Context
The problem
Healthcare prediction work is not only about producing a score. Models need reproducible experiments, understandable feature contributions, and clear boundaries around what a prediction can and cannot say.
My role: Designed and built as Machine Learning Intern (Remote) at Ezitech Institute (Eziline Software House); maintained and improved since.
Implementation
How it works
Prepare and compare
The workflow organizes dataset-specific preprocessing and classifier benchmarking, rather than treating different clinical prediction tasks as one interchangeable problem.
Track experiments
MLflow provides a record of model runs so evaluation can be compared and revisited instead of relying on notebook state.
Expose explanations
SHAP is used to inspect feature contributions. Explanations describe model behaviour; they are not clinical reasoning or proof of causality.
Serve the workflow
A FastAPI service connects model operations to a Next.js interface, with Supabase used for application data and supporting functionality.
Implementation highlights
- Model training and benchmarking across seven de-identified healthcare dataset domains
- SHAP-based explanation workflow exposed through a FastAPI service
- Experiment tracking with MLflow and an interface for reviewing model behaviour
- MLOps-oriented drift monitoring and retraining workflow documented in the project
Evidence
What can be verified
The points below summarize the public repository and owner-provided CV where noted. Read the source for implementation details and the evaluation setup before drawing further conclusions.
Evaluation0.8222 AUC–ROCLightGBM diabetes risk model on a held-out test split (BRFSS, 70,692 records); the live model card documents the full evaluation.
Score check0.8222, not 0.83Not the internship's 0.83. That was the best across seven datasets. This 0.8222 is the live model alone, on its held-out test split.
Datasets7 domainsDe-identified healthcare dataset domains described in project documentation.
ReproducibilityMLflowExperiment tracking is part of the documented stack.
Boundaries
Limitations and next questions
- Results depend on the dataset, target definition, split strategy, and class balance; a single score should not be generalized across clinical use cases.
- The system is a technical project, not a certified medical device or a substitute for professional judgement.
- Dataset-level details and evaluation protocols should be reviewed in the repository before using results for research conclusions.
Evidence-first note: a live demo demonstrates an interface and workflow; it does not by itself validate model quality, security, or real-world impact.