Preprint reports gains from neuro-symbolic clinical AI explanations

A preprint reports that NEURON, a neuro-symbolic system for acute heart-failure mortality prediction, improved AUC from 0.71-0.74 to 0.74-0.77 on the MIMIC-IV dataset. The authors combine SNOMED CT-informed representations with machine-learning models, then use a retrieval-augmented language model to turn SHAP feature attributions and patient notes into explanations. They report a human-aligned explanation score of 0.807, compared with 0.558 for SHAP-based explanations alone. These are preprint results, not evidence of clinical deployment.