{"id":"3e7c666b-94bc-4b06-bc6b-8238bcf8a822","arxiv_id":"2605.01189","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"NEURON integrates SNOMED CT, ML, and RAG LLM to raise AUC from 0.74-0.77 to 0.84-0.88 and human-aligned explainability scores from 0.50 to 0.85 on MIMIC-IV acute heart failure data.","lead":"NEURON is a neuro-symbolic system that combines SNOMED CT ontology with machine learning models and a RAG-grounded LLM to generate natural-language explanations for clinical predictions like heart failure mortality. A smart generalist might read it to see one engineering approach to making AI more trustworthy and usable in real medical settings.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"RAG-LLM synthesis into professional-level explanations lacks verification details","rationale":"The reader's weakest_assumption correctly isolates the least-supported step in the central claim. The AUC numbers would need separate ablation checks, but the novel contribution is the explanation layer, whose reliability is asserted without supporting evidence in the abstract. Full text availability does not alter this gap until the evaluation details are examined.","tokens_in":1709,"tokens_out":288,"duration_ms":49785,"concrete_test":"Extract the exact human evaluation protocol and rubric from the methods/results sections; re-run the clinician rating study on 30 held-out cases with the same rubric and report inter-rater kappa plus accuracy delta; if the 0.85 score drops below 0.70 or kappa < 0.6, the interpretability claim does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the RAG-grounded LLM layer reliably converts SHAP attributions plus patient notes into coherent, human-aligned explanations achieving professional clinical interpretability (0.85 vs 0.50). The abstract supplies no information on retrieval corpus construction, prompt design, hallucination mitigation, or the precise protocol behind the human-aligned metric. For the claim to hold, this component must demonstrably outperform raw SHAP without introducing factual errors or superficial fluency; the provided text leaves this untested.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents NEURON, a neuro-symbolic system that integrates SNOMED CT ontology-informed structural representations with ML models and a RAG-grounded LLM layer. The LLM synthesizes SHAP attributions and patient notes into natural-language explanations. On the MIMIC-IV dataset for Acute Heart Failure mortality prediction, it reports AUC improvement from 0.74-0.77 to 0.84-0.88 and human-aligned metrics of 0.85 versus 0.50 for raw SHAP visualizations.","tokens_in":1811,"tokens_out":393,"duration_ms":31634,"significance":"If the reported gains are supported by rigorous evaluation, the approach could advance clinical AI by grounding explanations in medical ontologies while using LLM synthesis for narrative transparency. The combination of symbolic structure with RAG-LLM offers a concrete engineering path toward human-centered interpretability in high-stakes domains.","major_comments":[{"comment":"Abstract and Methods: The central performance claims (AUC lift to 0.84-0.88 and human metric 0.85 vs 0.50) are stated without any description of the base predictive model architecture, training procedure, data splits, statistical testing, or controls. This information is load-bearing for assessing whether the ontology integration and RAG layer produce the claimed predictive reliability and interpretability gains.","section":"Abstract"},{"comment":"RAG-LLM component (likely §3 or Methods): No details are supplied on retrieval corpus construction, prompt design, hallucination mitigation strategies, or the exact protocol and inter-rater reliability for the human-aligned metric. Without these, it is impossible to verify that the LLM layer reliably produces coherent, professional-level explanations rather than superficial fluency.","section":"Methods / RAG-LLM layer"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. The comments correctly identify that key methodological details supporting the reported performance gains are insufficiently described in the current version. We will revise the manuscript to incorporate the requested information.","responses":[{"response":"We agree that the abstract and Methods section as submitted do not provide adequate detail on these elements. In the revised manuscript we will expand the Methods section to specify the base model architecture (including feature engineering from SNOMED CT embeddings), training procedure and hyperparameters, train/validation/test splits on MIMIC-IV, statistical testing (e.g., DeLong tests or bootstrap confidence intervals for AUC differences), and control experiments that isolate the contribution of the ontology and RAG components.","revision_made":"yes","referee_comment":"[Abstract] Abstract and Methods: The central performance claims (AUC lift to 0.84-0.88 and human metric 0.85 vs 0.50) are stated without any description of the base predictive model architecture, training procedure, data splits, statistical testing, or controls. This information is load-bearing for assessing whether the ontology integration and RAG layer produce the claimed predictive reliability and interpretability gains."},{"response":"We acknowledge the absence of these implementation details. The revised Methods section will describe: (1) retrieval corpus construction (SNOMED CT concepts plus de-identified MIMIC-IV notes indexed for RAG), (2) prompt templates and few-shot examples, (3) hallucination mitigation (document grounding, citation enforcement, and post-hoc fact-checking), and (4) the human evaluation protocol (number of clinicians, rating scale, inter-rater reliability via Cohen’s or Fleiss’ kappa, and exact computation of the 0.85 human-aligned score).","revision_made":"yes","referee_comment":"[Methods / RAG-LLM layer] RAG-LLM component (likely §3 or Methods): No details are supplied on retrieval corpus construction, prompt design, hallucination mitigation strategies, or the exact protocol and inter-rater reliability for the human-aligned metric. Without these, it is impossible to verify that the LLM layer reliably produces coherent, professional-level explanations rather than superficial fluency."}],"tokens_in":1374,"tokens_out":477,"duration_ms":33080,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The headline takeaway is that this paper describes a system called NEURON that wires SNOMED CT representations into an ML model, adds SHAP attributions, and routes the output through a RAG LLM to produce natural-language explanations. On the MIMIC-IV acute heart failure task it claims an AUC rise from the 0.74-0.77 range to 0.84-0.88 plus a jump in human-aligned scores from 0.50 to 0.85. That is the concrete claim.\n\nWhat the work actually does is assemble three existing pieces—ontology grounding, post-hoc attributions, and retrieval-augmented generation—into one pipeline and test it on a standard dataset. The numbers are stated plainly, which is better than many abstracts that stay at the level of “we improve interpretability.”\n\nThe problems are straightforward. The abstract supplies zero information on the base learner, feature construction, train-test split, hyper-parameter search, or any statistical test for the reported lift. There is no comparison to other neuro-symbolic or RAG-based clinical explainers, so it is impossible to know whether the gain is new or just a re-run of known tricks. The human evaluation metric is mentioned but not described—no protocol, no inter-rater numbers, no check for factual errors introduced by the LLM. The stress-test note is right on this point: without those details the central promise that the RAG layer produces professional-grade explanations remains untested.\n\nThis paper is aimed at applied clinical-AI groups that want a worked example of ontology-plus-LLM explainability. A reader looking for reproducible methods or a careful literature comparison will find little to use. Because the evidence for the headline improvements is missing, I would not send it to referees in its current form.","headline":"NEURON reports an AUC lift and better human metrics on MIMIC-IV heart failure data, but the abstract gives no architecture, training, or evaluation details so the claims cannot be assessed.","tokens_in":2331,"tokens_out":447,"would_cite":false,"duration_ms":24495,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"NEURON combines SNOMED CT ontology with RAG-LLM to raise clinical prediction AUC and human-aligned explainability.","keywords":["neuro-symbolic AI","clinical explainability","SNOMED CT","SHAP attributions","RAG","heart failure prediction","MIMIC-IV","LLM"],"falsifier":"A controlled evaluation in which clinical experts rate NEURON-generated explanations as no more interpretable or useful than raw SHAP visualizations or standard model outputs.","tokens_in":2626,"feed_emoji":"🩺","tokens_out":483,"duration_ms":30227,"temperature":0.7,"pith_summary":"The paper presents NEURON as a neuro-symbolic system that addresses the black-box problem in clinical AI by grounding predictions in medical ontology. It fuses SNOMED CT structural representations with machine learning models for improved accuracy. A RAG-grounded LLM then converts SHAP feature attributions and patient notes into natural language explanations. Validation on MIMIC-IV acute heart failure data shows AUC rising from 0.74-0.77 to 0.84-0.88 and superior human metrics over raw SHAP. This approach aims to enable trustworthy AI in connected health.","feed_headline":"Neuro-symbolic system lifts heart failure AUC to 0.84-0.88","feed_subtitle":"Ontology-grounded explanations from SHAP and notes outperform raw visualizations on human metrics for MIMIC-IV data.","key_machinery":"Neuro-symbolic integration of SNOMED CT ontology structural representations with ML models, followed by RAG-grounded LLM synthesis of SHAP attributions and notes into explanations.","core_discovery":"NEURON integrates SNOMED CT ontology-informed structural representations with machine learning models to bridge raw data and medical nomenclature, then employs a Retrieval-Augmented Generation grounded LLM layer to synthesize SHAP feature attributions and patient-specific clinical notes into coherent natural-language explanations, yielding both higher predictive performance and more clinically interpretable outputs on the MIMIC-IV acute heart failure mortality task.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["NEURON integrates SNOMED CT with ML for clinical explainability","NEURON uses RAG LLM to synthesize SHAP and patient notes","Heart failure mortality AUC reaches 0.84-0.88 with NEURON","Human metrics score 0.85 for NEURON explanations vs 0.50 raw SHAP","SNOMED CT ontology bridges raw data to medical nomenclature in NEURON"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The assumption that SNOMED CT ontology-informed structural representations combined with a RAG-grounded LLM layer will reliably synthesize SHAP attributions and patient notes into coherent, human-aligned natural-language explanations that deliver professional-level clinical interpretability.","fun_headline_variants_meta":{"raw":{"variants":["NEURON integrates SNOMED CT with ML for clinical explainability","NEURON uses RAG LLM to synthesize SHAP and patient notes","Heart failure mortality AUC reaches 0.84-0.88 with NEURON","Human metrics score 0.85 for NEURON explanations vs 0.50 raw SHAP","SNOMED CT ontology bridges raw data to medical nomenclature in NEURON"]},"model":"grok-4.3","cost_usd":0.006241,"raw_usage":{"total_tokens":2921,"prompt_tokens":634,"num_sources_used":0,"completion_tokens":104,"cost_in_usd_ticks":62412000,"prompt_tokens_details":{"text_tokens":634,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2183,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":634,"tokens_out":104,"duration_ms":23367,"temperature":1.0,"reasoning_tokens":2183,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T01:01:14.723859+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled evaluation in which clinical experts rate NEURON-generated explanations as no more interpretable or useful than raw SHAP visualizations or standard model outputs.","supporting_citations":[],"review_version":2}