# Machine learning-based prediction model and web calculator for postoperative LDVT in colorectal cancer

## TL;DR

This study created a machine learning model and web tool to predict the risk of postoperative deep vein thrombosis in colorectal cancer patients, aiming to improve early detection and prevention.

## Contribution

A robust machine learning model and accessible web calculator for predicting postoperative LDVT in CRC patients.

## Key findings

- The random forest model achieved high accuracy with AUCs of 0.942, 0.897, and 0.891 in training, testing, and validation sets.
- D-dimer, preoperative intestinal obstruction, and Caprini score were identified as major predictors of LDVT.
- A web-based calculator was developed to provide individualized risk estimation for clinical use.

## Abstract

Lower limb deep vein thrombosis (LDVT) is a common but often underdiagnosed complication after colorectal cancer (CRC) surgery. Its early symptoms are subtle, and delayed detection can lead to post-thrombotic syndrome or even life-threatening events. However, effective tools for early risk assessment are lacking.

To identify risk factors for postoperative LDVT in CRC patients and develop a machine learning (ML)-based risk prediction model with an accessible web calculator.

This retrospective study included 1,200 CRC patients undergoing radical surgery. A modeling cohort of 1,000 patients (January 2021–December 2022) was randomly split 8:2 into training and testing sets, and 200 patients (March–August 2024) formed an external validation cohort. Risk factors were screened using univariate analysis and least absolute shrinkage and selection operator (LASSO) regression. Eight ML models were constructed and compared based on area under the curve (AUC), accuracy, sensitivity, and F1-score. The best-performing model was interpreted using SHapley Additive exPlanations (SHAP), and a web-based calculator was developed.

Among 1,200 patients, 369 (30.75%) developed LDVT (31.5% in the modeling cohort, 27% in the validation cohort). Seventeen variables were associated with LDVT in univariate and LASSO analyses, and the top 10 were used to build models. The random forest (RF) model showed the best performance, with AUCs of 0.942, 0.897, and 0.891 in the training, testing, and validation sets, respectively, demonstrating high accuracy and generalizability. SHAP analysis identified D-dimer, preoperative intestinal obstruction, Caprini score, age, intraoperative blood loss, and diabetes as major predictors, with D-dimer having the strongest impact. A web-based calculator (https://crc-ldvt.shinyapps.io/RF-model/) was constructed to provide individualized risk estimation.

This study developed and validated a robust ML-based model for predicting postoperative LDVT in CRC patients. The RF model, incorporating key clinical predictors, demonstrated high predictive performance and clinical relevance. The online calculator enables rapid, individualized risk assessment and may help guide early prevention strategies, reducing postoperative complications and improving patient outcomes.