CL5D-ICU: A Deterministic Multi-Agent Framework for Liver Transplant Risk Minimization Toward Zero-Approaching Risk via Phase-Based Consciousness-Complexity Scoring

Abstract

Liver transplantation carries substantial residual risk even after graft allocation, and ex-isting severity scores (MELD, Child-Pugh) were developed for allocation prioritization ratherthan continuous, multi-organ risk minimization. We present CL5D-ICU, an adaptation of theCL5D Hybrid Model’s five-agent (At, Ab, Ex, T, Ct) architecture to liver transplant risk assess-ment, and validate it against three independent, real-world reference points. On the Dicksonet al. (1989) Mayo Clinic PBC cohort (n = 293 patients with complete data, obtained via theUCI ML Repository), CL5D’s Cn score achieves AUC = 0.850 for mortality discrimination,closely tracking the original, externally-validated Mayo Risk Score (AUC = 0.857, fit on thissame cohort) with correlation r = 0.925. On a second, independent population (the IndianLiver Patient Dataset, n = 584, a different task of disease-presence discrimination), CL5Dachieves a weaker but still-significant AUC = 0.674; a direct cross-dataset transfer test us-ing a conventional logistic regression shows a comparable generalization gap, suggesting thisreflects genuine population/task differences rather than a defect specific to CL5D. A small fea-sibility pilot on the openly-accessible MIMIC-IV Clinical Database Demo demonstrates thattrue, non-approximated MELD-Na computation and CL5D scoring can be executed end-to-endon real ICU lab data, though with only 3 death events this pilot is not statistically poweredand is reported as a methodological proof-of-concept only. We additionally identify and doc-ument, rather than conceal, two structural limitations in the current CL5D-ICU formula: anachievable-Cn floor (≈ 0.0067) that prevents the model from reaching its own canonical ”Excel-lent, risk-near-zero” threshold (0.000123) for any input, and a scope mismatch in which the Ct(Catalyst) agent and Phase II–IV cycle — appropriate for other CL5D application domains —do not correspond to the pathophysiology of native liver decompensation or to the discontinuousnature of organ transplantation. We restrict this model’s clinical scope to Phase I only pendinga formula recalibration, and outline the credentialed data access (PhysioNet/MIMIC-IV) andLLM-integration architecture needed for a fully-powered future validation.