisy health data into an interpretable baseline for cardiovascular-disease classification.

An end-to-end data-science project that explores, cleans, and models the Kaggle Cardiovascular Disease Dataset. The central question is simple: can demographic, clinical, anthropometric, and lifestyle variables help predict the binary cardio label?

Paper · Exploration · Cleaning · Baseline model

---

The finding, before the code

The raw dataset contains implausible heights, weights, and blood-pressure readings—including impossible pressure values and inconsistent systolic/diastolic pairs. After conservative cleaning and the addition of BMI and pulse pressure, a stratified logistic-regression baseline improves from 0.763 ROC-AUC to 0.795 ROC-AUC.

Evaluation on held-out data	Raw data	Cleaned data
Test observations	14,000	13,713
Accuracy	0.70	0.727
ROC-AUC	0.763	0.795
Recall — cardio = 1	0.67	0.67
F1 — cardio = 1	0.69	0.71

The lesson is not that the model is ready for care. It is that data quality changes what a model can learn: removing clear measurement errors makes the signal in blood pressure and related features substantially more interpretable.

---

What this project does

The work follows one connected analytical path:

flowchart LR
    A["Raw cardiovascular data<br/>70,000 records · 13 columns"] --> B["EDA<br/>distributions · target patterns · outliers"]
    B --> C["Cleaning & feature engineering<br/>valid ranges · age in years · BMI · pulse pressure"]
    C --> D["Stratified 80/20 split<br/>random_state = 42"]
    D --> E["StandardScaler<br/>fit on training data only"]
    E --> F["Logistic Regression<br/>interpretable baseline"]
    F --> G["Held-out evaluation<br/>accuracy · ROC-AUC · precision · recall · F1"]

Loading

Stage	Question answered	Main artifact
Exploratory analysis	What is in the data, and what looks implausible?	EDA.ipynb
Cleaning	Which rows and feature transformations make the data analysable?	DataCleaning.ipynb
Baseline modelling	How well does an interpretable classifier generalise?	MachineLearning.ipynb
Written report	What do the results mean—and where do they stop?	cardiovascular_disease_paper.pdf

---

The data

The source is the public Cardiovascular Disease Dataset on Kaggle. It contains 70,000 observations and a nearly balanced binary target: cardio = 0 indicates no recorded cardiovascular disease; cardio = 1 indicates its presence.

Feature family	Variables
Demographic	age (days in raw data), gender
Anthropometric	height, weight
Clinical	ap_hi (systolic BP), ap_lo (diastolic BP), cholesterol, gluc
Lifestyle	smoke, alco, active
Target	cardio

The original id is removed before modelling. The ordinal clinical categories use the dataset’s supplied encoding: 1 = normal, 2 = above normal, and 3 = well above normal.

What EDA revealed

• No missing values, but meaningful quality problems: implausible body measurements, impossible blood-pressure values, and pressure pairs where ap_lo > ap_hi.
• Cardiovascular-disease prevalence rises markedly with age in this dataset.
• Higher cholesterol and glucose categories are associated with a greater observed share of cardio = 1.
• Smoking, alcohol use, and physical activity have much weaker marginal relationships in this sample; that does not establish an absence of causal effect.

---

Cleaning & feature engineering

The cleaning rules are deliberately conservative: remove clear measurement errors while retaining plausible pathological values.

Operation	Rule / result
Age	Convert days to age_years for interpretability
Identifier	Drop id—it has no clinical meaning
Height	Keep 140–210 cm
Weight	Keep 30–200 kg
BMI	Create BMI = weight / height²; keep 15–60
Systolic pressure	Keep ap_hi between 80 and 250 mmHg
Diastolic pressure	Keep ap_lo between 40 and 200 mmHg
Pressure consistency	Correct inconsistent pairs before deriving features
Derived signal	Create pulse pressure: pp = ap_hi - ap_lo

The resulting clean dataset contains 68,562 rows and 13 model features plus the target.

The project is transparent about cleaning choices because they are part of the model: they affect the distributions, correlations, and evaluation that follow.

---

Modelling

Why logistic regression? The target is binary, and logistic regression gives an interpretable, reproducible baseline before considering more complex models.

Decision	Implementation
Split	80% training / 20% test, stratified by cardio
Reproducibility	random_state=42
Scaling	StandardScaler fit on training data only
Classifier	LogisticRegression(max_iter=1000, solver="lbfgs") with L2 regularisation
Evaluation	Accuracy, ROC-AUC, precision, recall, F1, and confusion matrix

On the cleaned test split, the classifier correctly separates the classes moderately well (ROC-AUC 0.795), but it identifies only 67% of positive cases at the evaluated threshold. That performance is unsuitable for medical decision-making.

Signals in the fitted baseline

Because features are standardised, coefficient magnitudes are comparable within this linear model. The strongest positive coefficients are associated with systolic pressure, diastolic pressure, age, pulse pressure, and cholesterol. These are model associations, not clinical causation or validated feature importance—especially because pulse pressure is derived from the two blood-pressure variables.

---

Repository map

.
├── data/
│   ├── raw/cardio_train.csv             # original dataset
│   └── clean/cardio_cleaned.csv         # cleaned, feature-engineered data
├── notebooks/
│   ├── EDA.ipynb                        # distributions, relationships, outliers
│   ├── DataCleaning.ipynb               # validation rules and transformations
│   └── MachineLearning.ipynb            # baseline training and evaluation
├── cardiovascular_disease_paper.pdf     # full methods and results report
└── Phasen.md                            # project plan and hypotheses

---

Quickstart

Use a Python environment with Jupyter and the core analysis libraries:

git clone https://github.com/theMindDeveloper/Cardiovascular_Disease_Prediction_Data_Science.git
cd Cardiovascular_Disease_Prediction_Data_Science

python -m venv .venv
# macOS / Linux
source .venv/bin/activate
# Windows PowerShell
# .\.venv\Scripts\Activate.ps1

pip install pandas numpy matplotlib seaborn scikit-learn jupyter
jupyter lab

Run the notebooks in this order:

1. notebooks/EDA.ipynb
2. notebooks/DataCleaning.ipynb
3. notebooks/MachineLearning.ipynb

The modelling notebook expects the cleaned file at data/clean/cardio_cleaned.csv. Re-running the cleaning notebook produces the analytical input used by the baseline.

---

Important scope and limitations

This is an educational data-science project, not a clinical risk calculator.

• The Kaggle dataset is incompletely documented and may not represent a broader population.
• The target is a simplified binary label; it does not describe disease type, severity, progression, or diagnosis.
• The sample has an imbalanced gender-code distribution and relies on coarse, self-reported lifestyle variables.
• Results have not been externally validated, calibrated for clinical use, or evaluated as a medical decision rule.
• A false negative could be harmful in real healthcare; the observed positive-class recall (~0.67) makes clinical deployment inappropriate.
• Observational correlations and model coefficients do not prove causality.

For the full methodological discussion, data-quality rationale, and references, read the project paper.

---

Built as a coursework project in exploratory analysis, data cleaning, feature engineering, and supervised machine learning.

If you use this work, please keep the educational and non-clinical scope visible.