Fig. 3
Feature importance of Neuro-TACTIC and its dependence on the cost-sensitivity parameter ζ. (A): Global importance of the ten most influential features in the development and validation cohorts (n = 1072, n = 81) at the optimal ζ = 0.9750. GAIN-based importance values (blue bars, plotted to the left) are computed on each training split during cross-validation and are therefore only available for the main dataset; features are sorted in descending order of mean GAIN. Mean absolute SHAP values are shown for both the development cohort (red bars) and the independent evaluation cohort (black bars), capturing each feature’s average contribution to individual predictions across all samples. Surgery duration ranks highest by both metrics, followed by tumor volume, surgical position, body mass index (BMI), and patient age. Error bars denote mean ± s.d. across 25 cross-validation folds. (B–E) Evolution of feature importance across ζ values between 0.5 and 1.5 (the “tuning” region) for the four top-ranked variables: surgical duration (B), tumor volume (C), Surgical position (D), and BMI (E). In each panel, red circles and solid lines indicate GAIN-based importance, while blue squares and dashed lines represent mean absolute SHAP values, both averaged over 25 cross-validation runs (error bars represent mean ± 1 s.d.). As ζ increases (greater emphasis on avoiding under-triage), gain importance for each feature rises—indicating that higher ζ settings drive the model to rely more heavily on these predictors to escalate care—whereas SHAP importance peaks at lower ζ, reflecting that when resource cost is prioritized, small changes in these features more strongly influence individual acuity assignments.
