Discriminative margin-sensitive autoencoder for collective multi-view disease analysis | JoVE Visualize

Area of Science:

Computational biology
Medical imaging analysis
Machine learning for healthcare

Background:

Medical prediction relies on integrating diverse data sources like bioimages and clinical features.
Learning from multiple heterogeneous data views (multi-view learning) is crucial but challenging for disease understanding.
Existing methods often process single data views independently, limiting comprehensive analysis.

Purpose of the Study:

To propose a novel Margin-Sensitive Autoencoder (MSAE) framework for automated Alzheimer's disease (AD) diagnosis and protein fold recognition.
To develop a method that collaboratively explores complementary properties of multi-view bioimage features.
To enhance multi-view disease understanding by constructing a discriminative semantic space.

Main Methods:

Developed a semantic-sensitive autoencoder with an encoder-decoder paradigm to project multi-view features into a common latent space.
Employed a margin-scalable support vector machine for a flexible semantic space to boost model discriminability.
Utilized a self-adjusting learning scheme for adaptive weighting of different views and correntropy induced metric for robust regularization against outliers.
Optimized the framework using half-quadratic minimization and an alternating learning strategy for closed-form solutions.

Main Results:

The MSAE framework achieved superior performance in both binary and multi-class classification for Alzheimer's disease diagnosis on ADNI datasets.
Evaluations on protein folds demonstrated highly encouraging performance in protein structure recognition.
The proposed method outperformed existing state-of-the-art methods in both application domains.

Conclusions:

The MSAE framework effectively integrates multi-view data for enhanced medical prediction tasks.
The method shows significant potential for improving automated disease diagnosis and biological structure recognition.
The proposed approach offers a robust and effective solution for leveraging complementary information across multiple data modalities.