Diabetes Risk Prediction

Early-stage diabetes risk classification using clinical health indicators — deployed as a live web application.

---

Problem Statement

Diabetes affects over 77 million people in India alone, and a large fraction of cases go undiagnosed until serious complications emerge. Early detection using routinely available clinical data can change outcomes dramatically. This project builds a binary classification model to predict diabetes risk from health survey indicators, optimized to minimize missed cases — because in healthcare, a false negative (missing a true diabetic) is far more costly than a false positive.

---

Pipeline Overview

Raw Clinical Data
    │
    ▼
Data Preprocessing
    ├── Missing value handling
    ├── Feature scaling (StandardScaler)
    └── Class imbalance treatment (balanced class weights)
    │
    ▼
Model Training
    └── Random Forest Classifier
    │
    ▼
Threshold Optimization
    └── Lowered classification threshold → improved recall
    │
    ▼
Evaluation
    ├── ROC-AUC Score
    ├── Precision / Recall / F1
    └── Confusion Matrix
    │
    ▼
Streamlit Web App → hackathon-c.streamlit.app

---

Dataset

Feature	Description
Size	~100,000 rows
Target	Binary: Diabetic / Non-diabetic
Features	Clinical indicators (BMI, age, blood pressure, glucose, etc.)
Challenge	Class imbalance — diabetic cases are a minority class

---

Tech Stack

Category	Tools
Language	Python 3.9+
ML	scikit-learn · Random Forest
Data	pandas · NumPy
Visualization	matplotlib · seaborn
Deployment	Streamlit

---

Key Design Decisions

Handling Class Imbalance
The dataset has significantly more non-diabetic cases than diabetic ones. Using class_weight='balanced' in Random Forest adjusts the model to treat minority class errors as more costly — pushing it toward better recall on diabetic cases.

Threshold Optimization
By default, a classifier predicts "positive" when probability ≥ 0.5. Lowering this threshold (e.g., to 0.35–0.40) increases sensitivity — the model flags more people as potentially diabetic, at the cost of slightly more false positives. For a screening tool, this tradeoff is the right one.

Why Random Forest?
Random Forest handles mixed feature types well, is robust to outliers, and provides feature importances — making it interpretable enough for healthcare contexts where understanding why a prediction is made matters.

---

Results

Metric	Score
ROC-AUC	0.95801
Precision	0.88
F1 Score	0.80

---

Getting Started

hackathon-c.streamlit.app

---

Repository Structure

diabetes-prediction/
├── app.py                # Streamlit web application
├── requirements.txt      # Dependencies
├── diab_pred.py          # EDA and model training
├── models/               # Saved model (.pkl)
└── data/                 # Dataset files

---