Systematic review of a machine learning model for prediction of venous thromboembolism risk

This systematic review aims to evaluate the methodological quality, performance, and clinical applicability of machine learning (ML) models for predicting the risk of venous thromboembolism (VTE) in hospitalized patients. Specifically, we aim to assess the methodological quality and reporting transparency of the included studies, with a focus on their risk of bias and adherence to reporting guidelines.

A systematic review by use of the databases Cochrane Library, Web of Science, Embase, PubMed, CNKI, VIP Journal Database, Wanfang Database, and the Chinese Biomedical Literature Database. The study was registered at PROSPERO before data collection and PRISMA guidelines were followed. The search was conducted to identify all relevant studies published from the inception of database up to August 1, 2024. Two independent researchers screened the literature and extracted data. Model quality was assessed using the PROBAST appraisal tool and a modified TRIPOD+AI framework, alongside reported model performance metrics.

A total of seventeen studies were included, comprising 65 VTE ML models with sample sizes ranging from 120 to 9213. All models demonstrated an area under the curve (AUC) >0.7. Twenty-three ML algorithms were employed, with logistic regression (LR) being the most frequently used (n = 11), followed by XGBoost (n = 10) and random forests (RF) (n = 9). Thirteen studies utilized various stochastic algorithms. Most studies used Bootstrap or 10-fold cross-validation for internal validation, but lacked external validation, leading to a high overall risk of bias. Key predictors in these models included D-dimer, history of thrombosis, history of hypertension, age, and complications.

Existing evidence suggests that ML models can effectively predict VTE outcomes. However, most models suffer from poor methodological quality, lack of external validation, and limited generalizability. Future research should focus on large-scale, multi-center prospective studies that are based on clinical practice, improve external validation, and develop optimized local VTE risk assessment and decision support tools for better integration into clinical practice.