OXA-MISS: A Robust Multimodal Architecture for Chemotherapy Response Prediction under Data Scarcity

In clinical oncology, tumor heterogeneity, data scarcity, and missing modalities are pervasive issues that significantly hinder the effectiveness of predictive models. Although multimodal integration of Whole Slide Imaging (WSI) and molecular data has shown promise in predicting overall survival (OS), current approaches often struggle when dealing with scarce and incomplete multimodal datasets, a scenario that reflects the norm rather than the exception in real-world clinical practice, especially in tasks like chemotherapy resistance prediction, where data collection is substantially more challenging than for OS. Accurately identifying patients who will not respond to chemotherapy is a critical clinical need, enabling the timely redirection to alternative therapeutic strategies and avoiding unnecessary toxicity. Hence, this paper introduces OXA-MISS, a novel multimodal model for chemotherapy response prediction designed to handle missing modalities. In the task of chemotherapy response prediction in ovarian cancer, OXA-MISS achieves a 20% absolute improvement in AUC over state-of-the-art models when trained on scarce and incomplete WSI–transcriptomics datasets. To evaluate its generalizability, we benchmarked OXA-MISS on OS prediction across three TCGA cancer types under both complete and missing-modality conditions. In these settings, the results demonstrate that OXA-MISS achieves performance comparable to that of state-of-the-art models. In conclusion, the proposed OXA-MISS is shown to be effective in OS prediction tasks, while substantially improving predictive accuracy in realistic clinical settings, such as the proposed prediction of chemotherapy response. The code for OXA-MISS is publicly available at https://github.com/AI-BioInformatics/OXA-MISS.