Machine Learning Model Predicts Catheter Thrombosis Risk in Older Cancer Patients

Peripherally inserted central catheters, better known as PICCs, have become one of the workhorses of modern cancer care. Threaded through a vein in the arm until the tip rests near the heart, they allow chemotherapy drugs, antibiotics, nutrition, and blood products to be delivered safely over weeks or months without repeatedly puncturing fragile peripheral veins. Yet the device carries a well-known danger: the foreign body sitting in the bloodstream can trigger clot formation, and when that clot develops around the catheter itself, the result is PICC-related thrombosis, a complication that can force treatment delays, catheter removal, and in the worst cases pulmonary embolism. A new study published in BMC Geriatrics by Huiping Zhang, Xiaomin Zhao, and colleagues at Shandong Medical and Pharmaceutical University and Shandong Cancer Hospital and Institute set out to answer a deceptively simple question with machine learning: which older adults with cancer are most likely to develop this clot before it happens?

The research team focused on a population that is particularly vulnerable. Older adults, defined in the study as patients aged 65 years or above, experience age-related physiological changes that tilt the hemostatic balance toward clotting, including increased platelet reactivity, elevated coagulation factors, and reduced fibrinolytic capacity. Layered on top of these changes come cancer itself, which is a prothrombotic state, and treatment-related factors such as chemotherapy that further amplify the risk. The retrospective study drew on 532 older adult cancer patients who underwent PICC placement, and among them 132 patients, or 24.8 percent, developed PICC-related thrombosis. That figure, roughly one in four, underscores why clinicians urgently need tools to stratify risk before the catheter is placed or soon after.

To build the prediction model, the researchers first had to decide which of the many variables recorded in the hospital information system actually mattered. They turned to least absolute shrinkage and selection operator regression, or LASSO, a statistical technique that compresses coefficients toward zero and effectively eliminates variables that contribute little to prediction. This approach is especially valuable in medical datasets where the number of candidate predictors can be large relative to the number of outcome events, a ratio known in the literature as events per variable. After the LASSO selection process, seven variables survived: chemotherapy, fibrinogen, D-dimer, thrombin time, albumin, white blood cell count, and platelet count. The list is a revealing snapshot of the biology at play, combining treatment exposure with coagulation markers, inflammation, and nutritional status.

With seven predictors in hand, the team trained eight different machine learning algorithms and compared their performance. The lineup included logistic regression, XGBoost, support vector machines, Random Forest, a neural network, LightGBM, K-nearest neighbours, and Adaboost. Each algorithm learns in a fundamentally different way. Logistic regression fits a linear decision boundary and produces coefficients that are directly interpretable. Tree-based ensembles such as Random Forest, XGBoost, and LightGBM build collections of decision trees that capture nonlinear interactions between variables. Support vector machines construct hyperplanes in high-dimensional feature space, K-nearest neighbours classifies patients by similarity to their closest counterparts in the training data, and neural networks learn layered representations of the input features. The researchers evaluated each model using the area under the receiver operating characteristic curve, or AUC, along with accuracy, sensitivity, specificity, and the F1 score.

What happened next is a cautionary tale that will resonate with anyone following the machine learning boom in medicine. In the training set, several of the complex models achieved high discrimination, suggesting they had learned the training data well. But when the models were tested on unseen data, some of them showed substantial performance deterioration, a classic signature of overfitting. Overfitting occurs when a flexible algorithm memorizes idiosyncrasies of the training sample rather than learning generalizable patterns, and it is a particular danger in clinical datasets of modest size. The 532-patient cohort, with 132 events, is respectable but not enormous, and complex models with many parameters can easily outgrow the information the data actually contain.

The winner, in the end, was the oldest and simplest algorithm on the list. Logistic regression demonstrated the best overall performance in the testing set, achieving an AUC of 0.724 with a 95 percent confidence interval of 0.620 to 0.828. An AUC of 0.724 means that if the model were given one patient who developed thrombosis and one who did not, it would assign the higher risk score to the patient who developed the clot about 72 percent of the time, which is moderate discrimination, useful for screening but not definitive. The model also achieved the lowest Brier score among the candidates, at 0.149, a metric that measures the mean squared difference between predicted probabilities and actual outcomes, where lower values indicate better overall accuracy and calibration. Calibration curves confirmed that the predicted risks tracked the observed event rates, and decision curve analysis showed the model delivered net clinical benefit across a clinically relevant range of risk thresholds.

To open the black box, the researchers applied SHAP, or SHapley Additive exPlanations, an explainable machine learning technique borrowed from cooperative game theory. SHAP assigns each predictor a contribution value for every individual prediction by computing how much the model’s output changes when that variable is included versus excluded, averaged across all possible orderings of the features. The analysis revealed that fibrinogen, thrombin time, chemotherapy, and white blood cell count were the major contributors to the model’s thrombosis risk predictions. Fibrinogen is a central coagulation protein that forms the scaffold of a clot, and elevated levels reflect both cancer-related inflammation and a hypercoagulable state. Thrombin time measures how quickly fibrinogen is converted to fibrin, the final step of the coagulation cascade. Chemotherapy exposure damages vascular endothelium and can activate the clotting system, while an elevated white blood cell count signals inflammation that further promotes thrombosis.

The clinical implications of the study are straightforward even if the statistics are not. A model that can be computed from seven routinely available laboratory values and one treatment variable could be integrated into hospital information systems to flag high-risk patients at the time of PICC placement or during routine follow-up. Clinicians could then tailor catheter management to the individual, choosing alternative vascular access for the highest-risk patients, scheduling more frequent monitoring for intermediate-risk patients, or considering prophylactic anticoagulation where guidelines permit. The authors suggest the model may assist in early risk stratification and individualised catheter management for older patients undergoing PICC placement, and the SHAP-based explanations give clinicians a window into why the model reached its conclusion for a particular patient, which is essential for building trust in algorithmic recommendations.

Limitations remain, as they always do in retrospective single-centre studies. The model was developed and tested within one institution, and its generalizability to other hospitals, other populations, and other catheter protocols needs external validation before widespread adoption. The AUC of 0.724, while clinically useful, leaves room for improvement, and the modest sample size constrained how much complexity the data could support, as the overfitting of the more elaborate models demonstrated. Still, the study offers a valuable lesson for the field: in clinical prediction, the most sophisticated algorithm is not always the best one, and a well-calibrated, interpretable model built on biologically meaningful variables can outperform its flashier competitors. As machine learning continues to spread through medicine, studies like this one remind us that the goal is not to impress with complexity but to deliver reliable, explainable, and actionable predictions that help clinicians protect their patients.

Subject of Research: Health

Article Title: Development of a machine learning–based prediction model for peripherally inserted central catheter (PICC)-related thrombosis in older adult cancer patients

Article References: Zhang, H., Hu, X., Chen, X., Xu, J., & Zhao, X. (2026). Development of a machine learning–based prediction model for peripherally inserted central catheter (PICC)-related thrombosis in older adult cancer patients. BMC Geriatrics. https://doi.org/10.1186/s12877-026-08458-8

Image Credits: AI Generated

DOI: 10.1186/s12877-026-08458-8

Keywords: machine learning, PICC, thrombosis, cancer, geriatrics, predictive model, logistic regression, SHAP, Development, machine, learning, based