Harmonization with Flow-Based Causal Inference.

Area of Science:

• Medical data analysis
• Machine learning
• Causal inference

Background:

• Heterogeneity in medical data from diverse sites and protocols hinders machine learning model generalization.
• Accurate prediction requires addressing data variability and confounding factors like site, gender, and age.

Purpose of the Study:

• To harmonize heterogeneous medical data using a normalizing-flow-based counterfactual inference method.
• To improve the cross-domain generalization of machine learning models on medical datasets.

Main Methods:

• Leveraging a normalizing-flow-based approach for counterfactual inference within a structural causal model (SCM).
• Modeling observed effects (e.g., brain MRI data) with confounders (site, gender, age) and exogenous noise.
• Exploiting the bijection of normalizing flows for data harmonization via posterior inference and intervention.

Main Results:

• The proposed method demonstrated superior cross-domain generalization compared to state-of-the-art algorithms on large, real-world medical datasets.
• Generated confounder-independent data showed high quality in downstream regression and classification tasks.

Conclusions:

• Normalizing-flow-based counterfactual inference offers an effective strategy for harmonizing heterogeneous medical data.
• This approach enhances the robustness and generalizability of machine learning models in medical applications.