Machine learning survival model for personalised prevention of catheter-related thrombosis in tumour patients

Central venous catheters for drug delivery introduce catheter-related thrombosis (CRT) and influence the survival of cancer patients. The key unmet needs to personalise CRT prevention include identifying high-risk patients and optimising extubation time. In this study, we aimed to develop a survival model to facilitate personalised CRT prevention strategies.

We prospectively collected tumour patient catheterization data across 4 centres. The SM-CRT survival model, which provides both continuous risk ranking (crank) predictions and the survival distribution (distr) predictions was constructed.

Here we include a total of 30,947 patients. The SM-CRT model exhibits robust performance in identifying high-risk patients, with c-indexes of 0.714 in the prospective test dataset and 0.678 and 0.779 in 2 external test datasets based on crank predictions. Femorally inserted central catheter (FICC), peripherally inserted central catheter (PICC), tumours in the thoracic cavity, and alkylating agents are identified as high-risk factors. Patients are subsequently divided into high-risk, low-risk, and long-term period groups on the basis of their distr predictions. The predicted low-risk and long-term groups present significantly fewer CRT events per day than the high-risk group in both the training dataset (odds ratio [OR] = 0.54, 95% CI [0.38-0.91], adjusted p-value [p_adj] <0.001; OR = 0.39, 95% CI [0.34-0.44], p_adj <0.001) and the test dataset (OR = 0.47, 95% CI [0.28-0.87, p_adj = 0.024; OR = 0.41, 95% CI [0.28-0.61], p_adj <0.001).

The high c-indexes based on crank predictions demonstrated the ability of the SM-CRT model to identify high-risk patients for thromboprophylaxis. Additionally, the SM-CRT model can guide extubation time by identifying high-risk periods through distr predictions.