Generalizing Major Adverse Cardiovascular Events (MACE) Prediction Across Study Designs Using Ensemble Transfer Learning in 4.2 million US Veterans

Heart disease remains the leading cause of death in the United States, claiming the lives of roughly one in four people. For decades, doctors have relied on established risk factors to predict who might suffer a major cardiac event, such as a heart attack, a stroke, or death from cardiovascular causes. These factors include age, sex, high blood pressure, diabetes, and cholesterol levels. While these guidelines have saved lives, they often struggle to adapt when applied to different groups of people or when the data they are built upon comes from different types of medical studies. A model trained on one group of patients might fail when used on another, simply because the way the data was collected created hidden biases. The challenge for modern medicine is to build prediction tools that are not only accurate but also robust enough to work across diverse populations and changing healthcare systems.

To address this, a team of researchers from the Los Alamos National Laboratory and various Veterans Affairs medical centers turned to a massive dataset containing the electronic medical records of 4.2 million U.S. veterans. Their goal was to create a new way of predicting major adverse cardiovascular events, or MACE, that could learn from different study designs and generalize its findings to real-world patients. They discovered that by combining two very different ways of looking at patient data—one focused on finding the biological causes of disease and the other focused on screening a broad population—they could build a much stronger prediction tool than either method could achieve alone.

The researchers worked with a dataset that spanned seven and a half years of medical history for millions of patients. They needed to predict who would experience a heart attack, a stroke, or a cardiovascular-related death in the future. To do this, they looked at a wide array of information, including lab results, vital signs like blood pressure, diagnoses, medications, and how often patients visited the doctor. They organized this information into time blocks leading up to the prediction date, creating a detailed timeline of each patient's health.

The team faced a fundamental problem: different types of medical studies capture different aspects of health. One type of study, known as a retrospective case-control study, is excellent for identifying the biological drivers of a disease. In this design, researchers match patients who have had a heart event with similar patients who have not, carefully controlling for factors like age and how often they visit the doctor. This matching removes the noise of healthcare usage patterns, allowing the biological signals to stand out clearly. However, these matched groups do not always look like the general population seen in a doctor's office.

The other type of study is a prospective screening cohort, which looks at a broad group of patients as they are seen in routine care. This design reflects the real world, including the fact that some people visit the doctor frequently while others rarely do. While this provides a realistic picture of the population, the data is often "noisy" because the frequency of visits is tied to how sick a person is, making it hard to separate the disease itself from the behavior of seeking care.

The researchers realized that trying to use just one of these approaches was limiting. Instead, they developed a framework called ensemble transfer learning. They first trained several different machine learning models on the retrospective case-control data to learn the pure biological drivers of heart disease. Then, they took these models and "fine-tuned" them using the prospective screening data. This process allowed the models to adjust to the real-world patterns of healthcare utilization while keeping the core biological insights they had learned. It is similar to how a musician might practice a piece in a quiet studio to master the notes, and then perform it in a noisy concert hall, adjusting their volume and timing to the room without losing the music itself.

The results were significant. When the researchers tested their new combined models on a prospective cohort of veterans from 2017, the models achieved a high level of accuracy, correctly distinguishing between those who would and would not have a cardiac event better than previous standard methods. The models were particularly good at identifying high-risk patients, a group that earlier tools often missed. Furthermore, the models remained accurate across different demographic groups, including men and women, and people of different races and ethnicities.

A key finding of the study was that the biological drivers of heart disease identified in the carefully matched retrospective groups were consistent with those found in the broader prospective groups. This suggests that the core mechanisms of the disease are stable and can be learned from one type of study and successfully applied to another. The study also highlighted that non-linear models, which can capture complex interactions between variables, performed better than traditional linear models. These advanced models were able to understand that a combination of factors, such as high blood pressure and a history of chest pain, created a risk that was greater than the sum of the individual parts.

The researchers also examined how the models handled missing data. In electronic medical records, it is common for certain tests to be missing. The study found that the absence of a test result was often informative; for instance, if a patient had not had a blood pressure check in a long time, it could indicate a lack of engagement with the healthcare system, which itself is a risk factor. The new models were able to use these patterns of missingness to improve their predictions, whereas older models often treated missing data as a simple error to be ignored.

This work demonstrates that it is possible to build clinical prediction tools that are both biologically grounded and practically useful. By acknowledging that different study designs offer different strengths, the researchers created a system that leverages the precision of matched studies and the realism of broad screening. The resulting models are not just more accurate; they are also more adaptable. As healthcare systems change, new treatments are introduced, and patient populations evolve, these models can be updated with relatively little effort, ensuring they remain relevant.

The study was conducted using data from the U.S. Department of Veterans Affairs, which provided a rich, longitudinal record of millions of patients. While the specific population of veterans is predominantly male and older, the methodology developed here offers a general strategy for improving risk prediction in any large healthcare system. The researchers suggest that this approach could be applied to other chronic diseases and complex medical problems, providing a path toward more reliable and personalized medicine. By combining the best of different study designs, the team has shown that machine learning can move beyond simple pattern matching to create tools that truly understand the complexity of human health.