Kidney Stone Prediction

#Project overview

I used a Python-based Kaggle kernel to run a full data analysis and predictive modelling workflow for a machine learning competition. The goal was to predict a binary target from a set of physiological features using an ensemble algorithm, CatBoost.

#Tools and libraries

• Python for data manipulation and modelling
• Pandas & NumPy for data processing and linear algebra
• scikit-learn for scaling and model evaluation
• Seaborn & Matplotlib for visual insight into the data
• CatBoost, an ensemble algorithm that is robust on categorical data

#Data handling

• Loading: imported the train and test sets within the Kaggle environment and explored their size and shape.
• Exploration: studied distributions and relationships using correlation matrices and plots.
• Visualisation: heatmaps to explore correlations between features, and scatter plots to compare key feature pairs across conditions.

#Feature engineering

I created ratios and products of existing features to uncover interaction effects, giving the model more nuanced information than the raw features alone.

#Modelling

• Preprocessing: normalised the feature set so every feature carried equal weight.
• Training: CatBoostRegressor with stratified k-fold cross-validation to make predictions more robust and reduce overfitting.
• Evaluation: ROC AUC, a single metric that captures how well the model separates the classes.

#Results

Training was refined iteratively to optimise ROC AUC across the validation folds, and I compared CatBoost against an XGBoost classifier. Final predictions on the test set were formatted into a competition submission, scoring 0.756 ROC AUC on the private leaderboard.

#What it shows

• Managing and preprocessing data efficiently with industry-standard Python libraries
• Using ensemble learning for robust predictive modelling
• Analysing and visualising data to uncover underlying patterns
• Optimising and evaluating models with cross-validation and the right metrics