Multi-modal dynamic risk prediction of thrombosis and bleeding in patients with cancer

Venous thromboembolism (VTE) is the second-leading cause of non-cancer death in ambulatory cancer patients receiving chemotherapy. Despite evidence supporting the prescription of low dose direct oral anticoagulants (DOACs) in ambulatory patients with cancer who are high-risk for VTE and low-risk for bleeding, its implementation is almost nonexistent due to the absence of accurate hemostatic risk prediction. The central hypothesis is that we can leverage population science and data science methodologies while guided by clinical and ethical principles to ascertain and predict thrombotic and hemorrhagic outcomes in ~330,000 patients with cancer from three different healthcare systems in the US. These sites are selected for their significant differences in age, sex, race, ethnicity, and cancer type to ensure fairness in model development and application. Our two specific aims are: 1) To ascertain and validate incident VTE and CRB outcomes from clinical notes in patients with cancer across three healthcare institutions; and 2) To develop, externally validate, and deploy a multimodal, dynamic risk prediction model for incident VTE and CRB following the initiation of anticancer systemic therapy to individualize preventive recommendations. We will pursue these aims using several innovative methods including natural language processing (NLP) algorithms fine-tuned from pre-trained masked language models, retrieval augmented generation (RAG) in large language model (LLM), federated learning (FL) framework to allow for decentralized privacy-preserving data sharing, and multimodal dynamic risk prediction with longitudinal trajectories from both structured and unstructured data. The proposed research is significant because accurate thrombo-hemorrhagic risk prediction will lead to increased adoption of risk-adaptive thromboprophylaxis to reduce complications in patients undergoing anticancer therapy. The expected outcome of this proposal is the creation of validated VTE and CRB dynamic risk prediction models in patients with cancer starting systemic therapy that are applicable for all racial and ethnic groups. The success of our proposal will empower other clinical investigators to validate our work against their own datasets, integrate it into clinical care, or join our ongoing FL network to further improve accuracy and applicability of the multimodal individualized treatment effects model for thrombosis and bleeding. Project Number: 1R01HL180402-01 | Fiscal Year: 2025 | NIH Institute/Center: National Heart Lung and Blood Institute (NHLBI) | Principal Investigator: Ang Li | Institution: BAYLOR COLLEGE OF MEDICINE, HOUSTON, TX | Award Amount: $721,826 | Activity Code: R01 | Study Section: Special Emphasis Panel[ZRG1 HSS-Y (90)] View on NIH RePORTER: https://reporter.nih.gov/project-details/1R01HL18040201