About This Project
Built and compared 7 supervised ML models to classify breast cancer using routine blood biomarkers and anthropometric data (age, BMI) from the Breast Cancer Coimbra Dataset (116 women; 64 cancer, 52 control). Models evaluated: Naive Bayes, LDA, KNN, Random Forest, Gradient Boosting, SVM, and Deep Neural Network. Applied z-normalization, factor encoding, stratified train-test splits, and repeated 10-fold cross-validation for robust evaluation. Performed hyperparameter tuning including KNN-k sweep and RF mtry optimization. Key insight: Glucose, Resistin, and Adiponectin were top predictors — showing non-invasive tests can assist early cancer detection.
Key Features
- Compared 7 models: Naive Bayes, LDA, KNN, RF, GBM, SVM, DNN
- Best result: Random Forest — Accuracy 87%, F1 0.88, AUC 0.91
- Sensitivity 85%, Specificity 90% — strong clinical balance
- Repeated 10-fold cross-validation with stratified splits
- Hyperparameter tuning: KNN-k sweep, RF mtry optimization
- Top predictors: Glucose, Resistin, Adiponectin
