Advanced data fusion: Random forest proximities and pseudo-sample principle towards increased prediction accuracy and...

Area of Science:

• Life Sciences
• Bioinformatics
• Computational Biology

Background:

• Biological sample analysis often requires integrating data from multiple complementary sources.
• Traditional data fusion methods (low-level, mid-level, high-level) face challenges with increasing data complexity and volume.
• Advanced fusion approaches are needed for comprehensive biological profiling.

Purpose of the Study:

• To present and evaluate an advanced data fusion approach: proximities stacking.
• To compare the performance of proximities stacking against traditional fusion methods and individual data platforms.
• To demonstrate the utility of proximities stacking in classifying Crohn's disease patient samples.

Main Methods:

• Developed a novel data fusion approach, proximities stacking, utilizing random forest proximities and the pseudo-sample principle.
• Applied the method to four diverse biological data platforms (faecal microbiome, blood, blood headspace, exhaled breath) from Crohn's disease patients.
• Compared classification performance using sensitivity, specificity, and principal component analysis visualization.

Main Results:

• Proximities stacking significantly outperformed mid-level, high-level fusion, and individual platform predictions in classifying patient samples.
• The pseudo-sample principle enabled identification of key variables and inter-platform relationships.
• The study demonstrated that more data does not always equate to better results, emphasizing careful fusion strategy.

Conclusions:

• Proximities stacking offers a superior alternative to traditional data fusion techniques in life sciences.
• This advanced method addresses limitations of existing fusion strategies and provides deeper biological insights.
• Careful consideration of data integration strategies is crucial for effective biological data analysis.