A research system called NEURON aims to solve one of clinical AI's most stubborn problems: models that predict well but can't explain themselves in terms doctors actually use.
The system, validated on the MIMIC-IV dataset for predicting Acute Heart Failure mortality, works by layering three components together. It grounds its data in SNOMED CT, a structured medical terminology standard, then runs machine learning on top of that, then uses a retrieval-augmented generation layer to convert raw SHAP feature attribution scores into coherent clinical notes. The result is a model that lifted its AUC score from 0.74-0.77 to 0.84-0.88 — a meaningful jump — while also scoring 0.85 versus 0.50 against raw SHAP outputs on human-aligned readability metrics.
The explainability gap has been the quiet killer of clinical AI deployments for years. Regulators and clinicians both resist black-box systems; NEURON's approach of bolting a structured medical vocabulary onto the model before the LLM translates findings into prose is a more principled fix than post-hoc plain-language wrappers applied to ungrounded outputs. It also sidesteps one common failure mode: LLMs hallucinating clinical rationale that wasn't in the model at all.
Neuro-symbolic hybrids have cycled in and out of fashion in AI research for decades. Whether NEURON's architecture survives contact with messy real-world hospital data — far noisier than MIMIC-IV — is the question the paper doesn't yet answer.