B.BRANDON FRANK HAN-CHWEEN HUANG← The notebook

RESEARCH LAB 07 / MEDICINE × MACHINE LEARNING

How much should
we trust it?

Three tests on public medical datasets: how reliable are the probabilities, how wide should prediction intervals be, and what happens when inputs are missing?

3 public datasetsHeld-out test resultsReproducible Python
Research, not a medical tool. These experiments use historic public datasets. They do not diagnose, predict your health, or establish clinical safety. The models were run for this notebook; the data were collected by the cited researchers.

Loading recorded experiment results…

METHODS & DOWNLOADS

Want to check the work?

Every split, parameter, prediction, and plot is available. No model was chosen or retuned from these test results. The numerical results shown here are from a completed run, not a live training animation.

Limits of these experiments.

The three cohorts contain 569, 442, and 299 records. Each result comes from one fixed split. Test-row intervals do not include uncertainty from retraining; randomly removed values do not reproduce the differences between hospitals. There is no prospective, external-site, or clinical-utility validation here.

Prepared for Brandon Frank Han-Chween Huang with AI-assisted analysis and writing. Experimental notebook, November 13, 2026. Not peer-reviewed.