<p>Accurate mortality prediction models for intensive care units (ICUs) have advanced significantly; however, translating probabilistic outputs into reliable clinical decisions remains challenging, particularly under predictive uncertainty. This study proposes an evolutionary-optimized adaptive neuro-fuzzy decision support framework that operates as a post-model reasoning layer, converting mortality risk and uncertainty estimates into conservative and interpretable escalation decisions. The framework integrates an Adaptive Neuro-Fuzzy Inference System (ANFIS) with genetic algorithms and particle swarm optimization to learn safety-oriented decision policies that prioritize false-negative minimization. Using the PhysioNet Challenge 2012 ICU dataset, evaluation is conducted at the decision level, emphasizing uncertainty-aware robustness and safety–workload trade-offs. Results show that static fuzzy and non-optimized neuro-fuzzy models exhibit unsafe under-escalation, missing the majority of high-risk patients. In contrast, evolutionary-optimized models substantially reduce false-negative rates, achieving near-complete elimination by converging to conservative escalation regimes with controlled increases in false positives. This work contributes: (i) a modular uncertainty-aware decision layer for ICU risk escalation, (ii) evolutionary optimization of neuro-fuzzy systems under explicit safety objectives, and (iii) a decision-centric evaluation framework focused on clinical risk and uncertainty. The proposed approach complements existing predictive models without requiring modification of their internal structure.</p>