Electronic Health Record-Based Machine Learning Models for Predicting Revascularization, Amputation, and Cardiovascular Mortality in At-Risk Patients

Every year, thousands of people face a terrifying choice: the loss of a limb or the loss of their life. This struggle often begins with a silent, creeping condition where blood vessels in the legs narrow and harden, starving the muscles and skin of oxygen. When this blockage becomes severe, it can lead to open sores that refuse to heal or a state where the limb is in constant danger of dying. Doctors call this peripheral artery disease, and in its most critical form, chronic limb-threatening ischemia. For patients with these conditions, or with diabetic foot ulcers, the path forward is fraught with uncertainty. A doctor must decide whether to perform a complex surgery to restore blood flow, amputate the damaged limb to save the patient's life, or focus on comfort care. These decisions are incredibly difficult because every patient is different, carrying a unique mix of other health problems, from diabetes to heart disease, and social factors that influence how they respond to treatment.

For decades, doctors have relied on scoring systems and risk calculators to guide these choices, but these tools have often failed to gain widespread trust. They are too rigid for a disease that is so messy and personal, or they rely on data that is hard to gather consistently. Now, a team of researchers at the University of North Carolina has turned to a different kind of guide: the massive, digital records of patient care itself. By using a new way to organize millions of electronic health records, they built a set of computer models designed to predict what will happen to a patient next. Their goal was not to replace the doctor, but to give them a clearer view of the future, helping to decide who needs surgery, who might lose a limb, and who is at risk of dying from heart-related causes.

The researchers started with a vast pool of nearly 179,000 adult patients who had received care within the UNC Health System between 2014 and 2024. These were people already known to be at risk of limb loss, including those with peripheral artery disease, chronic limb-threatening ischemia, or diabetic foot ulcers. To make sense of this enormous amount of information, the team used a standardized system called the OMOP Common Data Model, which acts like a universal translator for medical data, allowing different hospitals to speak the same language. They added a special extension to this system, filling it with 127 specific variables related to blood vessel disease that are usually missing from standard records. This allowed them to look at the patients not just as a collection of symptoms, but as whole people with thousands of data points, ranging from their blood pressure and lab results to their history of smoking, their mobility, and even their cognitive state.

Using this rich, harmonized dataset, the team trained computer algorithms to look for patterns that human doctors might miss. They built three separate prediction models. The first tried to guess whether a patient would receive a revascularization procedure, a surgery intended to open up blocked arteries. The second tried to predict if a patient would eventually need an amputation. The third attempted to forecast whether a patient would die from a cardiovascular event, such as a heart attack or stroke. The computer did not just look at the obvious signs of leg disease; it scanned through thousands of potential factors, from the presence of specific infections to the type of medications a patient took, to see which ones actually mattered most for the outcome.

The results revealed a story of surprising clarity for some outcomes and stubborn complexity for others. The model designed to predict who would get revascularization surgery was exceptionally accurate. It correctly identified the patients who would undergo the procedure with a level of precision that approached perfection. The factors driving this prediction were exactly what a vascular surgeon would expect: the model found that the strongest signals were the specific diseases of the arteries themselves, such as blockages in the lower limbs and blood clots. The computer learned that when these specific arterial problems are present, the medical system consistently responds with surgery. It was a confirmation that the decision to operate is driven by clear, objective signs of arterial failure.

The prediction for amputation was also strong, though the story behind it was more nuanced. The model showed that whether a patient loses a limb depends heavily on how their body and mind handle the stress of treatment. The most powerful predictors were not just the severity of the leg wound, but how the treatment interacted with the patient's overall frailty. For instance, if a patient with dementia or severe physical weakness underwent revascularization, the model predicted a higher likelihood of eventual amputation. Conversely, for patients who were too frail to undergo surgery, the model suggested they were less likely to have an amputation, likely because their care team steered them toward palliative options instead. This finding highlights a difficult reality: the path to losing a limb is often determined not just by the disease in the leg, but by the patient's ability to survive the journey of treatment.

However, the model for predicting cardiovascular death told a different story. While the computer could identify the usual suspects—conditions like heart failure, diabetes complications, and high blood sugar—it could not predict death with the same confidence it showed for the other outcomes. The accuracy of this model was modest, suggesting that while the risk factors are known, the actual moment of death is influenced by too many unpredictable variables to be captured by a single formula. The researchers noted that revascularization surgery, which fixes the local problem in the leg, did not appear to change the risk of dying from a heart event. This implies that saving a limb and saving a life from heart disease are two separate battles, and fixing one does not guarantee victory in the other.

The study also addressed a common hope in medical research: that artificial intelligence might discover entirely new, hidden risk factors that doctors have never considered. In this case, the computer did not find any secret codes. The most important variables it identified were the ones clinicians already knew were critical. The machine learning approach did not invent new knowledge; instead, it confirmed that the existing understanding of vascular disease is robust. It proved that if you gather enough data in a clean, organized way, the patterns of disease become clear. The fact that the model could predict amputation so well suggests that with the right data, we can foresee these outcomes, but it also shows that predicting death remains a much harder challenge because the human body can fail in so many different ways.

This work represents a significant step forward in how we use data to care for patients with limb-threatening conditions. By organizing electronic health records into a standard format and adding specific details about blood vessel disease, the researchers created a tool that can learn from the experiences of nearly 180,000 people. The models they built are not perfect, and they were tested on patients from a single health system, which means they need to be checked against data from other hospitals to ensure they work everywhere. Yet, the results offer a promising glimpse into a future where doctors can use the collective wisdom of millions of patient records to make better, more personalized decisions. For the patient facing the loss of a limb, this kind of clarity could mean the difference between a treatment plan that works and one that does not, turning a terrifying uncertainty into a manageable path forward.