Healthcare Analytics Project

Predicting Diabetes Risk Using Machine Learning

---

Overview

This project analyzes patient health metrics to predict diabetes risk using a modular, production-style machine learning pipeline.
The workflow includes:

• Data preprocessing and cleaning
• Exploratory data analysis (EDA)
• Model training using a recall-optimized Random Forest
• Standardized model evaluation
• Optional command-line prediction generation

The objective is to demonstrate an analytics engineering approach that is reproducible, interpretable, and clinically meaningful, especially for use cases where reducing false negatives matters.

---

Dataset

Source: Public diabetes dataset inspired by the Pima Indians dataset
Description: Includes features such as glucose, BMI, blood pressure, insulin, age, diabetes pedigree function, and an outcome label indicating diabetes status.

---

Objectives

1. Explore patient health metrics and identify meaningful predictors.
2. Build an ML classifier optimized for high recall (minimizing false negatives).
3. Use a modular, production-inspired project structure.
4. Provide clear evaluation metrics and interpretable insights.
5. Develop a scalable baseline for future healthcare analytics workflows.

---

Project Highlights

1. Data Preprocessing

• Converted non-physiological zero values to missing values
• Applied median imputation across relevant fields
• Centralized all cleaning logic in src/data/clean_data.py

2. Exploratory Data Analysis (EDA)

• Investigated variable distributions, correlations, and feature interactions
• Identified glucose, BMI, age, and pedigree function as key predictors
• Restricted notebooks to exploratory purposes only; production logic is modular

3. Machine Learning Model

• Used a RandomForestClassifier with class weighting to address imbalance
• Performed hyperparameter tuning via GridSearchCV (scoring = recall)
• Implemented training and evaluation pipelines in src/models/
• Saved the final trained model to the models/ directory

4. Visualization

Tools provided in src/visualization/ include:

• ROC curve generation
• Feature importance visualization
• Confusion matrix plotting

---

Tools and Technologies

• Python
• pandas
• numpy
• scikit-learn
• matplotlib
• joblib
• argparse
• Jupyter Notebook
• Modular analytics engineering structure

---

Results

Final Model Performance (Hold-Out Test Set)

• Accuracy: 0.747
• Precision: 0.632
• Recall: 0.667
• F1 Score: 0.649
• ROC AUC: 0.821

Interpretation

• Recall (0.667): The model captures most true diabetes cases, aligning with the priority of reducing false negatives.
• ROC AUC (0.821): Indicates strong class separability and reliable predictive discrimination.
• Precision and F1 remain balanced while recall is prioritized.

Earlier experimental approaches—threshold shifting, SMOTE oversampling, and initial hyperparameter tuning—helped guide the final implementation.
The structured pipeline improves clarity, reproducibility, and engineering quality without sacrificing clinical relevance.

---

Pipeline Usage

Training the Model

Run the following command:

python main.py train

This handles data cleaning, splitting, hyperparameter tuning, model evaluation, and saving the final model.

Generating Predictions

To create predictions:

python main.py predict --input data/raw/diabetes.csv --output predictions.csv

The output file will contain a new column named prediction.

---

Future Improvements

• Decision threshold tuning for recall/precision optimization
• Optional SMOTE or alternative class balancing
• SHAP for interpretability
• FastAPI for real-time scoring
• MLflow for experiment tracking
• Unit testing and CI/CD pipeline integration

---

Author

Julian Charlan Kelly
Analytics Engineer / Data Engineer
Los Angeles, CA