Skip to content

Neurology Intelligence Agent — Architecture Design Document

Author: Adam Jones Date: March 2026 License: Apache 2.0


1. Executive Summary

The Neurology Intelligence Agent extends the HCLS AI Factory platform to deliver RAG-powered clinical decision support across the full spectrum of neurological disease. It unifies fragmented neurological evidence — spanning cerebrovascular, neurodegenerative, epilepsy, movement disorders, multiple sclerosis, headache, neuromuscular, and neuro-oncology domains — into a single intelligence platform. Clinicians receive guideline-grounded, evidence-cited recommendations with validated clinical scale calculations in under five seconds.

The system implements 10 validated clinical scale calculators (NIHSS, GCS, MoCA, EDSS, Hoehn-Yahr, MDS-UPDRS Part III, HIT-6, ALSFRS-R, ASPECTS, mRS), provides 8 evidence-based clinical workflows, and searches 14 Milvus vector collections containing disease-specific literature, clinical trials, imaging protocols, treatment guidelines, and genomic correlations.

The platform enables time-critical queries like "Acute left MCA stroke, NIHSS 18, last known well 3 hours ago — treatment options?" that simultaneously search stroke protocols, imaging guidelines, treatment evidence, and clinical trials — returning grounded recommendations with guideline citations and scale calculations.

Key Results

Metric Value
Milvus collections 14 domain-specific collections (13 owned + 1 read-only)
Clinical scale calculators 10 validated instruments
Clinical workflows 8 + 1 general neurological query
Disease domains 10 (stroke, dementia, epilepsy, brain tumor, MS, Parkinson's, headache, neuromuscular, neurocritical, neuropathy)
Query expansion aliases 251+ across 16 synonym maps
Estimated total records ~855,000 across all collections
Test suite 209 automated tests across 12 modules

2. Architecture Overview

2.1 Mapping to VAST AI OS

VAST AI OS Component Neurology Agent Role
DataStore Raw files: PubMed XML, ClinicalTrials.gov JSON, guideline documents, imaging protocol specs
DataEngine Ingest pipelines for literature, trials, guidelines, imaging, drug, genomic, and pathway data
DataBase 14 Milvus collections (13 owned + 1 read-only) + knowledge base (disease taxonomy, drug catalog, guideline recommendations)
InsightEngine BGE-small embedding + multi-collection RAG + 10 clinical scale calculators + query expansion (251+ aliases)
AgentEngine NeurologyAgent orchestrator + Streamlit UI + FastAPI REST

2.2 System Diagram

+================================================================+
|                     PRESENTATION LAYER                          |
|  +---------------------+   +-----------------------------+     |
|  | Streamlit Chat UI   |   | FastAPI REST API             |     |
|  | Port 8534           |   | Port 8528                    |     |
|  | Interactive Q&A     |   | Versioned endpoints (v1)     |     |
|  | Scale calculators   |   | CORS, auth, rate limiting    |     |
|  +----------+----------+   +-------------+---------------+     |
+================================================================+
               |                            |
+================================================================+
|                    INTELLIGENCE LAYER                           |
|  +----------------+  +----------------+  +------------------+  |
|  | Query Expander |  | Workflow Engine |  | Clinical Scale   |  |
|  | 251+ aliases   |  | 8+1 workflows  |  | Calculators      |  |
|  | 16 synonym     |  | domain-specific|  | 10 validated     |  |
|  | maps           |  | weight boost   |  | instruments      |  |
|  +-------+--------+  +-------+--------+  +--------+---------+  |
|          |                   |                     |            |
|  +-------v-------------------v---------------------v---------+  |
|  |              Neurology RAG Engine                         |  |
|  | ThreadPoolExecutor parallel search across 14 collections  |  |
|  | Workflow-specific collection weight boosting              |  |
|  | Citation scoring (high/medium/low)                        |  |
|  | Conversation memory (24h TTL)                             |  |
|  +---------------------------+-------------------------------+  |
|                              |                                  |
|  +---------------------------v-------------------------------+  |
|  |              Claude Sonnet 4.6 (Anthropic)                |  |
|  | Evidence synthesis with clinical system prompt            |  |
|  | Guideline-grounded recommendations (AAN/AHA/ILAE/MDS)    |  |
|  +-----------------------------------------------------------+  |
+================================================================+
               |
+================================================================+
|                        DATA LAYER                              |
|  +-----------------------------------------------------------+  |
|  |              Milvus 2.4 Vector Database                   |  |
|  |  14 collections | BGE-small 384-dim | IVF_FLAT/COSINE    |  |
|  |  855K estimated records | etcd + MinIO backend            |  |
|  +-----------------------------------------------------------+  |
+================================================================+

3. Data Collections — Actual State

3.1 Collection Catalog

# Collection Est. Records Weight Primary Use
1 neuro_literature 500,000 0.12 PubMed neurological literature
2 neuro_trials 50,000 0.09 ClinicalTrials.gov neurology trials
3 neuro_guidelines 2,000 0.10 AAN/AHA/ILAE/MDS/IHS clinical guidelines
4 neuro_imaging 5,000 0.08 MRI/CT/PET protocols and findings
5 neuro_drugs 3,000 0.08 Neurological pharmacotherapy
6 neuro_genomics 10,000 0.07 Neurogenetic variants and associations
7 neuro_pathways 2,000 0.06 Neural pathways and circuits
8 neuro_scales 500 0.06 Clinical scale validation data
9 neuro_surgery 3,000 0.05 Neurosurgical procedures and outcomes
10 neuro_electrophysiology 5,000 0.06 EEG, EMG, NCS patterns and interpretation
11 neuro_biomarkers 2,000 0.05 CSF, serum, and imaging biomarkers
12 neuro_rehabilitation 3,000 0.04 Neurological rehabilitation protocols
13 neuro_case_reports 5,000 0.05 Published neurological case reports
14 genomic_evidence ~265,000 0.03 Shared genomic variant context

3.2 Index Configuration

Parameter Value
Index type IVF_FLAT
Metric COSINE
nlist 1024 (literature), 256 (trials), 128 (others)
nprobe 16
Embedding dim 384 (BGE-small-en-v1.5)

4. Clinical Scale Calculators

4.1 NIHSS (National Institutes of Health Stroke Scale)

Domain Items Score Range
Level of consciousness 3 items (LOC, LOC questions, LOC commands) 0-7
Gaze 1 item 0-2
Visual fields 1 item 0-3
Facial palsy 1 item 0-3
Motor arm 2 items (left, right) 0-8
Motor leg 2 items (left, right) 0-8
Ataxia 1 item 0-2
Sensory 1 item 0-2
Language 1 item 0-3
Dysarthria 1 item 0-2
Extinction/inattention 1 item 0-2
Total 15 items 0-42
Output Score, severity (minor/moderate/moderate-severe/severe), tPA eligibility window

4.2 GCS (Glasgow Coma Scale)

Component Range Best Response
Eye opening 1-4 Spontaneous
Verbal response 1-5 Oriented
Motor response 1-6 Obeys commands
Total 3-15 Severity: mild (13-15), moderate (9-12), severe (3-8)

4.3 MoCA (Montreal Cognitive Assessment)

Domain Max Points
Visuospatial/Executive 5
Naming 3
Memory 5 (delayed recall)
Attention 6
Language 3
Abstraction 2
Orientation 6
Total 30 (Normal >= 26, MCI 18-25, Dementia < 18)

4.4 Additional Scales

Scale Domain Range Key Thresholds
EDSS MS disability 0-10 (0.5 steps) 0 = normal, 6.0 = bilateral assistance, 10 = death
Hoehn-Yahr Parkinson's staging 1-5 1 = unilateral, 3 = bilateral with postural instability, 5 = wheelchair/bed
MDS-UPDRS III PD motor exam 0-132 18 items, each 0-4 severity scale
HIT-6 Headache impact 36-78 <= 49 little, 50-55 some, 56-59 substantial, >= 60 severe
ALSFRS-R ALS function 0-48 12 items: bulbar, fine motor, gross motor, respiratory
ASPECTS Stroke CT 0-10 >= 7 favorable, < 7 large core
mRS Disability outcome 0-6 0 = no symptoms, 6 = dead

5. Clinical Workflows

5.1 Workflow Catalog

# Workflow Clinical Question Key Scales Weight-Boosted Collections
1 Acute Stroke "tPA/thrombectomy eligibility for this stroke?" NIHSS, ASPECTS, mRS guidelines, imaging, drugs, literature
2 Dementia Evaluation "ATN staging and anti-amyloid eligibility?" MoCA, CDR guidelines, biomarkers, genomics, drugs
3 Epilepsy Classification "ILAE classification and drug-resistant assessment?" Seizure frequency guidelines, drugs, electrophysiology
4 Brain Tumor Grading "WHO 2021 molecular classification?" KPS guidelines, genomics, imaging, surgery
5 MS Monitoring "NEDA-3 status and DMT escalation?" EDSS guidelines, imaging, drugs, biomarkers
6 Parkinson's Assessment "Motor severity and DBS candidacy?" H&Y, UPDRS, MoCA guidelines, drugs, surgery
7 Headache Classification "ICHD-3 diagnosis and CGRP therapy guidance?" HIT-6 guidelines, drugs, imaging
8 Neuromuscular Evaluation "ALS vs. neuropathy differential?" ALSFRS-R electrophysiology, genomics, drugs

5.2 Workflow-Specific Weight Boosting

Each workflow dynamically adjusts collection weights. Example for Acute Stroke:

Collection Base Weight Stroke Boost Effective Weight
neuro_guidelines 0.10 2.0x 0.20
neuro_imaging 0.08 1.8x 0.14
neuro_drugs 0.08 1.5x 0.12
neuro_literature 0.12 1.0x 0.12
(others) varies 0.5-1.0x reduced

6. Multi-Collection RAG Engine

6.1 Search Flow

User Query: "Acute left MCA stroke, NIHSS 18, last known well 3h ago"
    ├── 1. Workflow classification: ACUTE_STROKE                  [< 1 ms]

    ├── 2. NIHSS calculation: Score 18 → Moderate-severe          [< 5 ms]
tPA window: Within 4.5hELIGIBLE
Thrombectomy: NIHSS >= 6, LVO suspectedEVALUATE
    ├── 3. Embed query with BGE asymmetric prefix                 [< 5 ms]

    ├── 4. Parallel search across 14 collections                  [12-18 ms]
    │      (with stroke workflow weight boosting)
    │   ├── neuro_guidelines:  AHA/ASA stroke guidelines (2x)    (score: 0.85-0.92)
    │   ├── neuro_imaging:     CTA/CTP/MRI stroke protocols (1.8x)(score: 0.80-0.88)
    │   ├── neuro_drugs:       tPA, TNK, antiplatelets (1.5x)    (score: 0.78-0.86)
    │   └── neuro_literature:  DAWN, DEFUSE-3 trials              (score: 0.75-0.84)
    ├── 5. Query expansion: "MCA stroke NIHSS 18" →               [< 1 ms]
    │      [middle cerebral artery, large vessel occlusion,
    │       alteplase, tenecteplase, thrombectomy, DAWN, ...]
    ├── 6. Knowledge base augmentation                            [< 1 ms]

    └── 7. Stream Claude Sonnet 4.6 response                     [~22-26 sec]
           AHA/ASA guideline-grounded recommendation:
           tPA eligibility, thrombectomy criteria (DAWN/DEFUSE-3),
           imaging protocol, BP management

Total: ~26 sec (retrieval + scales: ~30 ms; LLM: ~25 sec)

6.2 Citation Scoring

Level Threshold Display
High confidence >= 0.75 Full citation with source link
Medium confidence >= 0.60 Citation with caveat
Below threshold < 0.40 Filtered out

7. Performance Benchmarks

Measured on NVIDIA DGX Spark (GB10 GPU, 128GB unified LPDDR5x memory, 20 ARM cores).

7.1 Scale Calculator Performance

Scale Latency Validated Against
NIHSS (15 items) <5 ms NIH Stroke Scale training
GCS (3 components) <2 ms Standard GCS protocol
MoCA (7 domains) <5 ms MoCA validation study
EDSS (functional systems) <10 ms Neurostatus-eEDSS
Hoehn-Yahr <2 ms Original Hoehn & Yahr 1967
MDS-UPDRS III (18 items) <10 ms MDS-UPDRS validation
HIT-6 (6 items) <2 ms HIT-6 validation study
ALSFRS-R (12 items) <5 ms ALSFRS-R validation
ASPECTS (10 regions) <5 ms Original ASPECTS paper
mRS (single score) <1 ms Standard mRS protocol
All 10 scales <50 ms

7.2 RAG Query Performance

Operation Latency
Full query (retrieve + Claude generate) ~26 sec
Streaming query (time to first token) ~3 sec
14-collection parallel search 12-18 ms
Query expansion < 1 ms
Knowledge augmentation < 1 ms

8. Infrastructure

8.1 Technology Stack

Component Technology
Language Python 3.10+
Vector DB Milvus 2.4, localhost:19530
Embeddings BGE-small-en-v1.5 (BAAI) — 384-dim
LLM Claude Sonnet 4.6 (Anthropic API)
Web UI Streamlit (port 8534, NVIDIA black/green theme)
REST API FastAPI + Uvicorn (port 8528)
Configuration Pydantic BaseSettings with NEURO_ prefix
Testing pytest (209 tests)
Hardware target NVIDIA DGX Spark (GB10 GPU, 128GB unified, $4,699)

8.2 Service Ports

Port Service
8528 FastAPI REST API
8534 Streamlit Chat UI
19530 Milvus vector database (shared)

8.3 Dependencies on HCLS AI Factory

Dependency Usage
Milvus 2.4 instance Shared vector database — adds 13 owned collections alongside existing genomic_evidence (read-only)
ANTHROPIC_API_KEY Shared Anthropic API key
BGE-small-en-v1.5 Same embedding model as main RAG pipeline

9. Demo Scenarios

9.1 Validated Demo Queries

1. "Acute left MCA stroke, NIHSS 18, last known well 3 hours ago — treatment?" - NIHSS: 18 → Moderate-severe stroke - tPA: Within 4.5h window → Eligible (AHA/ASA Class I) - Thrombectomy: NIHSS >= 6 + suspected LVO → Evaluate with CTA (DAWN/DEFUSE-3 criteria)

2. "72-year-old with MoCA 22, amyloid PET positive — staging and treatment options?" - MoCA: 22 → Mild cognitive impairment - ATN staging: A+ (amyloid PET) → Evaluate tau and neurodegeneration - Anti-amyloid: Lecanemab eligibility criteria assessment

3. "Drug-resistant epilepsy, failed 3 ASMs — surgical candidacy evaluation?" - ILAE definition: Failed 2+ appropriately chosen ASMs → Drug-resistant - Surgical workup: Video-EEG monitoring, brain MRI (epilepsy protocol), neuropsych testing - Evidence: Cochrane review on epilepsy surgery outcomes

4. "WHO 2021 classification for IDH-mutant, 1p/19q codeleted brain tumor?" - Classification: Oligodendroglioma, IDH-mutant, 1p/19q-codeleted (WHO Grade 2 or 3) - Treatment: PCV chemotherapy + radiation (RTOG 9402, EORTC 26951 evidence)

5. "EDSS 4.0 MS patient on dimethyl fumarate with new enhancing lesions — escalation?" - EDSS: 4.0 → Ambulatory without aid, limited walking distance - NEDA-3: Failed (new MRI activity) - Escalation: Consider natalizumab, ocrelizumab, or ofatumumab (JCV status guides selection)


10. File Structure (Actual)

neurology_intelligence_agent/
├── src/
   ├── agent.py                     # Agent orchestrator
   ├── models.py                    # Enums and Pydantic models
   ├── collections.py               # 14 collection schemas
   ├── rag_engine.py                # Multi-collection RAG with weight boosting
   ├── clinical_scales.py           # 10 validated scale calculators
   ├── clinical_workflows.py        # 8 clinical workflows
   ├── knowledge.py                 # Domain knowledge base
   ├── query_expansion.py           # 251+ aliases, 16 synonym maps
   ├── cross_modal.py               # Cross-agent triggers
   ├── metrics.py                   # Prometheus metrics
   └── export.py                    # Report formats
├── app/
   └── neuro_ui.py                 # Streamlit chat interface
├── api/
   └── main.py                     # FastAPI (15 clinical endpoints)
├── config/
   └── settings.py                 # Pydantic BaseSettings (50+ params)
├── data/
   ├── cache/                      # Conversation persistence (24h TTL)
   └── reference/                  # Reference data files
├── scripts/
   ├── setup_collections.py
   ├── seed_knowledge.py
   └── run_ingest.py
├── tests/                          # 209 tests across 12 modules
├── requirements.txt
├── Dockerfile
├── docker-compose.yml
└── README.md

11. Implementation Status

Phase Status Details
Phase 1: Architecture Complete 14 collections, 10 scale calculators, 8 workflows, knowledge base, RAG engine
Phase 2: Data Complete 855K estimated records, 10 disease domains, guideline library (AAN/AHA/ILAE/MDS/IHS)
Phase 3: RAG Integration Complete Multi-collection parallel search with workflow-specific weight boosting, Claude streaming
Phase 4: Testing Complete 209 tests, all passing
Phase 5: UI + Demo Complete Streamlit UI on port 8534, NVIDIA theme, 5 demo scenarios validated

Remaining Work

Item Priority Effort
Real-time EEG pattern recognition integration Medium 1-2 weeks
NeuroImaging AI (MONAI integration for lesion detection) Low 2-3 weeks
Natural language EHR note parsing for auto-scale scoring Low 1 week
Integration with HCLS AI Factory landing page Low 1 hour

12. Relationship to HCLS AI Factory

The Neurology Intelligence Agent demonstrates the time-critical clinical extension of the HCLS AI Factory architecture. Neurological conditions — particularly acute stroke and status epilepticus — demand sub-minute decision support that combines imaging, clinical scales, genomics, and treatment guidelines simultaneously.

  • Same Milvus instance — 13 new owned collections alongside existing genomic_evidence (read-only)
  • Same embedding model — BGE-small-en-v1.5 (384-dim)
  • Same LLM — Claude via Anthropic API
  • Same hardware — NVIDIA DGX Spark ($4,699)
  • Same patterns — Pydantic models, BaseIngestPipeline, knowledge graph, query expansion

The neurogenomics integration connects Stage 1 (genomic variants in epilepsy genes like SCN1A, or movement disorder genes like GBA/LRRK2) directly to this agent's clinical workflows for genotype-informed treatment selection.


13. Credits

  • Adam Jones
  • Apache 2.0 License

Clinical Decision Support Disclaimer

The Neurology Intelligence Agent is a clinical decision support research tool for neurological evaluation. It is not FDA-cleared and is not intended as a standalone diagnostic device. Time-critical decisions (stroke, status epilepticus) must follow institutional protocols. All recommendations should be reviewed by qualified neurologists. Apache 2.0 License.