Machine learning-based identification of key risk factors for deep vein thrombosis within one week after total joint arthroplasty in elderly patients

Every year, millions of older adults undergo major surgery to replace worn-out hip or knee joints, a procedure that restores mobility and relieves pain. However, the recovery process carries a hidden danger: the formation of blood clots in the deep veins of the legs, a condition known as deep vein thrombosis. For elderly patients, this complication is particularly serious. It can extend hospital stays, increase medical costs, and, in severe cases, lead to life-threatening events if a clot travels to the lungs. Doctors have long relied on standard checklists to estimate who is most at risk, but these tools often treat all joint replacement patients the same, labeling them all as high-risk without distinguishing who is truly in danger. This lack of precision makes it difficult to provide the right level of protection for each individual.

To solve this, researchers turned to a branch of computer science called machine learning. Instead of following a fixed set of rules, these computer programs learn from vast amounts of past patient data to find complex patterns that human observers might miss. The goal is to build a digital tool that can look at a patient's medical history, blood test results, and details of their surgery to predict with high accuracy whether they will develop a blood clot within the first week after the operation. By identifying the specific factors that matter most, such a tool could help doctors tailor prevention strategies, ensuring that those who need extra care receive it while avoiding unnecessary treatment for those who do not.

A team of researchers from the University of Science and Technology of China set out to build and test such a tool. They gathered information on 3,050 elderly patients, all aged 60 or older, who had undergone total joint arthroplasty at a single hospital in 2024. The team collected a wide range of data points for each person, including their age, gender, body mass index, and medical history of conditions like diabetes or cancer. They also pulled specific numbers from preoperative blood tests, such as levels of inflammation and blood clotting factors, and recorded details from the operating room, including how long the surgery lasted, whether bone cement was used, and if a tourniquet was applied to stop blood flow during the procedure. The outcome they were trying to predict was simple but critical: did the patient develop a blood clot in their leg within seven days of the surgery, as confirmed by an ultrasound scan?

To train their computer models, the researchers had to address a common problem in medical data: most patients do not develop clots, while a smaller number do. This imbalance can confuse a learning algorithm. To fix this, they used a technique that creates synthetic examples of the less common group, effectively balancing the dataset so the computer could learn from both sides equally. They then split the data, using the majority of it to teach four different types of machine learning algorithms how to make predictions, and saving the rest to test how well those predictions worked on new, unseen patients. The four algorithms they tested included a traditional statistical method known as logistic regression, a decision tree that asks a series of yes-or-no questions, a random forest that combines many decision trees, and a powerful method called XGBoost, which builds a highly accurate model by learning from its own mistakes.

When the team compared the results, one model stood out clearly. The XGBoost algorithm proved to be the most effective, correctly identifying the vast majority of patients who would and would not develop a clot. It achieved a level of accuracy that far surpassed the other methods, including the traditional statistical approach which struggled to capture the complex relationships between the different risk factors. The random forest model performed well but was not quite as precise, while the decision tree and the traditional statistical model fell significantly behind. The superior performance of the XGBoost model suggests that the risk of developing a clot is not driven by a single factor, but by a complex web of interactions that only a sophisticated computer program can untangle.

To understand why this model worked so well, the researchers used a method to peel back the layers of the algorithm and see which specific pieces of information were driving the predictions. They found that the most influential factors were not the length of the surgery or the use of bone cement, as some might expect, but rather the patient's preoperative blood work. Specifically, the time it took for the blood to clot, the level of a protein called C-reactive protein which signals inflammation, and the amount of a substance called D-dimer were the strongest predictors. The patient's sex also played a significant role, with women showing a higher risk than men in this group. These findings align with what is known about how inflammation and blood clotting interact in the body, suggesting that a patient's internal state before the surgery is a more reliable indicator of early risk than the details of the operation itself.

The study concludes that this machine learning approach offers a promising way to improve patient care. By using a model that can accurately sort through complex medical data, doctors may soon be able to identify elderly patients at high risk for blood clots with much greater precision than current methods allow. This would allow for more targeted prevention strategies, focusing resources on those who need them most. While the researchers note that their work is based on data from a single hospital and requires further testing in larger groups to confirm its reliability, the results provide a strong foundation for a new era of personalized risk assessment in orthopedic surgery. The path forward involves validating these findings across different medical centers to ensure the tool works for everyone, but the potential to prevent a dangerous complication through better prediction is now within reach.