A dose-unit-aware predictive toxicology framework for literature-derived <i>in vitro</i> carbon nanotube cytotoxicity

Logistic Regression and Random Forest classifiers were trained using log-transformed dose-related variables and experimental-context predictors. Performance was assessed using stratified cross-validation, article-grouped cross-validation, leave-one-article-out validation, and leave-one-family-out validation with fold-level variability considered to reflect uncertainty and to mitigate overoptimistic estimates arising from within-study similarity.

Random Forest showed higher point estimates under conventional stratified validation, whereas Logistic Regression provided comparatively stable discrimination under stricter grouped-validation settings. Ablation-based stress testing identified high-dose nontoxic A549/MTT observations as a major source of false-positive toxicity predictions, particularly for Random Forest. A dose × dose-unit interaction sensitivity analysis did not consistently improve grouped or leave-one-out validation; therefore, the additive unit-aware model was retained as the primary unit-aware analysis.

The unit-aware analysis demonstrated that surface-dose records can be incorporated without assuming equivalence between nominal mass concentration and surface-area-normalized exposure. The framework provides a transparent strategy for curating, validating, stress-testing, and cautiously modeling heterogeneous literature-derived CNT viability data in nanosafety research.</p>