🩺 Diabetes Prediction using Classification Algorithms

This project focuses on predicting whether a person is diabetic or not using medical data, with a special focus on dealing with missing values, outliers, imbalanced data, and choosing the right classification model.

---

📌 Objective

To predict the likelihood of diabetes using various features such as glucose level, BMI, age, blood pressure, etc., from the Pima Indians Diabetes Dataset. I explored multiple machine learning classification models and techniques to handle data quality issues and imbalances.

---

📊 Dataset Overview

The dataset contains information on female patients including:

• Pregnancies
• Glucose
• Blood Pressure
• Skin Thickness
• Insulin
• BMI
• Diabetes Pedigree Function
• Age
• Outcome (0 = Non-Diabetic, 1 = Diabetic)

---

🧹 Data Cleaning & Imputation

I began by identifying missing values that were stored as 0s, particularly in columns like Glucose, Insulin, BMI, etc.

• If the distribution was symmetric (no outliers), I used the mean for imputation.
• If the distribution was skewed (had outliers), I used the median.
• For categorical features, I would use the mode, though in this dataset all features were numerical.

I used Seaborn’s displot and boxplots to determine symmetry and outliers:

• displot helped visualize the shape of distributions.
• KDE lines indicated skewness and helped guide the imputation strategy.

df['Glucose'] = df['Glucose'].replace(0, df['Glucose'].mean())
df['SkinThickness'] = df['SkinThickness'].replace(0, df['SkinThickness'].median())
...

---

📦 Outlier Detection and Removal

I applied Quantile-based filtering to remove outliers, particularly in the Insulin column which had high variance.

Instead of just removing rows from x_scaled, I made sure to apply the same mask to both features and labels to preserve alignment:

q = x_scaled['Insulin'].quantile(0.95)
mask = x_scaled['Insulin'] < q
dataNew = x_scaled[mask]
y_outlier_detection = y_outlier_detection[mask]

---

⚖️ Handling Imbalanced Data

The dataset was imbalanced with more non-diabetic cases. I tried:

Oversampling (to increase samples in minority class)

Undersampling (to reduce samples from majority class)

Then I used SMOTE (Synthetic Minority Oversampling Technique) to generate synthetic examples for the minority class.

🔴 Note: SMOTE was applied before normalization to maintain true geometry of data.

---

🚀 Models Used

I applied and compared several classification models:

1. Logistic Regression

   Good baseline model

   Predicts probability of class using a sigmoid function

   Suitable since the target variable is binary (0/1)

2. Gaussian Naive Bayes

   Assumes features follow Gaussian distribution

   Calculates P(class | features) based on Bayes' Theorem

   Simple yet powerful probabilistic model

   Works well with continuous features like age, glucose, etc.

3. K-Nearest Neighbors (KNN)

   Non-parametric (doesn’t assume any distribution)

   Classifies a point based on majority of K neighbors

   Needs feature scaling

   Performance was good after handling outliers and imbalances

📈 KNN Model Performance (after SMOTE)

          precision    recall  f1-score   support

       0       0.85      0.67      0.75       159
       1       0.54      0.76      0.63        79

accuracy                           0.70       238

macro avg 0.69 0.72 0.69 238 weighted avg 0.75 0.70 0.71 238

Class 0 (Non-Diabetic): High precision, moderate recall

Class 1 (Diabetic): Lower precision, but better recall

Good balance in F1-score

Overall accuracy: 70%

---

🧠 Key Takeaways

• Preprocessing matters just as much as the model!
• Imputation method depends on distribution shape (mean vs median)
• Removing outliers carefully prevents misaligned labels
• SMOTE helps a lot with imbalanced datasets
• KNN and Naive Bayes performed reasonably well
• Logistic regression is still a good baseline model