#29 · ML / AI methods

Calibration and controlled comparison now gate claims for biomedical predictive models

The dominant thread is scrutiny of how biomedical predictive models are evaluated rather than how they are built. A noise-limited study argues that measurement noise, not model capacity, often lets linear models match nonlinear ones on biomedical tabular data, and derives an excess-risk identity to formalise this. An endoExplain protocol audits colonoscopy image classifiers for discordance between class activation maps and scores across calibration, overlap and transfer. RepGene compares fusion operators for multi-view gene representations under a matched contract, including missing views. A Parkinson's study applies posterior and refit calibration to probabilistic motor states, while PBR and PIM metrics test whether biological data support state or direction-of-change prediction. What is new is treating calibration, controlled comparison and data sufficiency as prerequisites for claims.