When Are Multimodal Predictions Biologically Supported? A Diagnostic Evaluation Framework

Multimodal models in oncology can produce accurate predictions, but accurate prediction does not reveal whether the model has learned biology that is shared across modalities, biology confined to one modality, or spurious correlations that reflect confounders rather than genuine biology. We introduce DECAT, a model-agnostic post-hoc evaluation framework that classifies multimodal representations into four diagnostic scenarios for a given task and modality, using five null-referenced metrics and a rule-based decision procedure. The framework operates on learned representations, requires no knowledge of which specific confounder is present, and returns indeterminate when the evidence is insufficient. We validate DECAT on synthetic data across four multimodal model classes (over 2,500 trained representations) and on real data from 8,979 TCGA patients, evaluating both multimodal embeddings and five pretrained pathology foundation models. Entangled models (e.g., CLIP) achieve near-perfect shared biology detection but falsely claim shared biology in the majority of cases where it is absent on real foundation model embeddings. This false claim rate increases with confound strength so that larger cohorts and stronger representations produce more confident but still incorrect diagnoses. Applied to both multimodal TCGA embeddings and five pathology foundation models without paired RNA, DECAT detects confounding invisible to AUROC without requiring the confounder labels, as confirmed by post-hoc stratification.

Coupling Deep Imputation with Multitask Learning for Downstream Tasks on Genomics Data

Genomics data such as RNA gene expression, methylation and micro RNA expression are valuable sources of information for various clinical predictive tasks. For example, predicting survival outcomes, cancer histology type and other patients' related information is possible using not only clinical data but molecular data as well. Moreover, using these data sources together, for example in multitask learning, can boost the performance. However, in practice, there are many missing data points which leads to significantly lower patient numbers when analysing full cases, which in our setting refers to all modalities being present. In this paper we investigate how imputing data with missing values using deep learning coupled with multitask learning can help to reach state-of-the-art performance results using combined genomics modalities, RNA, micro RNA and methylation. We propose a generalised deep imputation method to impute values where a patient has all modalities present except one. Interestingly enough, deep imputation alone outperforms multitask learning alone for the classification and regression tasks across most combinations of modalities. In contrast, when using all modalities for survival prediction we observe that multitask learning alone outperforms deep imputation alone with statistical significance (adjusted p-value 0.03). Thus, both approaches are complementary when optimising performance for downstream predictive tasks.