1857P - Machine learning models (MLM) predict venous thromboembolism events (VTE) in Asian oncological inpatients better than the Khorana Score (KS)

Background

Cancer is the primary risk factor for VTE. Validated risk scores like the Khorana score (KS) can identify cancer pts that are more likely to develop VTE and has been validated in Asia. However, its performance is still suboptimal with a clinically unacceptably high false negative rate. We aim to improve on the prediction of VTE patient events by leveraging on MLMs.

Methods

We collected extensive clinical data of pts with solid organ tumors or lymphoma hospitalised from December 2014 to March 2019 (N = 5144). We used ensemble learning type MLMs to predict VTEs. To develop the MLMs, the pt cohort was split into a 80% training set and a 20% test set. The performance of the MLMs were compared to the performance of the Khorana score (cancer type, hemoglobin, platelet and total white count, body mass index (BMI) more than 35; high risk 3 or more) on the same test cohort.

Results

369/5144 hospitalised pts were diagnosed with VTE (incidence: 7.17%). The performance of KS was detailed in the table. Due to low incidence of pts with high BMI, we evaluated a modified Khorana Score removing BMI (high risk 2 or more), with slightly improved PPV and NPV. Developed MLM have superior accuracy, AUC ROC, PPV and NPV compared to the Khorana Score, as per table (AUC comparison only). The Gradient Boost Model performed the best. Table: 1857P

AUC comparison only

AUC: Area under Curve (Table): AUC ROC of the KS, Modified KS and developed MLM models to predict VTE events in oncological Inpatients

Conclusions

This is the largest study using MLM predicting VTE in oncological pts. All the MLMs evaluated outperformed the KS to predict VTEs in cancer pts. However, further optimisation of these MLMs such as the inclusion of additional features or unstructured data are required to improve performance before the MLM predictions can be clinically meaningful. In populations with a predominantly low BMI, the omission of BMI as part of the KS results in slightly improved predictive performance.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.