A machine learning pipeline that predicts heart disease from CDC behavioral survey data, achieving 83.5% ROC-AUC and 75.7% F1-score using logistic regression with SMOTE oversampling. Built on 319K+ patient records from the 2020 BRFSS survey, then applied to 445K unseen records from the 2022 survey.
Six classifiers were evaluated using 5-fold stratified cross-validation with SMOTE applied inside each fold to prevent data leakage. Logistic regression narrowly outperformed gradient boosting and XGBoost on F1-Macro, the lead metric chosen for its sensitivity to the minority class.
Final validation metrics (tuned logistic regression, 20% hold-out set):
| Metric | Score |
|---|---|
| ROC-AUC | 0.835 |
| F1-Macro | 0.757 |
| Accuracy | 0.757 |
| Precision (Yes) | 0.75 |
| Recall (Yes) | 0.77 |
Logistic regression coefficients reveal the strongest predictors after scaling. The chart below shows the top 20 features by absolute coefficient magnitude — pink bars increase heart-disease risk, blue bars decrease it.
Takeaways for clinicians:
- Age dominates — the 80+ coefficient is 5× larger than any medical condition
- Stroke history and poor self-reported health are the next-strongest signals
- Kidney disease and diabetes show elevated risk but smaller model weight than expected, likely due to correlation with age
- Being in the 25–29 age bracket is the strongest protective factor
The precision-recall curve (AP = 0.818) shows the model maintains >80% precision up to ~60% recall — meaning it can correctly flag 6 in 10 actual heart-disease cases while keeping false alarms manageable.
Source: CDC BRFSS 2020 via Kaggle
| Training (2020) | Test (2022) | |
|---|---|---|
| Records | 319,795 | 445,132 |
| Features | 18 | 40 → aligned to 17 |
| Target | HeartDisease (Yes/No) | Predicted |
| Class balance | 91.4% No / 8.6% Yes | — |
Features span three domains:
- Medical history — diabetes, kidney disease, stroke, asthma, skin cancer
- Demographics — age, sex, race
- Lifestyle — smoking, alcohol, physical activity, sleep, BMI, self-rated health
Full per-variable EDA with distributions, outlier detection, and heart-disease rate breakdowns is documented in the notebook.
Raw Data (319K records, 18 columns)
│
├─ Quality Audit ──── 0 nulls, 18K duplicates flagged, IQR outlier detection
│
├─ EDA ────────────── Per-variable distributions, target-rate analysis,
│ correlation heatmap, before/after rebalancing metrics
│
├─ Preprocessing ───── Median/mode imputation → one-hot encoding →
│ StandardScaler (fit on train only) → feature alignment
│
├─ Class Balancing ─── Random undersampling to 50/50 for training;
│ SMOTE inside CV folds for evaluation
│
├─ Model Selection ─── 6 classifiers × 5-fold CV × 3 metrics
│ (F1-Macro, Accuracy, ROC-AUC)
│
├─ Tuning ──────────── RandomizedSearchCV on best model
│ (C=0.1, penalty=L2, solver=liblinear)
│
└─ Prediction ──────── 445K test records → 31,089 predicted positive (7.0%)
Preventative-Pulse/
├── README.md
├── Preventative_Pulse.ipynb # Full pipeline: EDA → modeling → results
├── requirements.txt
├── .gitignore
├── data/
│ ├── heart_2020_cleaned.csv # Training data (2020 BRFSS, 320K rows)
│ └── heart_2022_with_nans.csv # Test data — download separately (see below)
└── figures/ # Auto-generated by notebook
git clone https://github.com/yourusername/Preventative-Pulse.git
cd Preventative-Pulse
pip install -r requirements.txtThe 2022 test dataset (139 MB) is too large for GitHub. Download it from Kaggle and place the file in data/heart_2022_with_nans.csv. Sections 1–8 of the notebook run without it; only Section 9 (test-set predictions) requires it.
jupyter notebook Preventative_Pulse.ipynbAll figures are saved to figures/ on execution.
Python · pandas · NumPy · scikit-learn · imbalanced-learn (SMOTE) · XGBoost · Matplotlib · Seaborn
Dataset published by Kamil Pytlak under CC BY-SA 4.0. Code in this repository is MIT licensed.