Visualizing diagnostic uncertainty in tabular data: an information-theoretic matrix approach

Official URL: https://dx.doi.org/10.1109/UncertaintyVisualization68947.2025.00011

We introduce the uncertainty matrix, an information-theoretic transformation that exposes patient- and feature-level uncertainty directly from raw tabular data, without relying on prior knowledge of measurement error or data provenance. Applied to 202 patients with myocarditis or acute coronary syndrome (ACS), the matrix quantifies diagnostic uncertainty through (i) atypicality relative to the nearest class, (ii) class-specific feature entropy, and (iii) residual class separability. A density-aware t-SNE embedding converts this high-dimensional matrix into a two-dimensional diagnostic landscape that reveals high-confidence class cores and highlights zones of diagnostic ambiguity. A hold-out experiment demonstrates that unseen patients can be projected into this space - without retraining - and consistently fall in clinically meaningful regions. Fully compatible with standard classification models, this framework provides a scalable, interpretable, and deployment-ready tool for real-time decision support in healthcare and beyond.