Machine Learning-Supported Efficient VTE Risk Assessment using Routinely Collected Electronic Health Record Data

Hospitals are places where the body's defenses are often at their lowest. When a person is admitted for surgery or a serious illness, the natural flow of blood through the veins can slow down, and the body's stress response can make the blood more likely to clot. These clots, known as venous thromboembolism, can form in the legs and travel to the lungs, causing a blockage that is a leading cause of preventable death in hospitals. To stop this from happening, medical guidelines require doctors and nurses to check every adult patient upon arrival. They must fill out a specific form that asks a series of questions about the patient's health history, their current condition, and their mobility. Based on the answers, the medical team decides whether the patient needs special stockings, blood-thinning medication, or both. This process is meant to be a safety net, catching risks before they become emergencies.

However, filling out these forms is a manual task that relies on a human being reading a chart, remembering a patient's history, and typing the answers into a computer. This creates a gap between what the hospital computer system already knows about a patient and what the doctor writes down on the form. The computer might have a record of a past surgery or a diagnosis of cancer, but if the doctor does not manually select those items on the checklist, the system treats the patient as lower risk than they actually are. This study, conducted across five major hospitals in London, asked a simple but critical question: could the hospital's own computer records automatically fill in the answers for these safety forms, saving doctors time and ensuring no risk factors are missed?

The researchers looked at a massive amount of data, examining nearly 580,000 hospital admissions over a ten-year period. They found that while doctors were very good at filling out the forms—completing them for almost 97% of patients—the timing was often off. In 2015, less than half of these forms were completed within the first 14 hours of a patient's arrival, a window of time that guidelines say is crucial for preventing clots. By 2024, that number had improved significantly to over 90%, but the delay remained a problem, especially for patients admitted for medical reasons rather than surgery. The study also revealed that even when forms were completed, the information written on them did not always match the facts already stored in the hospital's electronic records.

When the team compared what doctors wrote on the forms against what the computer records showed, they found a pattern of missed details. For common factors, like a patient being older than 60, the doctors and the computer records agreed most of the time. But for less common but serious risks, the doctors often forgot to write them down. For example, if a patient was admitted to a critical care unit, the computer knew it, but the doctor only wrote it down on the form about 14% of the time. Similarly, a history of blood clots in the family was recorded in the computer for many patients, yet doctors only noted it on the form for about 5% of those same patients. This meant that the safety forms were often missing the very clues that would trigger the most important protective treatments.

To see if this missing information mattered, the researchers checked if the risk scores on the forms actually predicted who would get a blood clot. They found that patients marked as high risk on the forms were indeed much more likely to develop clots than those marked as low risk, proving the forms were useful. However, they also found that when doctors wrote down that they were giving blood-thinning medication, it did not always match what the pharmacy records showed. Sometimes a doctor wrote that they gave a drug, but the pharmacy record showed nothing, and vice versa. This inconsistency made it hard to know exactly how well the current system was protecting patients.

The core of the study was an experiment to see if a computer could do the job of the doctor when filling out the form. The researchers built a machine learning model, a type of software that learns from past data, to look at the electronic health records and automatically select the risk factors. They tested this system on a group of patients and compared its performance to the forms actually filled out by doctors. The computer model, using only the information available within the first 14 hours of admission, performed just as well as the doctors in predicting who would develop a blood clot. In fact, the computer was able to identify the same high-risk patients that the doctors found, but it did so by using data that was already sitting in the system, data that the doctors had often overlooked when filling out the paper form.

The study suggests that the hospital's own data is rich enough to support these safety checks without requiring doctors to re-enter everything from scratch. The computer models showed that they could find the same patterns of risk as the human clinicians, and in some cases, they found risk factors that the humans had missed. The researchers concluded that the best path forward is not to replace the doctor, but to let the computer pre-fill the form with the information it already knows. The doctor would then simply review the pre-filled answers, confirm them, and add any new details. This approach would reduce the time doctors spend on paperwork, ensure that critical risk factors are not forgotten, and help hospitals meet the strict 14-hour deadline for safety assessments. By using the data that is already there, hospitals could make their safety nets stronger and more reliable, ensuring that the right protection reaches the right patients at the right time.