Water-Content-Aware Machine Learning for Screening Multiphysics Tissue Property Tables

Tissue property tables underpin radio frequency (RF) safety assessments, bioheat simulations, and magnetic resonance imaging (MRI) specific absorption rate (SAR) modelling. Because such tables consolidate measurements from many independent studies spanning different protocols and tissue conditions, cross-property patterns can be subtle and difficult to verify manually. We present a water-content-aware screening workflow that learns these cross-property relationships from the IT'IS Tissue Properties Database V5.0 (112 entries) and uses them to produce a ranked check-first list of entries that merit closer review. Three methodological contributions distinguish this work: (i) a $100 \%$ -missing feature column is explicitly identified and removed before modelling; (ii) binary missingness indicator features are added alongside median imputation, enabling the model to treat observed and imputed values differently; and (iii) a systematic duplicate featurevector audit is conducted before fitting, revealing 16 entry pairs that share identical post-imputation profiles. An Extremely Randomized Trees regressor achieves 5-fold cross-validation $\mathrm{R}^2=0.630$ (MAE $=7.48$ percentage points (pp), RMSE = 17.04 pp). Split conformal prediction wraps the model with distribution-free $90 \%$ uncertainty bands ( $\hat{\mathrm{q}}=14.0 \mathrm{pp}$ ). The resulting screening list and uncertainty estimates give simulation teams a principled, automated starting point for traceability review before tissue properties enter multi-physics workflows.