Analysis of Diabetes Predictors: Building Interpretable Machine Learning Models Using LASSO and PLSR

This project is a final project for a bioengineering course at the University of California, Berkeley. Its core goal is to build an interpretable diabetes prediction model. Unlike black-box models, this project emphasizes that the model should not only predict but also clearly identify key biomarkers. It mainly uses two methods: LASSO regression and partial least squares regression (PLSR), balancing predictive performance and interpretability to provide references for the clinical application of medical AI.

This project is from the final assignment of Berkeley's "BioE 175: Data-Driven Models and Machine Learning" course. The core objective is to build an interpretable diabetes prediction model—In medical AI, doctors and patients need to understand the basis of model decisions, so interpretability is crucial. This project addresses this need and aims to identify key diabetes predictors from biomarkers.

The project uses a real biomedical dataset that has undergone strict preprocessing. The feature engineering process includes: 1. Data cleaning (handling missing values and outliers); 2. Screening relevant biomarkers based on domain knowledge; 3. Logarithmic/power transformation of non-normal features; 4. Standardization. The relevant steps are recorded in feature_engineering_notebook.ipynb.

LASSO Regression: Achieves prediction and feature selection through L1 regularization, shrinks some coefficients to zero, screens important biomarkers, avoids overfitting, and the model is sparse and easy to interpret. PLSR: Handles multicollinearity in high-dimensional data, reduces dimensionality through latent variables, captures the joint effects between features, and complements LASSO. Relevant implementations can be found in LASSO_Data_Analysis.ipynb and plsr_notebook.ipynb.

The project achieves interpretability through three methods: 1. Feature importance visualization (ranking of non-zero coefficients in LASSO); 2. Statistical significance testing (cross-validation and confidence intervals to ensure reliable results); 3. Clinically interpretable expressions (e.g., "For every 1-unit increase in fasting blood glucose, the risk of diabetes increases by X%").

Insights: Medical AI needs to balance accuracy and interpretability; domain knowledge is crucial for feature engineering; interpretable models are more likely to meet regulatory requirements. Expansion Directions: Application to multiple diseases, longitudinal data analysis, ensemble learning, external validation, and development of clinical decision tools.