Diabetes Prediction: Comparing Random Forest & Logistic Regression

📌 Project Overview

This project compares the performance of Random Forest (RF) and Logistic Regression (LR) on the Pima Indians Diabetes Dataset to predict the presence of diabetes in women aged 21+ of Pima Indian heritage. The goal is to evaluate model accuracy, computational efficiency, and robustness in handling imbalanced data. Results are benchmarked against the study by Chang et al. (2022).

Key Questions Addressed:

• Which model (RF or LR) performs better in terms of accuracy, AUC, and computational speed?
• How do feature importance and correlation impact predictions?
• Does hyperparameter tuning significantly improve performance?

🏥 Dataset Details

• Source: Kaggle
• Samples: 768 women (8 predictors, 1 binary target)
• Features: Pregnancies, Glucose, Blood Pressure, Skin Thickness, Insulin, BMI, Diabetes Pedigree Function, Age
• Preprocessing:
  • Replaced missing values (encoded as 0) with mean/median.
  • Normalized predictors using min-max scaling.
  • Split into 60:20:20 (train/validation/test).

🛠️ Methodology

1. Data Cleaning & EDA:
   • Handled missing values (mean for Glucose/BP, median for Skin Thickness/Insulin/BMI).
   • Analyzed correlations (e.g., Glucose and BMI strongly linked to diabetes).
2. Model Training:
   • Random Forest: Hyperparameter tuning (trees=150, max splits=40, leaf size=30) via 10-fold CV.
   • Logistic Regression: Grid search for lambda (best=0.001).
3. Evaluation Metrics:
   • Accuracy, AUC, Precision, Recall, F1-Score.

📊 Results

Model Performance Comparison (Test Set)

Metric	Random Forest	Logistic Regression
Accuracy	78%	77%
AUC	0.82	0.75
Precision	0.74	0.68
Recall	0.65	0.72
F1-Score	0.69	0.70
Training Speed	Slow	Fast

Key Findings:

• RF Strengths: Higher AUC (0.82), better precision.
• LR Strengths: Faster training, better recall.
• Both models struggled with class imbalance (more "no diabetes" predictions).
• RF overfits slightly on training data but generalizes well on test data.
• Contradiction to Hypothesis: RF had lower TP/FP rates than expected.

📸 Key Visualizations

Figure 1: Frequency Distribution Before & After Imputation

Top: Original dataset with missing values (encoded as 0).
Bottom: After replacing missing values with mean/median.

Figure 2: Correlation Heatmap

Glucose, BMI, and Age show strong positive correlations with diabetes.

Figure 3: Feature Importance (Random Forest vs. Logistic Regression)

RF: Glucose and BMI are top predictors.

LR: Blood Pressure and Diabetes Pedigree Function dominate.

Figure 4: Class Distribution

Imbalanced dataset: ~65% "No Diabetes" vs. ~35% "Diabetes".

Figure 5: ROC Curves

RF AUC = 0.82: Better class separation.
LR AUC = 0.75: Moderate performance.

📹 Presentation Video

• Download the video here

📝 Lessons Learned & Future Work

Lessons:

• Feature Selection: Critical for LR (e.g., excluding negatively correlated features improved accuracy).
• Hyperparameter Tuning: RF requires careful tuning (OOB error may outperform CV for small datasets).
• Class Imbalance: Addressing imbalance (e.g., SMOTE) could improve recall.

Future Work:

• Experiment with feature engineering and advanced techniques (XGBoost, SVM).
• Compare OOB error vs. cross-validation for hyperparameter tuning.
• Expand dataset size and apply anomaly detection for outliers.

---

🔗 References

• [Chang, V., Bailey, J., Xu, Q.A. et al. Pima Indians diabetes mellitus classification based on machine learning (ML) algorithms. Neural Comput & Applic 35, 16157–16173 (2023). https://doi.org/10.1007/s00521-022-07049-z]
• Full references in docs/presentation.pptx