Overview

This section describes the step-by-step process used in the Neurodiagnoses project to develop a novel diagnostic framework for neurological diseases. The methodology integrates artificial intelligence (AI), biomedical ontologies, and computational neuroscience to create a structured, interpretable, and scalable diagnostic system.

1. Data Integration

Data Sources

• Biomedical Ontologies:
  • Human Phenotype Ontology (HPO) for phenotypic abnormalities.
  • Gene Ontology (GO) for molecular and cellular processes.
• Neuroimaging Datasets:
  • Example: Alzheimer’s Disease Neuroimaging Initiative (ADNI), OpenNeuro.
• Clinical and Biomarker Data:
  • Anonymized clinical reports, molecular biomarkers, and test results.

Data Preprocessing

1. Standardization: Ensure all data sources are normalized to a common format.
2. Feature Selection: Identify relevant features for diagnosis (e.g., biomarkers, imaging scores).
3. Data Cleaning: Handle missing values and remove duplicates.

2. AI-Based Analysis

Model Development

• Embedding Models: Use pre-trained models like BioBERT or BioLORD for text data.
• Classification Models:
  • Algorithms: Random Forest, Support Vector Machines (SVM), or neural networks.
  • Purpose: Predict the likelihood of specific neurological conditions based on input data.

Dimensionality Reduction and Interpretability

• Leverage DEIBO (Data-driven Embedding Interpretation Based on Ontologies) to connect model dimensions to ontology concepts.
• Evaluate interpretability using metrics like the Area Under the Interpretability Curve (AUIC).

3. Diagnostic Framework

Axes of Diagnosis

The framework organizes diagnostic data into three axes:

1. Etiology: Genetic and environmental risk factors.
2. Molecular Markers: Biomarkers such as amyloid-beta, tau, and alpha-synuclein.
3. Neuroanatomical Correlations: Results from neuroimaging (e.g., MRI, PET).

Recommendation System

• Suggests additional tests or biomarkers if gaps are detected in the data.
• Prioritizes tests based on clinical impact and cost-effectiveness.

4. Computational Workflow

1. Data Loading: Import data from storage (Drive or Bucket).
2. Feature Engineering: Generate derived features from the raw data.
3. Model Training:
   • Split data into training, validation, and test sets.
   • Train models with cross-validation to ensure robustness.
4. Evaluation:
   • Metrics: Accuracy, F1-Score, AUIC for interpretability.
   • Compare against baseline models and domain benchmarks.

5. Validation

Internal Validation

• Test the system using simulated datasets and known clinical cases.
• Fine-tune models based on validation results.

External Validation

• Collaborate with research institutions and hospitals to test the system in real-world settings.
• Use anonymized patient data to ensure privacy compliance.

6. Collaborative Development

The project is open to contributions from researchers, clinicians, and developers. Key tools include:

• Jupyter Notebooks: For data analysis and pipeline development.
  • Example: probabilistic imputation
• Wiki Pages: For documenting methods and results.
• Drive and Bucket: For sharing code, data, and outputs.
• Related projects: For instance: Beyond the hype: AI in dementia – from early risk detection to disease treatment

7. Tools and Technologies

• Programming Languages: Python for AI and data processing.
• Frameworks:
  • TensorFlow and PyTorch for machine learning.
  • Flask or FastAPI for backend services.
• Visualization: Plotly and Matplotlib for interactive and static visualizations.
• EBRAINS Services:
  • Collaboratory Lab for running Notebooks.
  • Buckets for storing large datasets.