A deep learning model to dynamically predict cancer-associated thromboembolism in large-scale healthcare systems.

Summary

A new deep learning model uses longitudinal electronic health records to predict the risk of blood clots in cancer patients throughout their first year of systemic treatment. This dynamic approach outperforms traditional static risk scores, allowing clinicians to identify high-risk patients more accurately as their medical history evolves.

Key result

A transformer model predicting venous thromboembolism risk in cancer patients achieved area under the curve (AUC) values ranging from 0.68 to 0.77 across quarterly intervals in a cohort of 80,808 patients.

Abstract

Venous thromboembolism (VTE) is a leading cause of preventable death among patients undergoing systemic treatment for cancer. Studies suggest that treatment strategies such as direct oral anticoagulant administration can significantly reduce the likelihood of VTE. Therefore, identifying people at high risk is of critical importance. Leveraging electronic health records (EHRs) from the U.S. Veterans Affairs (VA) healthcare system, we developed a transformer model to predict VTE risk in 80,808 cancer patients following the initiation of systemic treatment. The model uses longitudinal diagnostic codes, laboratory values, and demographic data. The proposed transformer model dynamically predicts VTE risk in 3-month quarterly intervals over the year following systemic treatment, achieving progressively improved performance across quarters (AUC: 0.68-0.77). The model is similarly performant on the external validation cohort from the Harris Health System (HHS) with 9752 patients (AUC: 0.68-0.74). By improving its predictions as a patient's history evolves, this dynamic model surpasses prior static risk scores and better supports actionable decisions deeper into the treatment course.