Genotype-Neurophenotype Modeling & Biomarker Discovery

Overview

This project investigates how genetic risk factors influence neurophysiological and cognitive phenotypes in healthy middle-aged peoples. The goal is to identify early biomarkers of neurodegenerative disease risk using multimodal data (genetics, EEG, fMRI, and psychometrics). A follow-up phase expands to clinical EEG data for disease classification.

• Primary Goal: Predict neurocognitive features and extract biomarkers from genotype and neuroimaging data.
• Expansion Goal: Apply findings to clinical data for diagnostic use cases (e.g., Alzheimer's vs. FTD classification).
• Technologies:

---

Phase 1: Genotype-Neurophenotype Modeling (Healthy Aging)

Dataset: PEARL-Neuro (N=192)

• Genetic markers: APOE, PICALM
• Neuroimaging: EEG (resting-state, cognitive tasks), fMRI
• Cognitive tasks: Sternberg Memory, MSIT
• Psychometrics: memory, intelligence, mood, personality
• Health data: blood tests, demographics

Objectives

• Model the relationship between genotype and neurophenotype
• Extract EEG/fMRI biomarkers associated with risk alleles
• Identify possible predictors of cognitive decline
• Apply feature attribution methods (e.g., SHAP, GradCAM) for biological interpretability

---

Phase 2: Clinical Expansion (Alzheimer’s & FTD EEG Classification)

Dataset: AHEPA General Hospital EEG Dataset

• Subjects: 36 with Alzheimer’s, 23 with FTD, 29 healthy controls
• 10–20 EEG, 500Hz, >19 hours of recordings

Objectives

• Build deep learning models to classify Alzheimer’s, FTD, and controls
• Validate if features from Phase 1 generalize to clinical data
• Use explainable AI to localize and understand neural changes in dementia
• Compare feature spaces between healthy aging and disease

---

Methodology

• Preprocessing: MNE, fMRIPrep, artifact rejection, source localization
• Feature Engineering:
  • EEG: Mean, Standard Deviation, Maximum value, Minimum value, Variance, Skewnewss, Kurtosis, Shannon entropy, Pwelwich, Band Power
  • For more info read
  • fMRI connectivity, ICA components
• Modeling:
  • ML: Logistic Regression, SVMs, Random Forests
  • DL: CNNs for EEG spectrograms, Transformer variants
  • Interpretability: SHAP, GradCAM, PCA
• Evaluation:
  • Cross-validation, AUROC, F1 score, model robustness tests

---

Data Sources

Dataset	Description	Usage
A Polish Electroencephalography, Alzheimer’s Risk-genes, Lifestyle and Neuroimaging (PEARL-Neuro) Database	N=192, includes EEG/fMRI, genetics, psychometrics	Phase 1 – Biomarker Discovery
AHEPA EEG	Alzheimer’s, FTD, Controls EEG recordings	Phase 2 – Clinical Classification

• https://www.nature.com/articles/s41597-024-03106-5 = https://www.medrxiv.org/content/10.1101/2024.08.05.24311520v1

---

Future Directions

• Integrate additional omics data (e.g., transcriptomics)
• Longitudinal modeling with follow-up neurodegeneration studies
• Investigate gene-environment interactions (stress, mood vs. risk)
• Train contrastive/self-supervised models on EEG/fMRI for generalized neurofeature learning