Diabetes Prediction Web App: Multi-Algorithm Comparison and Complete Machine Learning Workflow Practice

This open-source project focuses on diabetes prediction, using the Pima Indians dataset to compare four machine learning algorithms (Logistic Regression, KNN, Decision Tree, Random Forest). It covers the complete workflow of data preprocessing, model training, and evaluation, and builds an interactive web dashboard via Streamlit, providing engineering practice references for machine learning beginners.

The project uses the classic Pima Indians Diabetes Dataset (available from the UCI Machine Learning Repository or Kaggle), containing 768 records, 8 medical features (number of pregnancies, glucose concentration, blood pressure, skin fold thickness, insulin level, BMI, diabetes pedigree function, age), and the target variable Outcome (0 = no diabetes / 1 = has diabetes). Each feature is related to diabetes risk—for example, glucose concentration is a core diagnostic indicator, and BMI reflects obesity risk.

The project demonstrates the complete workflow from raw data to deployment:

1. Data Preprocessing: Treat zero values of indicators like Glucose and BloodPressure as missing values and convert them to NaN, fill with category mean (calculate mean separately by Outcome category), and standardize features using StandardScaler;
2. Exploratory Data Analysis (EDA): Analyze category distribution, detect outliers, draw correlation heatmaps, and calculate descriptive statistics.

The project compares four classic classification algorithms:

• Logistic Regression (baseline model): Accuracy 75.97%;
• K-Nearest Neighbors: Accuracy 85.71%;
• Decision Tree: Test set accuracy 89.61% (highest);
• Random Forest: Accuracy 86.36%, cross-validation shows better stability.

Evaluation uses multi-dimensional metrics: accuracy, precision, recall, F1-Score, confusion matrix, and K-fold cross-validation (Random Forest cross-validation score: 0.8763); hyperparameter tuning is performed for Random Forest, optimizing parameters via grid/random search.

The project uses Streamlit to build an interactive prediction dashboard:

• Streamlit advantages: Build interfaces with pure Python code, real-time interactive components, fast deployment;
• Features: Users input medical indicators to get real-time diabetes risk prediction results, realizing the transformation from model to application.

Learning Value: Covers the full machine learning lifecycle, understanding of multi-algorithm comparisons, application of evaluation metrics, engineering code organization, and model-to-application capabilities; Improvement Directions: Expand dataset size, optimize class imbalance, enhance feature engineering, try advanced algorithms like XGBoost/neural networks, and need clinical experts to validate prediction results.