How Firm Is an Earwax Test for Cancer Detection? Reading the Test at Several of Its Own Settings

Earwax gives off chemicals that studies in people and in dogs have linked to cancer. In the largest study so far, of 751 volunteers, the test had a sensitivity of 0.904 and a specificity of 0.880 after cross-validation. These figures describe how the test does on a group of people. They do not say how firm the verdict is for one person's sample. This note describes a way to find out. The measured data of the sample is kept exactly as it is. The settings of the analysis, such as the level at which a peak counts as present, and the settings of the classifier, are moved a little, and the verdict is read at each setting. The spread of the verdicts is the answer. A narrow spread means the verdict is firm. A wide spread means the verdict depends on a setting that should not matter, and the apparatus must be rethought before it is relied on. A computer-decided shunting action follows a wide spread. The computer decides without waiting for a person, and it moves that sample aside, for example to a repeat measurement or a confirming test, while the other samples carry on. The note covers the earwax studies, work on analysis settings in mass-spectrometry data, and methods that give a spread or a confidence from inside a model, such as Bayesian methods, Monte Carlo dropout, deep ensembles, the bootstrap and conformal prediction. For each one it explains how this method differs. The same reading applies to other tests built from the chemicals in earwax, saliva or breath. Keywords: earwax, cerumen, cancer detection, gas chromatography-mass spectrometry, uncertainty, firmness of a verdict, shunting action