Enhancing prediction and stratifying risk: machine learning and bayesian-learning models for catheter-related thrombosis in chemotherapy patients.

Abstract

Background: Catheter-related thrombosis (CRT) is a serious complication in cancer patients undergoing chemotherapy, yet existing risk prediction models demonstrate limited accuracy. This study aimed to evaluate the clinical utility of machine learning (ML) and Bayesian-learning models for CRT prediction in a large cohort of breast cancer patients undergoing catheterization.

Methods: A total of 3337 breast cancer patients with central venous catheters (Cohort 1) were included to develop and test ML models. Given the suboptimal clinical feasibility of ML models, the Bayesian-learning model was constructed using odds ratio analysis and Gaussian distribution. The hazard ratio for the high-risk and low-risk groups was calculated using Cox proportional hazards regression analysis, and the model was validated in an independent cohort of 1274 patients (Cohort 2).

Results: In Cohort 1, 246 patients (7.37%) developed CRT. Among the eight ML algorithms tested, WeightedEnsemble model exhibited relatively stable performance, achieving area under the receiver operating characteristic curves of 0.89 in the training set and 0.69 in the test set. WeightedEnsemble improved generalization by integrating multiple base models. The odds ratio analysis and Bayesian-learning modeling identified 4 independent risk factors: hemoglobin (threshold point [TP]: 134.63 g/L), activated partial thromboplastin time (TP: 31.71 s), total cholesterol (TP: 11.19 mmol/L), and catheterization approach (TP: peripherally inserted central catheters). A simplified risk stratification system was developed, categorizing patients into low-risk (0-1 factors) and high-risk (2-4 factors) groups. This system exhibited strong CRT risk discriminative ability, as confirmed through survival analysis (P < 0.001 in both cohorts). In Cohort 1, cox regression analysis showed that the high-risk group had hazard ratio (HR) of 1.60 (95% confidence interval [CI], 1.15-2.22) for both catheter indwelling time and catheter use duration. In Cohort 2, the system maintained stable discriminative ability, with an HR of 5.63 (95% CI, 3.46-9.21) for catheter indwelling time and 5.62 (95% CI, 3.46-9.12) for catheter use duration.

Conclusions: While ML models demonstrated high predictive performance, their clinical applicability was limited due to complexity. The Bayesian-learning-based risk stratification model provided a simplified yet robust alternative, effectively predicting CRT risk and offering a clinically feasible tool for risk assessment in breast cancer patients with chemotherapy. Further validation in diverse cancer populations is warranted to refine its generalizability.