AI-Powered Epilepsy Diagnosis Platform Reducing Diagnostic Time Through Multimodal Data Processing

australian_epilepsy_project · Healthcare · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Clinical epilepsy diagnosis and treatment planning

The Australian Epilepsy Project needed to make specialist epilepsy diagnostic expertise available nationally while processing multimodal clinical data safely for decision support.

Constraints:

• The platform handles high-stakes healthcare data and clinical decisions
• Inputs include MRI scans, medical histories, assessments, and other multimodal patient data
• AI processes must fit regulatory, ethics, and clinician-verification requirements

Architecture

The Australian Epilepsy Project runs a cloud-based AWS decision-support platform that standardizes multimodal patient data, stores unstructured data in S3 and structured data in RDS/data nexus, uses Textract to extract text from medical PDFs, applies Bedrock/Mistral and SageMaker/OpenChat models for structured conversion and summaries, uses PGVector/RAG for medical-history exploration, and returns reports through a secure clinician portal with neurologist feedback.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  secure-clinician-portal["secure-clinician-portal"]
  standardized-data-acquisition["standardized-data-acquisition"]
  s3-unstructured-storage["s3-unstructured-storage"]
  textract-medical-ocr["textract-medical-ocr"]
  rds-data-nexus["rds-data-nexus"]
  bedrock-mistral-text-structuring["bedrock-mistral-text-structuring"]
  medical-history-rag["medical-history-rag"]
  sagemaker-language-laterality-model["sagemaker-language-laterality-model"]
  openchat-report-generator["openchat-report-generator"]
  neurologist-sense-check["neurologist-sense-check"]
  secure-clinician-portal --> rds-data-nexus
  standardized-data-acquisition --> s3-unstructured-storage
  standardized-data-acquisition --> rds-data-nexus
  s3-unstructured-storage --> textract-medical-ocr
  textract-medical-ocr --> medical-history-rag
  rds-data-nexus --> bedrock-mistral-text-structuring
  rds-data-nexus --> sagemaker-language-laterality-model
  rds-data-nexus --> openchat-report-generator
  sagemaker-language-laterality-model --> neurologist-sense-check
  openchat-report-generator --> secure-clinician-portal

Components

Data flow

Patients enter a standardized protocol covering MRI, neuropsychology, genetics, and medical history. Unstructured files and structured data are stored in AWS; Textract extracts PDF text, Bedrock/Mistral converts medical free text, the RDS data nexus tracks model versions and processes, language-laterality ML processes MRI-derived features, medical-history text is embedded for RAG questions, and neurologists review baseline reports and provide feedback on diagnostic influence.

Why this architecture

AEP is explicitly decision support, not autonomous diagnosis: the system standardizes data nationally, centralizes it in a data nexus, runs multiple AI/ML workflows over different modalities, and returns reports to neurologists for treatment decisions and sense-checking. PGVector fit because it integrates with the existing database layer instead of adding a separate vector stack.

Models used: Mistral · Len AI Research LLM · OpenChat 3.6 · machine-learning language-laterality models

Scale

• Diagnosis time for language-area mapping can be reduced by 70%.
  “epilepsy care. How a IML on AWS cuts diagnosis time by 70%.”transcript, stated
• Roughly one in ten people will have a seizure, and about half of those are diagnosed with epilepsy.
  “one in 10 people will have a seizure in their lifetime? And roughly half of those will go on to be diagnosed with epilepsy.”transcript, stated
• Patients are followed up two years later.
  “we also follow them up two years later.”transcript, stated
• Measured outcomes include 9% work-productivity improvement, 8% quality-of-life improvement, and 8% seizure reduction.
  3 supporting quotes

  “we have a 9% improvement in work productivity”transcript, stated

  “there we already see an 8% improvement which is amazing.”transcript, stated

  “our patients report a decrease of 8% in seizures.”transcript, stated
• MRI imaging delivered a 10% higher brain-lesion detection rate.
  “a 10% higher detection rate of brain lesions based on our very advanced MRI imaging”transcript, stated
• The neuropsychological assessment lasts two hours.
  “Second is an in-house developed neurossychological assessment that again everyone undergoes for two hours.”transcript, stated
• The laterality model automatically classifies 70% of cases, leaving 30% for in-house neurologists.
  “in 70% of cases, you can do an automata automated classification”transcript, stated

  “the the other 30% still require our in-house neurologists to look at the scans and to make the call.”transcript, stated

Alternatives considered (roads not taken)

Siloed hospital-center expertise

Rejected because: AEP moves specialist diagnostic capabilities to a cloud platform available to neurologists across the country.

Manual-only language-area MRI review

Rejected because: The ML model automates classification for most cases while preserving neurologist review for the remaining cases and sense checks.

Autonomous AI diagnosis

Rejected because: The system is framed as decision support and asks neurologists whether the report influenced their final diagnosis.

Separate specialized vector database

Rejected because: The team tested other vector databases but chose PGVector because it fit the existing Amazon database setup.

Patterns

rag-retrieval (clinical-source-attribution)

The Medical History Explorer lets neurologists ask questions against patient histories and inspect the source paragraphs and PDFs behind answers.

model-routing (task-specific-clinical-models)

AEP uses different models for free-text numeric extraction, medical sentence embeddings, and patient summary generation.

structured-output-contract (clinical-data-normalization)

Free-text patient responses are converted into structured numeric database fields for analysis.

safety-policy-enforcement (clinical-guardrails)

Patient summary generation uses a model chosen for lower hallucination in medical context plus custom guardrails.

human-in-the-loop-gate (clinical-decision-support)

The platform surfaces analyses and reports to neurologists, who retain treatment decision authority.

Tradeoffs

Which vector store to use for clinical RAG.

Chose PGVector integrated with the existing Amazon database over introducing a separate specialized vector database.

Rationale: Operational integration with the existing database outweighed adding a new standalone vector component.

How to frame AI in clinical workflow.

Chose decision support for neurologists over autonomous diagnosis.

Rationale: The platform makes analysis available to clinicians but does not replace their treatment decisions.

Gotchas (what bit them)

Healthcare AI adoption depends on the data foundation, not only the models.

Root cause: The platform must collect, transform, version, and keep multiple clinical data modalities current before AI can be useful.

Resolution: AEP uses a centralized RDS data nexus for model versioning, orchestration, and keeping data updated.

Clinical AI requires regulatory and ethics governance.

Root cause: Outputs are used in healthcare decision-support contexts with patient safety implications.

Resolution: The AI processes are listed with Australia's TGA and governed under clinical trial ethics.

Automated language-area classification cannot cover every surgical case.

Root cause: Functional MRI language mapping remains clinically complex and not all cases are suitable for automated classification.

Resolution: The platform automates classification in many cases while keeping neurologists in the review and decision loop.

Evaluation

AEP measures clinical outcomes over two years, reports neurologist-facing benefits, and benchmarks model choices for specific medical tasks.

Metrics: diagnosis time reduction · lesion detection rate · work productivity · quality of life · seizure rate reduction · hallucination behavior

Outcomes

• AEP reports 70% diagnosis-time savings for a pre-surgical language-area use case.
• The project reports 10% higher brain lesion detection, 9% work-productivity improvement, 8% quality-of-life improvement, and 8% seizure-rate reduction.
• The clinician portal serves more than 100 clinicians in Australia.

Each participant contributes about 8GB of data, and the platform makes specialist diagnostic expertise available through a secure cloud portal.