NEEMRs: Standardized Neurology EMR Module

• NEEMRs is a standardized module that integrates structured neurology EMRs with mechanistic and statistical models to enhance diagnostic accuracy.
• The preprocessing pipeline denoises 1,001 EMRs by tokenizing clinical narratives and employing expert-driven constrained reasoning to extract diagnosis–evidence pairs.
• NEEMRs leverages a comprehensive knowledge graph and multi-agent reasoning to achieve improved F1 scores and robust, clinically validated decision support.

The Neurology Subset of the Clinical Model-based Electronic Medical Record (CMEMR), known as NEEMRs, serves as a standardized computational and informatics module for representing, analyzing, and supporting clinical decision making in neurology. NEEMRs draws upon neuro-structured EMR data, mechanistic and statistical models, and explicit logical relations between neurological diseases, providing an integrated framework for research and diagnostics in both cerebrovascular and neurodegenerative domains.

1. Dataset Composition and Preprocessing

NEEMRs comprises 1,001 neurology EMRs, each containing on average 2.75 confirmed diagnoses and a mean tokenized input length of 582.4. Clinical narratives are structured into a concatenated format spanning Chief Complaint, Basic Information, History of Present Illness (HPI), Past Medical History (PMH), Physical Examination, and Auxiliary Examination (including lab and imaging summaries). Structured fields are extracted with a named entity recognizer (RaNER), yielding annotations in 9 semantic categories (symptoms, diseases, drugs, body parts, items/tests, equipment/imaging, microbiology, department, procedure).

The preprocessing pipeline for NEEMRs emphasizes denoising and standardization of heterogeneous input. Steps include:

1. Summarization of the free-text EMR into a distilled token sequence $s$.
2. Constrained Chain-of-Thought (CoT) reasoning by a primary expert agent, yielding sets of diagnosis–evidence pairs,

$$\mathcal{K}_{\mathrm{pri}} = \{(d_i, \mathcal{E}_i)\}_{i=1}^N, \quad \mathcal{E}_i = \{e_{i1}, \dots, e_{im}\}.$$

1. Mapping all evidence entities $\mathcal{E} = \bigcup_i \mathcal{E}_i$ to a standard comparison table $\mathcal{U}$, yielding $\mathcal{E}_\mathrm{std}$.
2. Embedding-based fuzzy matching (Bge-small-zh) for normalization and noise suppression.

This pipeline reduces input sparsity and harmonizes variable notations, allowing subsequent AI models to process more consistent neurological EMR representations (Shen et al., 1 Feb 2026).

2. Disease Taxonomy and Inter-Disease Logical Relations

NEEMRs encodes a detailed taxonomic hierarchy of neurological disorders, prominently featuring cerebrovascular events (e.g., ischemic stroke subtypes, hemorrhagic stroke, transient ischemic attacks) and neurodegenerative diseases (e.g., Alzheimer's disease, Parkinson's disease). The taxonomy follows established clinical ontologies and cross-references sources such as Zhou et al. 2024 via hierarchical relationships (Shen et al., 1 Feb 2026).

Three principal relation types are instantiated from the Medical Knowledge Graph (MKG):

• Mutual Exclusivity: Enforced when pathophysiology dictates that certain diagnoses cannot co-occur. For diseases $d_i$, $d_j$,

$$(d_i, d_j) \in \mathcal{R}_{\mathrm{exc}} \implies \neg(d_i \wedge d_j).$$

The exclusivity expert $\mathcal{A}_\mathrm{exc}$ nullifies the logic score $\mathcal{F}_\mathrm{logic}(d_i) \to 0$ if both exclusive diagnoses are predicted.

• Clinical Confusion: Defined by explicit “differs-from” edges in the MKG or significant feature overlap (e.g., $|\Phi(d_i) \cap \Phi(d_j)| \geq 2$). The confusion expert $\mathcal{A}_\mathrm{conf}$ penalizes ambiguous disease pairs based on shared evidence and uncertainty scores.
• Comorbidity/Causality: Recognizes permitted coexistence or causal sequence between diagnostic classes.

Explicit formalization of these relations enables NEEMRs-based systems to systematically reject clinically implausible hypotheses and manage differential diagnoses, bringing computational models closer to expert clinical reasoning (Shen et al., 1 Feb 2026).

3. Knowledge Graph Construction and Integration

NEEMRs leverages CPubMed-KGv2 as its backbone MKG, encompassing 1.7 million entities and 3.9 million relations across diseases, symptoms, tests, drugs, and anatomical sites. Entity–disease connections are computed by path distance in the graph: $$\Gamma(e, d) = \begin{cases} 1, & \mathrm{dist}_{\mathcal{G}}(e, d) \le 3 \ 0, & \text{otherwise} \end{cases}$$

Expert agents employ the MKG in several ways:

• Laboratory Expert ($\mathcal{A}_\mathrm{lab}$): Predicts dynamic weights $w_\mathrm{type}(e)$ for each clinical entity via LLM-based reasoning $\psi_\mathrm{wei}(\cdot)$. Abnormal findings ($e \in \mathcal{E}_\mathrm{abn}$) are upweighted:

$$w'_e = \begin{cases} 1.5\,w_\mathrm{type}(e), & e \in \mathcal{E}_\mathrm{abn} \ w_\mathrm{type}(e), & \text{otherwise} \end{cases}$$

• Disease Scoring: Integrated across KG evidence with

$$S(d) = \alpha_1 V(d) + \alpha_2 C(d), \quad \alpha_1 + \alpha_2 = 1,$$

where $V(d)$ and $C(d)$ aggregate evidence proximity and connectivity, respectively.

Abnormality-driven weighting and graph-based evidence aggregation directly address the challenge of heterogeneous, noisy neurology EMRs by amplifying salient clinical signals while suppressing confounders (Shen et al., 1 Feb 2026).

4. Evaluation Metrics and Diagnostic Workflow

Quantitative results on NEEMRs benchmark the efficacy of multi-agent reasoning and logical validation. Using the Qwen2.5-7B-Instruct LLM backbone:

Method	Recall	Precision	F1
Chain-of-Thought (CoT)	37.80	39.03	38.40
MindMap (LLM⊕KG)	45.54	41.91	43.65
Graph-CoT (LLM⊗KG)	45.43	41.66	43.46
RE-MCDF (NEEMRs)	46.13	42.28	44.11

Ablation studies on NEEMRs F1:

Configuration	NEEMRs F1
Full RE-MCDF	44.11
w/o MKG supplement ($\mathcal{G}$)	42.20
w/o laboratory expert	41.70
w/o relation experts	43.72
Direct LLM prediction (no pipeline)	41.28

These results demonstrate that explicit KG supplementation, dynamic laboratory weighting, and multi-relation reasoning each contribute essential performance gains (+1.0–1.5 percentage points F1 over KG-enhanced baselines) (Shen et al., 1 Feb 2026).

The closed-loop workflow—generation, verification, revision—enables dynamic self-correction. When the adjustment agent ($\mathcal{A}_\mathrm{adj}$) detects logical inconsistencies in candidate diagnoses, it triggers re-examination of the evidence, refining the reasoning chain without manual intervention.

5. Neurology-Specific Computational Challenges and Solutions

Neurology EMRs are typified by variable data quality, sparse labels, and complex comorbidities. NEEMRs-based pipelines address domain-specific issues via:

• Heterogeneous/Noisy Indicators: Laboratory expert’s real-time reweighting amplifies high-value abnormalities, normalizes across diverse test types, and mitigates missing/uncertain values.
• Implausible Comorbidities: Multi-relation experts enumerate mutual exclusivity and confusion edges, enforcing

$$\forall (d_i, d_j) \in \mathcal{R}_\mathrm{exc}: \neg(d_i \wedge d_j)$$

and attenuating ambiguous pairs.

• Absence of Self-Correction in LLMs: The closed-loop “generation → verification → revision” enables detection and reconciliation of logical conflicts.

An illustrative case demonstrates this pipeline: for a 78-year-old female with vertigo and CT-confirmed hemorrhage, the laboratory expert assigns a decisive weight to “cerebellar hemorrhage,” mutual exclusivity nullifies “infarction,” and the MKG surfaces “hypertension” as a relevant comorbidity. The system’s final ranking aligns exactly with clinical hierarchies, reflecting the successful operationalization of NEEMRs logic (Shen et al., 1 Feb 2026).

6. NEEMRs as a Computational Neurology Module

Mechanistic neural models and data-driven classifiers are both supported in NEEMRs. Mechanistic modules include compartmental neuronal models: $$C\,\frac{dV_{\rm soma}}{dt} = -I_{\rm DC} - \sum_{i}I_{i} - g\,(1-p)\,(V_{\rm soma}-V_{\rm dend}) - I_{\rm syn}$$ with population-level ODEs (Wilson–Cowan-type) for oscillatory/functional network phenomena. Data-driven components leverage multi-kernel SVMs, kernel ridge regression, and deep neural networks for diagnostic classification and severity scoring, integrating multi-modal EMR features spanning demographics, genomics, imaging, and cognitive scores (Wong-Lin et al., 2020).

The standardized schema for NEEMRs modules encompasses:

• Demographics, genetic risk factors (e.g., APOE genotype), cognitive test batteries, neuroimaging (GM/CSF volumes, PET SUVRs), laboratory biomarkers, and comorbid conditions.
• Outputs include probabilistic diagnostic stages, continuous severity indices, and decision support recommendations.
• Validation is performed via internal cross-validation and external multi-centre prospective studies, with regulatory compliance to Software as a Medical Device standards (e.g., FDA, ISO 13485, IEC 62304, GDPR/HIPAA).

7. Integration and Future Directions

NEEMRs is architected for modular, standards-based interoperability with EHR platforms and clinical decision support systems. Technical stack recommendations include microservices for data ingestion, preprocessing, mechanistic simulation, ML inference, and visualization; containerized deployments (Docker/Kubernetes); GPU/HPC use for resource-intensive computation.

Ongoing recommendations include hybrid modeling—linking mechanistic predictions (e.g., oscillatory disruptions) to ML pipelines—data harmonization using common data elements and ontologies (SNOMED CT, LOINC), federated learning for privacy-preserving cross-site training, and tight workflow integration via HL7/FHIR and SMART platforms (Wong-Lin et al., 2020).

Research opportunities lie in multi-scale modeling from molecular to network to behavioral levels, the integration of real-time digital biomarkers (e.g., gait, sleep, remote digital exams), and continual model recalibration with outcome-driven feedback loops.

NEEMRs thus constitutes a rigorously validated, extensible, and clinically actionable framework for computational neurology research and practice, enabling reproducible and interoperable workflows across heterogeneous clinical datasets (Shen et al., 1 Feb 2026, Wong-Lin et al., 2020).