🧠 Diabetes Risk Prediction & Random Projection Analysis 📌 Overview

This project implements a machine learning pipeline for diabetes prediction combined with Random Projection techniques to analyze dimensionality reduction trade-offs. It evaluates multiple models and optimizes performance using GPU-accelerated XGBoost.

📊 Dataset Samples: 70,692 Features: 22 → 23 (after feature engineering) Balanced dataset: Diabetic: 35,346 Non-diabetic: 35,346 Duplicates removed: 1,635 ⚙️ Key Features Advanced feature engineering (HealthScore, RiskFactorCount) Outlier detection & capping Random Projection (Gaussian, Sparse, Achlioptas) GPU-based training using CUDA + XGBoost Statistical validation + SHAP interpretability 🚀 Model Performance 🔹 Baseline Models Model Accuracy Precision Recall F1 ROC-AUC Logistic Regression 0.7456 0.7384 0.7735 0.7555 0.8182 Random Forest 0.7436 0.7256 0.7967 0.7595 0.8192 XGBoost 0.7498 0.7317 0.8016 0.7651 0.8235

✅ Best Model: XGBoost

🔹 Final Tuned Model Accuracy: 0.7488 Precision: 0.7307 Recall: 0.8009 F1-Score: 0.7642 ROC-AUC: 0.8246 ⚡ Random Projection Results 📉 Dimensionality Reduction Impact Original features: 23 Reduced features: 15 (−34.8%) Memory reduction: 34.8% 🔹 XGBoost Comparison Metric Original Projected Accuracy 0.7498 0.7333 ROC-AUC 0.8235 0.8026 Precision 0.7317 0.7151 Recall 0.8016 0.7900

👉 Only ~1–2% performance drop with significant dimensionality reduction.

🔹 Classification Trade-off (Projection) x=5 → Accuracy: 0.6734 x=15 → Accuracy: 0.7215 x=20 → Accuracy: 0.7457 (≈ original) 🔹 K-Means Cost Reduction Cost ratio improved from 2.56 → 1.69 as dimension increased ⚡ Performance Optimization XGBoost training time: 0.41s Logistic Regression: 1.08s Random Forest: 1.23s 🔧 Hyperparameter Tuning Best CV Score: 0.8266 Strategy: Fine-Tuned Grid Search GPU acceleration used (Tesla T4)