Dimensionality reduction-based fusion approaches for imaging and non-imaging biomedical data: concepts, workflow, and...

Area of Science:

• Biomedical data science
• Medical informatics
• Computational biology

Background:

• Biomedical data is increasingly multi-modal, multi-protocol, and multi-scale.
• There is a need for quantitative tools to integrate diverse data channels for improved disease characterization.
• Existing data fusion methods often struggle with differences in dimensionality and scale across modalities.

Purpose of the Study:

• To identify optimal methodological choices for building data fusion techniques.
• To focus on data fusion approaches employing dimensionality reduction (DR).
• To quantitatively evaluate existing DR-based data fusion instantiations.

Main Methods:

• Evaluated 4 non-overlapping DR-based data fusion instantiations.
• Applied methods to 3 distinct biomedical applications (prostate cancer, Alzheimer's Disease).
• Integrated data from MRI, MR spectroscopy, histopathology, and mass spectrometry.

Main Results:

• Weighted multi-kernel DR-based fusion achieved the highest predictive performance (AUC > 0.8).
• Non-optimized DR methods yielded the worst predictive performance across applications.
• Data fusion methods may fail if individual modalities have poor discriminability.

Conclusions:

• Methodological choices in data fusion are critical for improving predictive ability.
• DR-based fusion, particularly with kernel representations, shows significant potential.
• Accounting for feature space sparsity and noise is essential for effective data fusion.