This project focuses on building a binary classification system to predict whether a patient has diabetes or not using machine learning models, while properly handling imbalanced data and deploying the final model using Streamlit.

• Problem Type: Binary Classification
• Target Variable: Diabetes (0 = No, 1 = Yes)
• Main Challenges:
  • Imbalanced dataset
  • Model selection & comparison
  • Proper preprocessing
  • Deployment

The project follows a clean and modular ML pipeline:

• Data cleaning using FunctionTransformer
• Handling missing values using SimpleImputer
• Feature scaling with RobustScaler
• Feature transformation using PowerTransformer

To solve class imbalance, SMOTE-based techniques were applied:

• SMOTE
• SMOTETomek

These techniques were applied only on the training data to avoid data leakage.

The following models were trained and compared:

• Logistic Regression
• K-Nearest Neighbors (KNN)
• Decision Tree
• Random Forest
• Support Vector Machine (SVM)
• Naive Bayes (Gaussian)
• XGBoost

Each model was evaluated using multiple metrics.

Because the dataset is imbalanced, accuracy alone is not enough.
The following metrics were used:

• Accuracy (Train & Test)
• Precision
• Recall
• F1-score
• ROC-AUC

The F1-score was the primary metric for model selection.

• Technique: GridSearchCV
• Cross Validation: 5-Fold
• Scoring Metric: F1-score

Random Forest hyperparameters tuned include:

• n_estimators
• max_depth
• min_samples_split
• min_samples_leaf
• max_features

• Model: Random Forest Classifier
• Why?
  • Best balance between precision & recall
  • High F1-score
  • Stable performance after SMOTE

The trained artifacts were saved using pickle:

• rf_model.pkl → trained model
• preprocessor.pkl → preprocessing pipeline

This allows:

• Reusing the model without retraining
• Easy deployment

A simple and interactive Streamlit web app was created where users can:

• Input patient medical data
• Get real-time diabetes prediction
• View prediction confidence

streamlit run app.py