Multimodal AI Fusion Model for Early Detection for Pancreatic Cancer

Pancreatic ductal adenocarcinoma (PDAC), accounting for 90% of all pancreatic cancers, is among the deadliest, due largely to late-stage diagnosis and the aggressive nature of the disease. The critical challenge lies in early detection, which is currently not viable for the general population due to low annual incidence and a significant risk of false positives even with highly specific tests. While current risk assessment tools rely on static factors such as age, obesity, and diabetes, recent studies suggest the potential for imaging biomarkers derived from pre-cancerous computed tomography (CT) scans to predict PDAC. Our project aims to develop a comprehensive and scalable risk prediction model that fuses imaging and non-imaging data to enable early detection of PDAC in asymptomatic individuals. The model, termed "PRECISE" (PancREas Cancer multImodal riSk prEdiction), will employ novel algorithmic adversarial debiasing techniques to ensure fairness, meaning it should perform accurately across different demographic and socioeconomic subgroups. In Aim 1, we will develop deep learning models that segment imaging biomarkers from abdominal CTs, applying adversarial debiasing to ensure fair representation across diverse patient factors and image acquisition techniques. Validation will be done using data from Mayo Clinic, Cornell University, and UCSF. Aim 2 involves the creation of the PRECISE fusion model. It will combine imaging biomarkers from CTs with clinical data from electronic medical records (EMRs) to predict the risk of PDAC. We will employ a graph neural network model to capture the semantic relations between multimodal data. The model's prognostic performance will be compared with baseline models. In Aim 3, we plan to deploy and evaluate the PRECISE model prospectively across disparate geographical sites. The model's performance will be assessed by comparing its predictions with patient outcomes collected at regular intervals. This proposal's overall goal is to create a fair and effective PDAC risk prediction tool, PRECISE, that leverages both imaging and non-imaging data to calculate unbiased risk estimates. If successful, our scalable automated risk stratification could potentially transform PDAC early detection, enabling opportunistic screening for patients undergoing routine abdominopelvic CT scans for non-pancreatic cancer indications. This could significantly improve PDAC survival rates by enabling earlier intervention and treatment.