Neurodiagnoses: AI-Powered CNS Diagnosis Framework

Neurodiagnoses is an open-source AI diagnostic framework for complex central nervous system (CNS) disorders. It integrates multi-modal biomarkers, neuroimaging, and AI-based annotation to improve precision diagnostics and advance research into the underlying pathophysiology of neurological diseases.

Ecosystem & Integrations

Neurodiagnoses AI operates across three core platforms:

• GitHub: Stores all scripts, pipelines, and model training workflows.
• EBRAINS: Provides HPC resources, neuroimaging, EEG, and biomarker data for AI training.
• Hugging Face: Hosts pre-trained AI models & datasets for easy deployment.

By utilizing these platforms, we ensure seamless integration of AI models, datasets, and neuroimaging resources.

AI-Assisted Diagnosis Approaches

Neurodiagnoses offers two complementary AI-powered diagnostic models:

1. Probabilistic Diagnosis (Differential Diagnosis):

   • Generates multiple possible diagnoses, each with an associated probability percentage.
   • Useful for differential diagnosis and ranking possible conditions.
   • Example Output:
     • 80% Prion Disease
     • 15% Autoimmune Encephalitis
     • 5% Neurodegenerative Disorder

2. Tridimensional Diagnosis (Structured):

   • Provides structured diagnostic outputs based on three axes:
     • Axis 1: Etiology (e.g., genetic, autoimmune, prion, vascular, toxic, inflammatory)
     • Axis 2: Molecular Markers (e.g., CSF biomarkers, PET findings, EEG patterns, MRI signatures)
     • Axis 3: Neuroanatomoclinical Correlations (e.g., regional atrophy, functional impairment, metabolic alterations)
   • Enhances precision medicine and biomarker-guided diagnosis.

Research Applications: CNS Computational Modeling

• Multi-omics Integration: Combines proteomics, genomics, lipidomics, and transcriptomics data.
• Personalized Simulations: Models disease progression tailored to individual patients.
• Computational Modeling: Aids in biomarker discovery and therapeutic target identification.

Project Components

• Data Processing (EBRAINS):

  • Stores and processes raw EEG, MRI, and biomarker data.
  • Feature extraction pipelines convert raw data into structured datasets.
  • Employs federated learning for secure multi-center AI training.

• AI Model Training & Hosting (Hugging Face):

  • Fine-tunes pre-trained models using Hugging Face resources.
  • Stores model artifacts on the Hugging Face Model Hub.
  • Provides access to models via the Hugging Face API.

• Codebase & Pipelines (GitHub):

  • Hosts all scripts and training workflows.
  • Implements continuous integration (CI/CD) for automated model updates and deployment.

Getting Started

To set up the project locally:

# Clone the repository
git clone https://github.com/Fundacion-de-Neurociencias/neurodiagnoses.git
cd neurodiagnoses

# Install dependencies
pip install -r requirements.txt

# Connect to Hugging Face
huggingface-cli login

# Train a Model
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

---