Hierarchical confounder discovery in the experiment-machine learning cycle.

Area of Science:

• Bioinformatics
• Machine Learning
• Statistical Modeling

Background:

• High-dimensional bioscience datasets often contain confounding variables that hinder accurate machine learning (ML) analysis.
• Nested hierarchies in biological data obscure confounder origins, rendering traditional bias mitigation ineffective.
• Ensuring ML models extract true insights rather than biases is crucial for scientific validity.

Purpose of the Study:

• To introduce a novel statistical method for detecting hierarchical confounding effects in complex datasets.
• To demonstrate the efficacy of this method in both raw data and ML-derived embeddings.
• To enhance the robustness and interpretability of ML models in bioscience research.

Main Methods:

• Development of the non-parametric rank-to-group (RTG) score.
• Application of RTG scores to identify hierarchical confounder effects.
• Validation using public biomedical image and multi-phenotypic biological datasets.

Main Results:

• RTG scores successfully identified hierarchical confounder effects where linear methods failed.
• Unreported experimental design effects were discovered in a public biomedical image dataset.
• Crossmodal correlated variability was identified in a multi-phenotypic biological dataset using RTG scores.

Conclusions:

• The RTG score is a powerful tool for identifying hierarchical confounding in bioscience data.
• This method improves the reliability of ML models by addressing hidden biases.
• RTG scores facilitate more robust experimental design and analysis cycles.