Robust Multimodal Healthcare with Missing Modalities

PhD Qualifying Examination


Title: "Robust Multimodal Healthcare with Missing Modalities"

by

Mr. Fengtao ZHOU


Abstract:

The integration of diverse data modalities has become a cornerstone of
precision medicine. However, real-world multimodal healthcare datasets are
frequently incomplete, with critical modalities missing due to economic
constraints, patient compliance, technical failures, or privacy regulations.
This pervasive incompleteness severely undermines the robustness,
generalizability, and clinical utility of current AI systems, which typically
assume full modality availability during both training and inference. In this
survey, we systematically review methodological approaches to address missing
modalities in multimodal healthcare, categorizing them into two broad
paradigms: imputation-based and imputation-free strategies.
Imputation-based methods reconstruct missing modalities either at the raw
data level, at the latent feature level, or through sample retrieval from
reference cohorts. In contrast, imputation-free approaches operate directly
on available modalities, leveraging flexible architectures and specialized
training protocols to enhance robustness without explicit data synthesis. We
further discuss critical challenges, including the gap between simulated and
clinically realistic missingness patterns, the lack of standardized
benchmarks, limited interpretability under incompleteness, and heavy
reliance on labeled multimodal data. Finally, we outline promising
directions toward equitable, robust, and deployable multimodal AI in
real-world clinical settings.


Date:                   Monday, 13 January 2026

Time:                   2:00pm - 4:00pm

Venue:                  Room 4472
                        Lifts 25/26

Committee Members:      Dr. Hao Chen (Supervisor)
                        Dr. Dan Xu (Chairperson)
                        Dr. Xiaomin Ouyang