Skip to Main Content

An official website of the United States government

AI-based Pan-Cancer Survival Outcome Prediction and Biomarker Discovery

Principal Investigator
Name
Jana Lipkova
Degrees
PH.D.
Institution
University of California, Irvine
Position Title
Assistant Professor
Email
About this CDAS Project
Study
PLCO (Learn more about this study)
Project ID
PLCOI-1715
Initial CDAS Request Approval
Oct 17, 2024
Title
AI-based Pan-Cancer Survival Outcome Prediction and Biomarker Discovery
Summary
We have developed a pan-cancer AI model for improving survival outcome prediction in oncology. Our model utilize histopathology images and related clinical data, with the capacity to handle missing or incomplete information, making it adaptable for deployment in diverse medical settings with variable data availability. We have further implemented interpretability methods to explore the predictive features, facilitating the exploration of prognostic biomarkers in complex medical data. To further validate and rigorously evaluate our model, we wish to gain access to the Prostate, Lung, Colorectal, and Ovarian (PLCO) pathology images, along with the associated clinical, pathological, and follow-up data. The ultimate goal is to provide a robust prognostic tool for survival prediction and to identify biomarkers related to patient outcomes, both within and across cancer subtypes.
Aims

Aim 1: Benchmark Existing Foundational Models in Pathology for Survival Prediction
- This aim focuses on identifying the best foundational model for preprocessing histopathology data for survival prediction. The PLCO dataset will be utilized for external validation of the tested foundational models.

Aim 2: Multimodal Pan-Cancer Model for Survival Outcome Prediction with Missing Modalities
- Building on the best histology model identified in Aim 1, we will enhance our multimodal fusion model to integrate histopathology and available clinical data to improve survival outcome prediction.
- We will evaluate the robustness of the proposed model in handling missing data both during training and inference.
- The PLCO cohort will be used for fine-tuning and validation of the model.

Aim 3: Prognostic Biomarker Discovery Within and Across Cancer Subtypes
- Leveraging the fine-tuned models from Aims 1 and 2, along with interpretability methods, we will explore potential prognostic biomarkers associated with patient outcomes across various cancer subtypes.

Collaborators

Jana Lipkova (PI), Sami Yavuz, Richard Chen, Jerry Lou, Ethan Ton Niu, Sabri Kahya (all with University of California, Irvine)