Cancer prognosis prediction involves forecasting the development process and ultimate outcomes of cancer via patients’ clinical presentations, pathology, and other related information. The significance of cancer prognosis prediction lies in its ability to effectively prevent over-treatment and the wastage of medical resources while providing a scientific reference for medical decision-making for physicians.
The project utilizes multimodal data from the UK Biobank to establish a multimodal foundation model handling missing modality through prompt learning, employing two stages of self-supervised pre-training and task fine-tuning, which allows the model to learn information-rich multimodal representations before learning task-specific representations. Thus, strong representations further lead to better cancer prognosis prediction performance.
Over a two-year duration, the project will build a cancer prognosis prediction multimodal foundation model that generates multimodal solid representations and ensures more accurate prediction and model generalizability. The model can guide tailored treatment strategies by providing clinicians with more accurate and personalized prognostic assessments, maximizing therapeutic efficacy and minimizing unnecessary interventions. Moreover, the interpretability analysis via the multimodal foundation model can contribute to advancing the understanding of cancer biology, potentially uncovering novel prognostic markers and informing future research and therapeutic innovations.
We expect the multimodal foundation model generated by this project will be valuable research resources and improve the lives of cancer patients worldwide.
