Diabetes Classification Pipeline
LightGBM + Optuna classification reaching 90.07% accuracy with SHAP explainability.
Model accuracy
Uplift vs baseline
Cross-validation folds
Optuna trials
Overview
An end-to-end binary classification pipeline that predicts diabetes risk. The workflow covers preprocessing, feature engineering, LightGBM modelling, Optuna hyperparameter tuning and SHAP interpretability — pushing accuracy roughly 14 percentage points above the baseline.
Features
- End-to-end pipeline: preprocessing → model → evaluation
- Feature engineering to boost signal
- LightGBM model tuned with Optuna
- Cross-validated model selection to avoid overfitting
- SHAP interpretability for per-feature impact
Challenges
- Lifting accuracy ~14 pp above the baseline model.
- Preventing overfitting during aggressive hyperparameter search.
- Making the model's decisions explainable with SHAP.
Results
Engineering the right features and tuning LightGBM with Optuna delivered 90.07% accuracy — about a 14 percentage-point improvement over the baseline — with SHAP confirming which features drove predictions.
Screenshots
Architecture
The end-to-end flow from raw data to a shipped result.
Preprocess
Clean & encode the dataset
Engineer
Derive predictive features
Tune
LightGBM + Optuna, CV-validated
Explain
SHAP feature attribution