Biomedical Machine Learning / Shiny App
Autoimmune Disease Classification using Machine Learning
AutoimmuneMap
Overview
AutoimmuneMap is a machine learning research prototype and Shiny application for exploring autoimmune disease classification using whole-blood gene expression data. The project combines model benchmarking, reduced-gene analysis, and an interactive interface for sample selection, prediction visualisation, and model interpretation.
Shiny App Preview
My Contribution
- Developed and deployed the AutoimmuneMap Shiny app for interactive disease prediction from gene expression data.
- Implemented uploaded data handling and gene-feature matching for user-provided datasets.
- Integrated trained machine learning models into the app prediction workflow, including prediction visualisation and explanation-oriented result display.
- Designed the user-facing workflow to make model outputs easier to inspect and interpret within a biomedical research prototype setting.
- Supported model training workflow, model artifact validation, and external GEO validation audits, including communication of dataset shift and validation constraints.
Team Attribution
AutoimmuneMap was completed as a group research prototype. Team members are acknowledged below, with my specific contributions summarised separately.
Team members: Ivy Chen, Mikhail Mundackal, Kenny Lu, Jemima Siu, Summer Hao, and Lucy Bao.
Model Performance
| Model | Accuracy | Weighted F1 |
|---|---|---|
| KNN | 0.709 | 0.767 |
| SVM | 0.707 | 0.794 |
| LASSO / Elastic Net | 0.562 | 0.497 |
Nested cross-validation results were used for model comparison. These results should be interpreted as internal model evaluation rather than external validation.
Saved best model internal validation accuracy: 0.88
Feature Reduction Story
The reduced-gene analysis explored whether smaller gene subsets could preserve useful classification performance. This is useful for interpreting model compactness and potential downstream feasibility.
| Top N Genes | Accuracy | Weighted F1 |
|---|---|---|
| Top 5 | 0.613 | 0.563 |
| Top 10 | 0.676 | 0.660 |
| Top 20 | 0.717 | 0.701 |
| Top 50 | 0.774 | 0.765 |
| Top 90 | 0.786 | 0.778 |
Limitations and Validation
External validation experiments identified dataset shift and validation limitations, including prediction collapse under some external data settings. This highlighted the need for stronger cohort-level validation, preprocessing alignment, and careful interpretation before any clinical or biomedical use.
- The system is a research prototype and not a clinical diagnostic tool.
- Model performance depends on dataset composition, preprocessing, and cohort compatibility.
- External validation showed dataset shift risks, so results should be interpreted cautiously.
- Further validation across independent cohorts would be needed before any practical biomedical use.
Project Ownership and Reuse Notice
AutoimmuneMap was developed as a group research prototype for academic coursework. The project materials shown here are provided only for portfolio demonstration and attribution of contribution.
Please do not copy, reproduce, redistribute, or submit any part of this project, including screenshots, report content, code structure, model workflow, or written descriptions, for coursework or assessment purposes without permission from the original contributors.
This page intentionally does not provide the full source code, report, dataset, or complete implementation details.