Evaluating Multiomics Integration Architectures for Training With Structured Missingness
Multimodal bioinformatics datasets are increasingly common in biomedical research, for tasks such as cancer subtyping and outcome prediction. It is feasible that data from a given patient, or even all data from a given institution, does not have coverage across all modalities; the available data is contingent on both the assay choice at the institution alongside technical aspects and associated drop-out. Consequently, algorithms for machine learning models must be tolerant to structured data missingness (or occurrence) for entire modalities when training. In this paper, we compare general strategies for training multimodal models in the context of structured modality missingness, employing suitable strategies for the stage of modality integration: early by concatenation of features, intermediate by max pooling of latent features, and late by aggregating model predictions probabilistically. We evaluate our strategies on a real-world bioinformatics dataset for the task of breast cancer subtyping, constructing a range of structured missingness scenarios. We highlight that, despite their inability to learn cross-modality interactions, late integration models outperform against early and intermediate integration strategies across a range of scenarios according to the level and nature of missingness. Logistic regression models, although simple, also outperform neural networks within the same settings. Fundamentally, we show that understanding the structure of missingness within a dataset is necessary when selecting a method of integration, and that simple models and approaches should not be dismissed when working with structured missingness.