Classification Models Comparison: Diabetes Prediction
Kaedyn Cruickshank
Like this project
Posted Aug 28, 2025
Compared models for diabetes prediction using Pima Indians dataset.
Classification Models Comparison: Diabetes Prediction
Log Reg, Random Forest, Gradient Boosting
Purpose
To predict diabetes risk from recorded health metrics such as glucose, BMI, insulin, etc. and compare Logistic Regression vs tree-based models among other observations.
Dataset
Pima Indians Diabetes dataset (768 rows, 9 columns). Target:
Outcome (1 = diabetes, 0 = no diabetes).Public, well-known educational dataset.
Process
Cleaning
Replaced impossible zeros in table fields (
Glucose, BloodPressure, SkinThickness, Insulin, BMI) using median imputation.Models
Logistic Regression (with StandardScaler in pipeline)
Random Forest (300 trees)
Gradient Boosting
Model Evaluation
Metrics: Accuracy, Precision, Recall, F1, ROC-AUC
Visuals: Confusion Matrix per model, ROC Curves
Interpretability: Feature importance
Results — Test Set
Logistic Regression: Accuracy=0.760, Precision=0.698, Recall=0.556, F1Score=0.619, ROC_AUC=0.813
Random Forest: Accuracy=0.747, Precision=0.653, Recall=0.593, F1Score=0.621, ROC_AUC=0.839
Gradient Boosting: Accuracy=0.753, Precision=0.660, Recall=0.611, F1Score=0.635, ROC_AUC=0.829
Best by ROC-AUC: Random Forest
Feature Importance (For Tree Models)
Random Forest and Gradient Boosting (the tree-based models) provide feature importance scores.
Generally, Glucose, BMI, and Age have most feature importance. Some differences are present as the gradient boosting model has a surprising importance score of approx. 0.4 for glucose compared to just 0.25 for the random forest model.