Dynamic Risk Stratification in Thromboembolism Prevention for Cancer Patients

Dynamic risk stratification enables more precise forecasting of venous thromboembolism risk in cancer patients by continuously updating predictions as new clinical data emerge. A transformer-based deep learning model was developed using longitudinal electronic health records from the U.S. Veterans Affairs healthcare system to forecast venous thromboembolism risk among 80,808 cancer patients after systemic treatment initiation. The approach incorporates diagnostic codes, laboratory values, and demographic information to generate quarterly predictions over the subsequent year, with area under the curve values rising from 0.68 to 0.77 across successive intervals. External validation in the Harris Health System cohort of 9,752 patients yielded comparable performance (AUC 0.68–0.74), demonstrating that the model outperforms conventional static risk assessment tools by refining estimates as patient data evolve.

Evolving Prediction Accuracy

The model’s progressive improvement in discrimination across quarterly windows illustrates its capacity to integrate accumulating clinical history, thereby supporting more precise identification of individuals who may benefit from preventive strategies such as direct oral anticoagulant administration. Specific performance gains are evident in the reported AUC trajectory, which reflects enhanced calibration as additional longitudinal observations become available. Comparable results in the independent Harris Health System cohort reinforce the robustness of these findings, indicating that dynamic risk stratification maintains reliability when applied to distinct large-scale healthcare environments with differing patient demographics and data capture practices.

Sequential Data Processing

The underlying methodology relies on sequential processing of structured data elements drawn from routine clinical encounters, enabling the model to update risk estimates at three-month intervals without requiring manual feature engineering. This design choice aligns with the temporal nature of cancer treatment trajectories, where laboratory trends and diagnostic events accumulate over time. By leveraging the transformer’s attention mechanisms on these evolving inputs, the framework captures dependencies that static scoring systems overlook, thereby grounding its predictions in the actual progression of patient status within real-world healthcare datasets.

Targeted Therapy Allocation

These dynamic predictions facilitate risk-stratified allocation of preventive therapies, which may improve the efficiency of resource utilization in oncology settings. In contexts of market access and reimbursement, payers could consider incorporating such models into coverage criteria to prioritize direct oral anticoagulant use among patients whose updated risk profiles exceed defined thresholds, potentially reducing preventable events and associated downstream costs. The observed performance stability across two large systems further suggests scalability for broader adoption in value-based frameworks that reward proactive management of treatment-related complications.