Modular Meta-Evolutionary AI Architecture Enables Interpretable Stratification in Heterogeneous Clinical Trials

Foundation models are poorly matched to small, heterogeneous clinical datasets where transparency and auditability are required. We describe a modular, meta-evolutionary architecture that pairs an interpretable dynamical-systems learner (NetraAI) that leverages a long-range memory (LRM) mechanism to identify stable, outcome-linked Model-Derived Subgroups (MDS) and abstain when evidence is insufficient, coupled to a literature-grounded large language model (LLM) Strategist used for structured scientific critique and robustness testing. Across three datasets: CATIE schizophrenia (olanzapine vs perphenazine), CAN-BIND depression (escitalopram response), and COMPASS pancreatic cancer (GnP vs FOLFIRINOX), standard classifiers achieved near-chance whole-cohort prediction (AUC 0.51-0.57), whereas NetraAI identified compact 2-4 variable MDS yielding high discrimination within high-confidence patients (AUC up to ~1.0), including a 3-SNV signature with high C-for-benefit (0.92) for regimen selection in PDAC. Combining dynamical-systems subgroup discovery with LLM-guided scientific critique embodies the AI modularity hypothesis – showing how distinct computational models can jointly transform small, heterogeneous clinical datasets into concise, interpretable patient stratifications.