Machine Learning Prediction by Streamlit : https://multiclassdiabetesdata-machinelearning-prediction.streamlit.app/

🧪 Multiclass Diabetes Classification

📌 Project Overview

This project aims to build a multiclass classification model to predict diabetes categories based on clinical and laboratory features. The dataset contains demographic, metabolic, and kidney-related biomarkers.

The objective is not only to detect diabetes, but also to analyze feature interactions that may help differentiate metabolic severity.

---

🔎 1. Data Cleaning

In a medical context, Total Cholesterol cannot be 0 (normally >100 mg/dL or >2 mmol/L).

A value of 0 likely indicates:

• Input error
• Laboratory result not recorded
• Missing value incorrectly stored as 0

Therefore, cholesterol values equal to 0 were treated as invalid entries and handled appropriately during preprocessing.

---

📊 2. Target Class Distribution

The dataset contains three classes with the following proportions:

Class	Proportion
2	47.89%
0	36.78%
1	15.33%

The dataset is moderately imbalanced, particularly for Class 1, which motivates the use of SMOTE in the modeling stage.

---

⚙️ 3. Feature Engineering

To enhance biological interpretability and capture interactions, new features were created:

🧪 New Features

• Urea_Cr_Ratio Clinically relevant ratio used to assess kidney function.

• BMI_HbA1c Interaction between obesity (BMI) and blood sugar control (HbA1c). Helps capture metabolic risk patterns.

• AGE_BMI Combines age and obesity, both important diabetes risk factors.

These engineered features significantly improved class separability.

---

📈 4. Feature Selection

🔹 ANOVA Test

Top significant features based on F-Score:

Feature	F-Score	p-value
BMI_HbA1c	263.26	4.95e-63
HbA1c	195.42	2.15e-52
BMI	184.78	1.58e-50
AGE_BMI	171.47	4.26e-48
AGE	58.39	1.20e-21

Features like HDL and LDL showed low significance.

🔹 Chi-Square Test

Feature	Score	p-value
BMI_HbA1c	7.73e+06	4.25e-229
HbA1c	1.88e+04	1.82e-115

Both tests confirm that HbA1c and BMI_HbA1c are extremely strong predictors.

⭐ Feature Importance Insight

• HbA1c indicates whether someone has diabetes, but does not fully differentiate metabolic severity.

• BMI_HbA1c (interaction feature) reveals:

  • High BMI + High HbA1c → likely Type 2 diabetes (insulin resistance)
  • Low BMI + High HbA1c → possible Type 1 diabetes

This interaction provides better clinical interpretability.

---

🤖 5. Model Comparison

Algorithm	ROC AUC	Accuracy	STD
Random Forest	99.50	97.12	2.79
Gradient Boosting	99.31	98.07	2.36
Decision Tree	98.10	97.60	2.16
SVM	96.06	87.03	4.66
Logistic Regression	92.43	87.02	5.36
KNN	89.79	79.85	6.23

🏆 Best Model: Random Forest

Random Forest achieved the highest ROC AUC with strong stability.

---

⚖️ 6. Handling Class Imbalance (SMOTE)

Oversampling was performed after train-test split and during cross-validation.

Important principle:

• SMOTE applied only to training data
• Validation/test data kept in original distribution

This prevents data leakage and ensures realistic model evaluation.

---

🔧 7. Hyperparameter Tuning

Best Parameters:

Best Parameters: {'n_estimators': 200}
Best CV Accuracy: 0.9766666666666666

Tuning improved model generalization and stability.

---

📉 8. Overfitting Check

🔹 Gap (Train – Test)

• Accuracy Gap: 1.89%
• ROC AUC Gap: 0.34%

The small gap indicates very low overfitting and strong generalization performance.

---

📊 9. Final Model & ROC Curve

The final selected model is Random Forest with 200 trees.

The ROC Curve demonstrates excellent class separability with near-perfect AUC values.

---

🎯 Conclusion

• Engineered interaction features significantly improved performance.
• HbA1c alone is powerful, but combining it with BMI enhances predictive power.
• Random Forest provides robust performance with minimal overfitting.
• Proper SMOTE implementation ensured fair validation.

This project highlights how combining clinical domain knowledge + feature engineering + proper validation strategy leads to highly accurate multiclass classification models.

---

🚀 Deployment Ready

The multiclass diabetes classification model has been fully trained, validated, and evaluated.

After performing:

• Data cleaning
• Feature engineering
• Feature selection (ANOVA & Chi-Square)
• Class balancing using SMOTE
• Hyperparameter tuning
• Overfitting validation

The final Random Forest model demonstrates strong generalization performance and stable metrics.

• ✅ The multiclass dataset pipeline is production-ready
• ✅ The trained model is ready for deployment
• ✅ Successfully deployed using Streamlit for interactive prediction

The Streamlit application allows users to:

• Input clinical parameters (Age, BMI, HbA1c, Cholesterol, etc.)
• Automatically generate engineered features
• Predict diabetes class in real-time
• Display prediction probabilities

This ensures the model is not only accurate in experimentation, but also practical for real-world usage.

---

📚 Concepts Covered

• 🧾 Multiclass Classification
• 🧪 Clinical Data Analysis
• 🧹 Data Cleaning & Missing Value Handling
• ⚙️ Feature Engineering (Biological Feature Interaction)
• 📊 Statistical Feature Selection (ANOVA & Chi-Square)
• ⚖️ Handling Imbalanced Data (SMOTE)
• 🤖 Ensemble Learning (Random Forest, Gradient Boosting)
• 🔧 Hyperparameter Tuning
• 📈 ROC AUC Evaluation & Model Comparison
• 🚀 Model Deployment with Streamlit