On the increase of predictive performance with high-level data fusion.

Area of Science:

• Multivariate data analysis
• Bioinformatics
• Machine learning

Background:

• Data fusion combines diverse data sources for improved classification.
• Fusion can occur at low, medium, or high levels.
• High-level fusion integrates model outputs.

Purpose of the Study:

• Investigate the predictive performance of high-level data fusion.
• Analyze the impact of within-group correlations on fusion effectiveness.
• Evaluate fusion performance with simulated and real-world data.

Main Methods:

• Partial least squares (PLS) applied to individual datasets.
• Gaussian distribution fitting to estimated class responses.
• Simulation study with two datasets and two classes.
• Analysis of within-group correlations and predictive abilities.

Main Results:

• Error rate of high-level fusion is less than or equal to the best individual model.
• Negative within-group correlations consistently improve predictive performance.
• Fusion benefits increase when combining non-predictive and discriminative classifiers, especially with strong correlations.

Conclusions:

• High-level data fusion offers a powerful approach for classification tasks.
• Understanding within-group correlations is crucial for optimizing fusion strategies.
• The method demonstrates practical applicability across various data types, including metabolomics.