Machine-Learning Foundation Model For The Early Detection Of Neurodegenerative Diseases

A machine learning technique that identifies biomarkers from neuroimaging data, allowing for identification and treatment of populations vulnerable to neurodegenerative diseases.
Problem:
Neurodegenerative diseases such as Alzheimer’s disease (AD) impact an estimated twenty-four million people worldwide, with prevalence increasing with global human life expectancy. Machine learning (ML) represents a powerful tool for rapid processing of large volumes of data, such as MRI imaging, to extract biomarkers for AD and other neurodegenerative diseases, allowing for early detection and treatment of AD. Current ML approaches suffer from instability and are difficult to cross-validate findings across various datasets. These models are also often “black boxes,” with no conclusive explanations of their outputs, and often do not return anatomically interpretable biomarker results. These attributes can make it difficult for clinicians to interpret and validate the biological relevance of identified biomarkers.
Solution:
An AI covariance neural network (VNN) platform technology that is not limited to specific health conditions. VNN is inherently stable compared to other approaches, and returns anatomically interpretable results linked to specific regions, allowing for easy clinician validation. The VNN model outputs do not depend on the size or processing technique used to process imaging data, allowing for a wide range of data input sets.
Technology:
A VNN trained exclusively on healthy neuroimaging MRI data to extract biomarkers for any neurodegenerative condition. The covariance matrices employed by the VNN model are characterized by the covariance between sets of data elements, thus incorporating the redundancy of a given dataset and stability in the model. The model predicts the “brain age” of an individual using brain volumetric and thickness data derived from structural MRI scans. Large gaps between brain age and chronological age are indicators of accelerated aging and increased vulnerability to adverse health conditions. VNNs are pre-trained to exploit the anatomical covariance matrix to yield a level of biological explainability and insight into a neurodegenerative condition typically not present in ML models for biomarker extraction.
Advantages:

VNN trained in healthy neuroimaging data allows for generalizable biomarker extraction
The method identifies contributing brain regions, allowing for explainable anatomical interpretability to results
ML model is inherently scale-free, allowing for dataset of varying sizes and manners of processing analysis
Scale-free technique is stable to perturbations in the model input, increasing confidence and reproducibility to results

Stage of Development:

Proof of Concept

A.

B.

Figure A.: Flow diagram illustrating the application of same VNN model for identifying biomarkers and brain age indicative of neurodegeneration in two distinct scenarios. Figure B.: Examples of the application of same VNN model for identifying biomarkers and brain age indicative of neurodegeneration in a healthy individual and three individuals with neurodegenerative conditions, which include aphasia, Alzheimer’s disease, and mild cognitive impairment.
Intellectual Property: