Diabetes Prediction

A beginner binary-classification project on the Pima Indians diabetes dataset. It predicts diabetes from eight clinical measurements with logistic regression. The point of interest is that several columns use 0 to mean "missing".

This project is educational only. It is not a medical diagnostic tool.

Dataset

Download Kaggle's Pima Indians Diabetes Database and place diabetes.csv at:

data/raw/diabetes.csv

With the Kaggle CLI:

kaggle datasets download -d uciml/pima-indians-diabetes-database -p data/raw --unzip

The file has 768 rows. In Glucose, BloodPressure, SkinThickness, Insulin, and BMI a value of 0 is physically impossible and really means the reading is missing, so those zeros are converted to missing values and imputed with the median. About 35% of patients are diabetic.

Workflow

1. Load the raw CSV and remove exact duplicate rows.
2. Replace impossible zeros with missing values in the five affected columns.
3. Make a stratified 80/20 train/test split.
4. Impute missing values with the median and standardize - inside the pipeline, so the median is computed on training data only.
5. Train logistic regression.
6. Compare against a majority-class dummy model.
7. Evaluate with accuracy, precision, recall, F1, a confusion matrix, and five-fold stratified cross-validation.

Results

Measurement	Result
Dummy test accuracy	64.9%
Logistic regression test accuracy	70.8%
Test recall (diabetics found)	0.50
CV mean accuracy	78.8%

Accuracy beats the baseline, but recall of 0.50 means the model misses half the diabetic patients - the metric that matters most in a screening context. Raising recall (class weighting or a lower decision threshold) is the natural next experiment. Note the single test split (0.71) reads lower than cross-validation (0.79), so the CV number is the more reliable estimate.

Run it

pip install -r requirements.txt
python main.py

The confusion matrix is written to reports/figures/confusion_matrix.png.

Layout

data/raw/         downloaded dataset (git-ignored - see Dataset above)
src/preprocess.py loading, cleaning, zero-as-missing, imputing scaler
src/train.py      pipeline, training loop, evaluation, figure
src/evaluate.py   classification metric helpers
main.py           entry point

Why this project

Inspired by classmate medical-classification projects (e.g. PranavSingla122's diagnostic notebooks, adityaxdubey's health models). It sits next to the heart-disease classifier but adds the zero-as-missing cleaning step and median imputation.