What it is
A research project I led at UW's Lutz Lab to improve the sensitivity of influenza rapid diagnostic tests by combining the binary test result with patient biometrics (demographics, symptoms, test conditions) through an ensemble of classical ML models.
Why I built it
Lateral flow tests have a known false-negative rate, especially at lower viral loads. But the test result isn't the only signal available. By the time a patient takes a test, we typically know their age, symptoms, and exposure context. The hypothesis: combining the test result with that context, with a known false-positive baseline as a constraint, should outperform either signal alone.
How it works
- Inputs: binary LFA result + patient biometrics
- Models: an ensemble of Random Forests, Support Vector Machines, and K-Nearest Neighbors classifiers
- Constraint: maintain a fixed false-positive rate (matching the underlying test) while maximizing sensitivity
- Statistical evaluation: likelihood ratios, odds ratios, and Wilcoxon signed-rank tests for paired comparisons against the LFA-only baseline
The ensemble vote was tuned to honor the false-positive constraint, not chasing raw accuracy, but lifting sensitivity at a fixed specificity.
What I learned
- Constrained optimization changes which model wins. A model that's more accurate on average can be a worse choice if it can't hit the false-positive ceiling reliably.
- Classical ML ensembles can be surprisingly strong on small clinical datasets. Deep learning isn't always the right answer when N is small and feature interpretation matters.
- Cross-validation rigor and statistical testing are non-negotiable in diagnostics. Small effect sizes need careful measurement to be defensible.
Links
- Lab: lutzlab.org (UW Bioengineering)