Conformalized Multiview Learning

Abstract

SO_SCPLOWUMMARYC_SCPLOWThere is an overwhelmingly large body of literature and numerous algorithms already available on "multimodal AI", based on different modeling and fusion paradigms. As multimodal AI becomes fully integrated into healthcare applications, the ability to say "Im not sure" or "I dont know" when uncertain is a necessary capability for safe clinical deployment, requiring the augmentation of current multimodal AI techniques. Our goal in this work is not to add yet another method to the growing toolbox of multimodal AI, but rather to (i) clarify how existing supervised multimodal AI methods can be understood through the lens of uncertainty quantification, and (ii) develop a unified framework that provides valid and interpretable uncertainty estimates for multimodal AI practitioners. We introduce Coracle, a conformalized framework for multimodal AI that seamlessly adapts to early, late, and intermediate fusion. Coracle provides theoretical marginal confidence guarantees and achieves valid finite-sample coverage without relying on distributional assumptions. Through extensive simulations and analyses of diverse biomedical multiview datasets, we demonstrate that Coracle consistently outperforms existing non-conformalized methods in both uncertainty estimation and calibration. By equipping published multiview models with valid, interpretable confidence measures, Coracle advances the development of trustworthy, uncertainty-aware statistical tools for biomedical decision support. Coracle is publicly available as an open-source R package at https://github.com/himelmallick/Coracle.