Beyond Padua and IMPROVE: Machine Learning Outperforms Guideline Risk Scores for Prediction of Radiologically Confirmed Hospital-Acquired Venous Thromboembolism
This study demonstrates that machine learning models trained on routine admission data significantly outperform traditional guideline-based risk scores (Padua and IMPROVE) in predicting radiologically confirmed hospital-acquired venous thromboembolism across a large, unselected cohort of nearly 400,000 admissions.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Every year, millions of people enter a hospital for treatment, and for many, the act of lying still or recovering from an illness carries a hidden danger: the formation of blood clots. These clots can travel through the veins and block blood flow to the lungs, a condition known as venous thromboembolism, which is a leading cause of preventable death in hospitals. Doctors have long relied on simple checklists to decide who needs preventive medication. These checklists ask a few basic questions: Is the patient over a certain age? Do they have cancer? Are they unable to move around? Based on the answers, a score is calculated to estimate risk. While these tools are standard practice, they often miss the mark, failing to identify many patients who will develop clots while flagging others who are actually safe. The challenge has been to find a way to see the danger more clearly without adding more work for overburdened medical staff.
A team of researchers set out to test whether a computer program could learn from hospital records to predict these clots better than the traditional checklists. They turned to a massive collection of de-identified medical records from a single hospital in the United States, covering over 399,000 adult admissions between 2008 and 2022. To ensure they were studying new cases rather than old history, they focused only on patients who had never had a blood clot before. The researchers did not rely on simple billing codes to find the sick patients, as these codes can be inaccurate. Instead, they built a system to read the full text of radiology reports, such as CT scans and ultrasound images, looking for the specific words doctors use when they confirm a new clot. This careful process identified 1,915 patients who developed a new clot during their stay, a rate of roughly five cases for every 1,000 admissions.
The researchers then trained two types of computer models to predict these events. The first type looked at a snapshot of the patient's data from the first 24 hours of their stay, including their age, medical history, lab test results, and whether they had surgery. The second type was dynamic, meaning it updated its prediction every day as new lab results and vital signs came in. They compared these computer models against the two most common guideline checklists used in hospitals today. The results were striking. The traditional checklists performed poorly, correctly identifying the risk in only about 62% of cases when tested on new data. In contrast, the computer models, which used the same routine data available to doctors, achieved an AUC of 0.89. The improvement was so large that it was statistically undeniable, showing that the computer models could spot patterns in the data that the simple checklists completely missed.
The study also explored whether more complex, deep-learning systems that try to read the entire sequence of a patient's daily changes would do even better. Surprisingly, these complex systems offered no advantage over the simpler models that just looked at the daily summaries. The researchers found that the simpler models were just as effective at capturing the risk signals hidden in the data. Furthermore, the computer models remained superior even when the researchers tested them under strict conditions, such as looking only at patients who developed clots more than a day after arriving, or including patients who had a history of clots. The advantage of the computer models held steady across all these scenarios.
This work suggests that the future of hospital safety may lie in software that runs quietly in the background, automatically calculating risk the moment a patient is admitted. Unlike the current checklists, which require a nurse or doctor to manually fill out a form, these computer models can pull the necessary information directly from the electronic health record. While the study was conducted on historical data and requires further testing in real-time hospital settings before it can be widely adopted, the findings indicate that a significant leap in accuracy is possible. By using the data already sitting in hospital computers, it may soon be possible to protect more patients from blood clots without adding a single extra step to the daily routine of medical care.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.