Clinical Interpretability

Anatomically informed whitening improved interpretation of linear neuroimaging classifiers without materially changing predictive accuracy. Pairwise whitening reduced left-right homologous coupling and GM-CSF inverse relationships that complicate coefficient interpretation. After whitening, region-importance rankings were interpreted as more consistent with psychiatric neuroimaging literature. Compact fixed-size tokens may make downstream clinical explanations more traceable. A right bundle branch block case study aligned dominant fingerprint activation with physiologically meaningful QRS deflections. Projecting coefficients back to original feature space preserved clinically recognizable ROI terms. The framework is not presented as a diagnostic system replacing clinicians. The approach addressed interpretability rather than predictive performance. Attention pooling is intended to probe whether pathology depends on sparse token-level biomarkers. The intended result of TS-Fingerprint is to support compact digital biomarkers and interpretability.