Multimodal survival and recurrence prediction in head and neck oncology

Clinical endpoints, such as predicting survival and recurrence risk, are critical in guiding oncological treatment decisions. In practice, clinicians combine diverse patient information to tailor treatment strategies. However, classical AI tools often fall short by relying on a single source of data, limiting their utility in complex clinical scenarios, especially in treatment planning and decision-making. The novel HANCOCK dataset addresses this gap by offering a comprehensive multimodal resource, encompassing six data types: clinical data, pathology reports, histopathology images (primary tumor and lymph nodes), tissue microarrays, tabular blood data, and free-text surgery reports. With data from 763 head and neck cancer patients collected at a single center to minimize technical biases, HANCOCK places the focus squarely on patient-specific insights. This challenge invites participants to harness the full potential of these diverse modalities, maximizing their predictive power for critical endpoints like survival and recurrence prediction for precision oncology. Challenge participants need to derive strategies for how to fuse the different data streams and combine large image datasets with structured and free text data in order to perform two binary classification tasks.

The two tasks use the same data to (i) determine the 5-year-survival to potentially help in therapy and adjuvant therapy prediction and (ii) estimate the recurrence risk to ensure timely follow-up intervals for monitoring purposes.