MRI, fMRI, EEG: similar performance, but different explanations depending on the geometry chosen for comparing patients

Changing the way in which the distance between brain matrices is measured has little effect on a classifier’s accuracy, but may alter its individual decisions and the explanations it provides for them.

In neuroimaging, the data often take the form of matrices (connectivity, covariance) that are compared between patients using a particular geometry. An Italian team led by Nicola Amoroso tested on arXiv whether this choice, generally seen as a technical detail, affects the models. It compared two geometries, Frobenius and trace, on the same matrices from morphometric MRI, functional MRI and EEG.

Predictive performance remains broadly the same, but the agreement ends there: the decision scores per patient, and sometimes the predicted diagnoses, differ. Local neighbourhoods, sensitivity to disturbances and the ranking of variables deemed explanatory also change. Reducing the dimension brings the two geometries closer together, at the cost of a decrease in performance and agreement on decisions. For the authors, the geometry of data is a hidden degree of freedom from explainability: two models as precise as each other can justify the case of the same patient differently, a point to be verified before any clinical use.