NEURON: Neuro-Symbolic AI for Clinical Explainability
NEURON, a novel neuro-symbolic system, seeks to enhance the explainability of clinical AI by combining machine learning with the SNOMED CT ontology. This system incorporates a Retrieval-Augmented Generation (RAG) grounded LLM layer to transform SHAP feature attributions and patient notes into coherent natural-language explanations. When tested on the MIMIC-IV dataset for predicting mortality in Acute Heart Failure, NEURON increased the AUC from 0.74-0.77 to 0.84-0.88, surpassing the effectiveness of raw SHAP visualizations in human-aligned metrics (0.85 compared to the baseline). This development tackles the opaque nature of high-performing models used in clinical environments.
Key facts
- NEURON is a neuro-symbolic system for clinical explainability
- Integrates SNOMED CT ontology with machine learning
- Uses RAG-grounded LLM layer to generate natural-language explanations
- Validated on MIMIC-IV dataset for Acute Heart Failure mortality prediction
- Improved AUC from 0.74-0.77 to 0.84-0.88
- Outperformed raw SHAP visualizations in human-aligned metrics (0.85 vs. baseline)
- Addresses black-box nature of clinical AI models
- Published on arXiv as preprint 2605.01189
Entities
Institutions
- arXiv
- SNOMED CT
- MIMIC-IV