Multimodal integration strategies for clinical application in oncology

In clinical practice, a variety of techniques are employed to generate diverse data types for each cancer patient. These data types, spanning clinical, genomics, imaging, and other modalities, exhibit significant differences and possess distinct data structures. Therefore, most current analyses focus on a single data modality, limiting the potential of fully utilizing all available data and providing comprehensive insights. Artificial intelligence (AI) methods, adept at handling complex data structures, offer a powerful approach to efficiently integrate multimodal data. The insights derived from such models may ultimately expedite advancements in patient diagnosis, prognosis, and treatment responses. Here, we provide an overview of current advanced multimodal integration strategies and the related clinical potential in oncology field. We start from the key processing methods for single data modalities such as multi-omics, imaging data, and clinical notes. We then include diverse AI methods, covering traditional machine learning, representation learning, and vision language model, tailored to each distinct data modality. We further elaborate on popular multimodal integration strategies and discuss the related strength and weakness. Finally, we explore potential clinical applications including early detection/diagnosis, biomarker discovery, and prediction of clinical outcome. Additionally, we discuss ongoing challenges and outline potential future directions in the field.