Diabetes Prediction Model and Web App

Project Overview

This repository contains a Diabetes Prediction Model and a Streamlit Web App designed to predict diabetes risk using a well-known health dataset. The project combines exploratory data analysis (EDA), machine learning model comparison, and a user-friendly web interface to deliver actionable insights and predictions. It demonstrates skills in data preprocessing, classification, visualization, and web deployment—ideal for showcasing data science proficiency.

Dataset:

Source: Diabetes Prediction Dataset on Kaggle

Size: 768 rows, 9 columns

Features:
-Age: Age of the individual (years)

-Gender: Gender of the individual (binary)

-Body Mass Index (BMI): Measure of body fat based on height and weight

-Blood Pressure: Diastolic blood pressure (mmHg)

-Insulin Level: 2-hour serum insulin (mu U/ml)

-Skin Thickness: Triceps skin fold thickness (mm)

-Glucose Level: Plasma glucose concentration (mg/dl)

-Diabetes Pedigree Function: Genetic diabetes risk score

-Number of Pregnancies: Number of times pregnant

-Target: Outcome (0 = No Diabetes, 1 = Diabetes)

Purpose: These features capture physiological and genetic factors critical for predicting diabetes likelihood.

Workflow:

Exploratory Data Analysis (EDA):

-Inspected data structure, distributions, and class balance using pandas, seaborn, and matplotlib.

Key findings:

-Glucose and BMI show significant variance; dataset is imbalanced (more non-diabetic cases).

Preprocessing:

-Standardized column names (str.lower()).

-Scaled features with RobustScaler to handle outliers.

-Split data into 70% training and 30% testing sets (train_test_split).

Modeling:

Tested five classifiers from sklearn:
-Random Forest (n_estimators=200, max_depth=5)

-K-Nearest Neighbors (n_neighbors=7)

-Support Vector Machine (SVM)

-Logistic Regression (max_iter=200)

-Gradient Boosting (n_estimators=50, max_depth=5)

-Evaluated with accuracy, precision, recall, and F1-score due to class imbalance.

Visualization:

-Distribution plots for numerical features.

-Count plot for outcome balance.

-Bar chart comparing model performance with accuracy labels.

Model Deployment:

-Saved the best-performing model (Random Forest) using pickle for integration into the web app.

Web App:

-Built an interactive Streamlit application to allow users to input health data and receive diabetes risk predictions.

Live Demo:

Diabetes Web App

Key Insights:

Feature Importance:

• Random Forest identified Glucose Level and BMI as top predictors of diabetes risk.

Model Performance:

• Random Forest achieved ~75% accuracy, with balanced precision/recall (F1 ~0.70), outperforming others due to its robustness to imbalanced data.

EDA Highlights:

-Higher glucose levels and BMI correlate strongly with positive diabetes outcomes (correlation heatmap analysis).

Web App Features:

-Input Fields: Users enter values for all 9 features (e.g., Age, BMI, Glucose Level).

Prediction:

• Displays diabetes risk (Yes/No) based on the Random Forest model.

Accessibility:

-Hosted on Streamlit Cloud for easy access—no local setup required.

Results

-Best Model: Random Forest (Accuracy: ~75%, F1: ~0.70).

Impact: The web app enables non-technical users (e.g., healthcare providers) to assess diabetes risk instantly.