Embedded Files
Classification is a supervised machine learning technique that uses labeled outcomes to learn patterns and make predictions. Unlike clustering, which identifies groups based on feature similarity without labels, classification models are trained on known outcomes (e.g., diabetic or non-diabetic) to predict future cases. In this project, classification was used to predict the likelihood of diabetes in patients based on health-related features such as glucose, insulin, and BMI, using the Outcome variable as the target label.
Classification is a supervised machine learning technique that uses labeled outcomes to learn patterns and make predictions. Unlike clustering, which identifies groups based on feature similarity without labels, classification models are trained on known outcomes (e.g., diabetic or non-diabetic) to predict future cases. In this project, classification was used to predict the likelihood of diabetes in patients based on health-related features such as glucose, insulin, and BMI, using the Outcome variable as the target label.
Objectives of Classification in This Project
Objectives of Classification in This Project
Predict diabetes status (diabetic or non-diabetic) based on patient health metrics
Identify key features (e.g., glucose, BMI, insulin) most strongly associated with diabetes
Build a reliable supervised model that supports early diagnosis and clinical decision-making
Enable personalized healthcare planning by accurately classifying high-risk individuals
Data Splitting
Data Splitting
We use a 67-33 train-test split because, after testing multiple ratios including 90-10, 80-20, and 67-33, we found that the 67-33 split yielded the highest accuracy across most models. It provided a good balance between giving the model enough data to learn and enough unseen data to evaluate its generalization performance effectively.
We use a 67-33 train-test split because, after testing multiple ratios including 90-10, 80-20, and 67-33, we found that the 67-33 split yielded the highest accuracy across most models. It provided a good balance between giving the model enough data to learn and enough unseen data to evaluate its generalization performance effectively.
KNN as baseline model
KNN as baseline model
Overall Accuracy: 84.38%
Overall Accuracy: 84.38%
Model Type: Supervised classification using KNN with k = 13
Model Type: Supervised classification using KNN with k = 13
The model performs very well for non-diabetic cases, with high precision and recall.
For diabetic patients, performance is slightly lower, indicating a modest trade-off in sensitivity.
Overall, the model is balanced and accurate, but may benefit from further tuning or alternative algorithms for better recall on diabetic cases.
Comparison with other models
Comparison with other models
Seven classification models were evaluated to predict diabetes. Among them, Random Forest delivered the highest performance with an accuracy of 88.29% and an AUC of 94.65, making it the top-performing model. AdaBoost followed closely, showing comparable recall and the highest AUC (94.67), despite slightly lower specificity.
Seven classification models were evaluated to predict diabetes. Among them, Random Forest delivered the highest performance with an accuracy of 88.29% and an AUC of 94.65, making it the top-performing model. AdaBoost followed closely, showing comparable recall and the highest AUC (94.67), despite slightly lower specificity.
Support Vector Machine (SVM) also performed well, achieving 85.67% accuracy and AUC of 91.19, leading all non-ensemble models. K-Nearest Neighbors (KNN) maintained balanced performance with 84.38% accuracy and good recall, making it a reliable baseline. Simpler models like Logistic Regression and GaussianNB offered high recall but fell behind in specificity and overall AUC, highlighting the superior effectiveness of ensemble methods for this task.
Support Vector Machine (SVM) also performed well, achieving 85.67% accuracy and AUC of 91.19, leading all non-ensemble models. K-Nearest Neighbors (KNN) maintained balanced performance with 84.38% accuracy and good recall, making it a reliable baseline. Simpler models like Logistic Regression and GaussianNB offered high recall but fell behind in specificity and overall AUC, highlighting the superior effectiveness of ensemble methods for this task.
Top 3 Models for Hyperparameter Tuning
Top 3 Models for Hyperparameter Tuning
Optimization Techniques Used:
Optimization Techniques Used:
GridSearchCV for exhaustive hyperparameter tuning
GridSearchCV for exhaustive hyperparameter tuning
Cross-validation to ensure model stability and generalization
Cross-validation to ensure model stability and generalization
Feature importance analysis to focus on key predictors (e.g., Glucose, Insulin)
Feature importance analysis to focus on key predictors (e.g., Glucose, Insulin)
After tuning the top three models—Random Forest, AdaBoost, and K-Nearest Neighbors (KNN)—all showed strong, balanced classification performance.
After tuning the top three models—Random Forest, AdaBoost, and K-Nearest Neighbors (KNN)—all showed strong, balanced classification performance.
- Random Forest achieved the highest accuracy (88.67%), best recall (0.92), and strong overall balance across all metrics, making it the most reliable and consistent model.
- AdaBoost had the highest AUC (95.11%) and slightly better specificity, but slightly lower recall than Random Forest.
- KNN stood out with the highest precision, but had the lowest accuracy and AUC, indicating weaker overall performance compared to the ensemble methods.
Selected Model
Selected Model
Based on the optimization results, the Random Forest Classifier was selected as the final model for deployment. It consistently delivered the best trade-off between accuracy, recall, and robustness, making it the most suitable choice for real-world diabetes prediction.
Page updated
Google Sites
Report abuse