Evaluating an information theoretic approach for selecting multimodal data fusion methods.

Area of Science:

• Biomedical informatics
• Machine learning
• Computational biology

Background:

• Precision health initiatives increasingly integrate diverse data types (radiology, pathology, genomics, clinical) for enhanced diagnostic and prognostic accuracy.
• Current methods for selecting multimodal datasets and modeling approaches are often empirical, lacking a theoretical foundation.
• Partial Information Decomposition (PID) offers a framework to theoretically understand multimodal data interactions, quantifying redundancy, uniqueness, and synergy.

Purpose of the Study:

• To evaluate the efficacy of existing PID-based metrics in understanding multimodal data interactions across a broader range of biomedical datasets.
• To investigate the impact of parameter selection on PID metric calculations.
• To propose improvements to PID metrics for more reliable application in multimodal data fusion for precision health.

Main Methods:

• Applied four PID-based metrics to seven distinct modality pairs across four independent biomedical cohorts.
• Evaluated downstream machine learning model performance for prognostic prediction (overall survival, recurrence) in non-small cell lung cancer, prostate cancer, and glioblastoma.
• Compared trends between PID metric values and predictive model performance.

Main Results:

• PID metrics provided informative insights but did not consistently predict optimal multimodal data fusion strategies.
• Consistency between PID values and model performance varied: 0% for three pairs, 66%-89% for three pairs, and 100% for one pair.
• Two key improvements to PID metrics were proposed: optimal parameter determination and uncertainty estimation.

Conclusions:

• Current PID metrics require refinement to accurately estimate multimodal data interactions and reliably guide data fusion for predictive modeling.
• Proposed improvements aim to enhance the utility of PID metrics as a robust tool in precision health.
• Further research is needed to optimize PID metrics for clinical applications.