Project / 02 / Mayur Patil
Clinical Lab Abnormality Predictor
Explore diabetes-risk modeling with data curation, Random Forest and PyTorch models, age-group fairness audits, and an explainable API.
The problem
A score alone tells little about how a model behaves. This project builds the surrounding workflow: data checks, evaluation, explanations, and repeatable deployment.
How it works
Use a small, well-understood dataset to focus on ML engineering. Keep training, evaluation, governance, and serving as separate parts of the system.
Evidence and evaluation
Independent Random Forest baseline: ROC AUC 0.822 and accuracy 75.3% on 154 held-out records. Stratified 80/20 split and training-only preprocessing. This baseline is separate from the deployed model; research only.
Limitations
Research only. The displayed evaluation is a separate eight-feature Random Forest baseline with training-only median imputation. It does not reproduce the deployed model or the README scores. A single split from a specific population cannot establish clinical validity or fairness.
About the author
Mayur Patil is a data scientist and ML engineer pursuing an M.S. in Data Science at The George Washington University, based in Falls Church, Virginia. His work spans data engineering, applied ML, language models, and sports forecasting.
More about Mayur Patil ↗