Diabetes Prediction System
A Machine Learning system for predicting diabetes risk based on medical records, featuring in-depth Exploratory Data Analysis.
Project Overview
The Challenge
Raw medical datasets are often noisy and contain illogical values—such as zero for blood pressure or glucose levels—which can severely mislead a machine learning model. Building a reliable predictive system requires a robust method to handle these data anomalies before training.
The Solution
This project emphasizes rigorous Exploratory Data Analysis (EDA) to uncover and address data quality issues. K-Means clustering was employed to intelligently impute invalid zero-values, replacing them with more realistic estimates based on patient groupings. This cleaned and engineered data was then used to train and evaluate multiple models, including L1-Regularized Regression and Voting Ensembles, resulting in a more accurate and trustworthy prediction system.
Key Features
- EDA-driven cleaning of invalid medical records.
- K-Means clustering for intelligent data imputation.
- Feature engineering and binning for better generalization.
- High F1-score and ROC-AUC achieved across models.
- Interactive prediction interface built with Streamlit.