TL;DR
This review synthesizes the latest research on Venous Thromboembolism (VTE) risk factors and assessment tools for colorectal cancer (CRC) patients undergoing surgery. It highlights a critical gap between high patient risk and low prevention rates, advocating for a shift from traditional scoring systems to specialized, high-accuracy machine learning and nomogram-based prediction models.
What Problem Does This Paper Address?
Venous Thromboembolism—comprising deep vein thrombosis (DVT) and pulmonary embolism (PE)—is the second leading cause of cancer-related death. Colorectal cancer patients are particularly vulnerable, with VTE occurring in up to 8.9% of cases. While surgery is the primary treatment for CRC, the "perioperative" period (the time around the surgery) creates a perfect storm of risk factors: malignancy-induced hypercoagulability, surgical trauma, and postoperative immobility.
The core challenge is that current risk assessment is inconsistent. While studies suggest over 80% of postoperative CRC patients are at high risk for VTE, only about 24% actually receive adequate preventive care. There is currently no "gold standard" for identifying which patients need aggressive intervention versus those who don't, leading to both missed diagnoses and unnecessary over-treatment.
Methods
As a systematic review, the authors analyze and categorize existing literature into three distinct perioperative phases:
- Preoperative: Patient-specific factors like age (≥60), BMI, and baseline D-dimer levels.
- Intraoperative: Procedure-specific factors including surgical duration (≥3 hours), blood loss, and the choice between open vs. laparoscopic approaches.
- Postoperative: Recovery factors such as immobilization duration (≥3 days) and complications like infections or anastomotic leaks.
The paper then evaluates the evolution of assessment tools, moving from general clinical scoring sheets (Caprini, Khorana) to advanced diagnostic tests (Thromboelastography) and specialized statistical/machine learning models.
Datasets
The paper reviews findings from several large-scale and specialized cohorts, including:
- The Schlick et al. study: A massive dataset of 51,139 patients undergoing CRC resection.
- The CRC-VTE model cohort: 1,836 patients across 46 centers in China.
- Machine Learning (ML) training sets: A cohort of 1,191 CRC patients used to develop XGBoost-based models.
- Localized models: Smaller, single-center retrospective datasets (ranging from 213 to 982 patients) used to validate tools like the ABAD and CRSPOT nomograms.
Key Insights and Ideas
- The Dynamic Nature of Risk: Risk isn't a static number assigned at admission. It fluctuates. For example, a patient might be low-risk pre-op but move to high-risk due to a 4-hour surgical duration or a postoperative infection.
- The Laparoscopic Debate: Interestingly, the paper notes conflicting data on surgical approach. While laparoscopic surgery is generally less invasive, some studies suggest it may increase VTE risk due to longer operative times and specific patient positioning that can obstruct venous return.
- ML Outperforms Traditional Scales: The "gold standard" Caprini score is criticized for being overly complex (40 items) and often over-classifying patients as "high risk," which dilutes its clinical utility. In contrast, machine learning models (specifically XGBoost) and nomograms provide higher predictive accuracy (AUC) with fewer variables.
- Biomarker Integration: D-dimer levels remain a cornerstone of assessment, but the paper emphasizes that they are most effective when integrated into multi-variable models rather than used as a standalone "yes/no" trigger.
Key Results
The paper summarizes the performance of several modern prediction models (detailed in Table 1):
- Machine Learning (ML) Models: Demonstrated the highest performance, with a validation set AUC of 0.908.
- CRSPOT Nomogram: Achieved an AUC of 0.826 using seven postoperative factors (e.g., hemoglobin < 10 g/L, BMI ≥ 25).
- Sir Run Run Shaw (SRSV) Model: A localized Chinese tool that achieved an AUC of 0.769 using only four routine variables, making it highly practical for fast-paced clinical settings.
- The Prevention Gap: The review highlights that while 82.5% of patients are technically "high risk," the lack of precise tools contributes to the fact that only 24.3% receive VTE prophylaxis.
Limitations and Future Work
The authors identify several hurdles to universal adoption:
- Retrospective Bias: Most current models are built on retrospective data, which may not account for real-time clinical nuances.
- Lack of External Validation: Many of the highest-performing models are single-center studies. They require large-scale, multi-center prospective validation before they can be considered "standard of care."
- Non-Unified Standards: There is a lack of consensus on which variables are most critical, leading to a fragmented landscape of different models for different regions or hospitals.
- Future Direction: The authors suggest that future research should focus on "personalized, precision thrombosis management," likely through the integration of real-time electronic health record (EHR) data into automated ML risk-scoring systems.
Summary updated