Personalised thrombo-embolic risk prediction after endometrial cancer surgery: an explainable AI approach using SHAP

Postoperative lower extremity deep vein thrombosis (LEDVT) remains a significant clinical challenge for patients undergoing treatment for endometrial cancer. Researchers have now developed an explainable machine learning framework designed to offer personalized risk predictions, aiming to improve patient outcomes through more precise postoperative management.

Did You Know? The researchers evaluated 26 different machine learning algorithms before identifying the Support Vector Machine (SVM) model as the most robust architecture for this specific clinical application.

Refining Risk Prediction

The study utilized perioperative data from 841 patients across multiple centers to train and validate the new model. By applying recursive feature elimination, the research team narrowed down a vast array of data to a concise set of four critical variables: postoperative D-dimer, age, fibrinogen, and clinical stage.

This streamlined approach proved highly effective in clinical testing. The model achieved an area under the curve (AUC) of 0.828 during internal validation and maintained a strong performance of 0.819 when tested against an independent external cohort.

Expert Insight: The integration of SHAP (SHapley Additive exPlanations) is a turning point for clinical AI. By moving away from “black-box” models and toward transparent, interpretable logic, clinicians can finally see the “why” behind a risk assessment, which is essential for building the trust required to adopt these tools at the bedside.

Bridging the Gap Between AI and Clinical Trust

A primary hurdle in medical AI is the “black-box” nature of complex algorithms, which often obscures how a final risk score is calculated. To address this, the team integrated SHAP to quantify how individual factors contribute to a patient’s risk profile.

This transparency revealed non-linear associations, most notably the critical risk threshold for D-dimer levels. To make these insights actionable, a web-based decision support interface was implemented, allowing for real-time, interpretable risk assessments during the recovery process.

Future Implications for Patient Care

Looking ahead, this framework could fundamentally change how surgical teams monitor recovery. If successfully integrated into broader clinical workflows, this tool may allow hospitals to identify high-risk patients earlier, potentially enabling more proactive and personalized interventions.

Future iterations of this technology could see wider adoption across oncology departments, provided the model continues to demonstrate high discriminative performance in diverse patient populations. Analysts expect that as these decision support interfaces become more common, the standardization of postoperative management for endometrial cancer patients may improve significantly.

Frequently Asked Questions

What four variables does the model use to predict risk?

The model uses a concise set of four variables: postoperative D-dimer levels, patient age, fibrinogen levels, and the clinical stage of the cancer.

Day6 Lecture – Explainable AI with SHAP – Demystifying Model Predictions

How does the model overcome the “black-box” nature of AI?

The researchers integrated SHapley Additive exPlanations (SHAP) to quantify individual feature contributions, providing a transparent view of the decision-making logic behind each risk assessment.

Is this tool currently available for clinical use?

The researchers have implemented a web-based decision support interface designed to provide real-time, interpretable risk assessments for clinicians managing endometrial cancer care.

How do you think the integration of real-time AI tools will change the way patients perceive their own postoperative recovery process?