Missing Data Governance in Real-World Cardiovascular Risk Prediction: Effects on Machine Learning Performance and Interpretability
This study demonstrates that in real-world cardiovascular risk prediction, appropriate handling of missing data through imputation or hybrid strategies is critical for maintaining model reliability and interpretability, whereas uncritical deletion of incomplete records significantly degrades performance and overlooks the informative value of missingness itself.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Heart disease remains the leading cause of death worldwide, and doctors rely on early detection to save lives. To do this, they gather a vast amount of information about each patient: their age, blood pressure, cholesterol levels, and whether they have a history of smoking or diabetes. In recent years, computer programs known as machine learning have been trained to look at this information and predict who is at risk. These programs are excellent at finding patterns in clean, perfect data. However, the real world is rarely perfect. In actual hospitals, records are often messy. A doctor might forget to write down a blood pressure reading, a lab result might be missing, or a patient might have had a test done at a different clinic that never made it into the file. For a long time, scientists testing these computer programs have mostly used old, cleaned-up datasets that do not reflect this reality. They have focused on which computer algorithm is the smartest, often ignoring the fact that the data itself is incomplete. This raises a critical question: if a computer program is trained on perfect data, can it still work when it faces the messy, incomplete records found in a real hospital?
A team of researchers in Denizli, Türkiye, decided to answer this by building a study around the data itself rather than the computer program. They went directly to the cardiology department of a private hospital and pulled every single record from patients seen between 2019 and 2024. They did not clean the data first. They did not throw away the records with missing numbers. Instead, they took the raw, unfiltered information exactly as it sat in the hospital database, with all its gaps and inconsistencies. From these records, they selected nineteen key health parameters, such as age, sex, cholesterol, and various heart measurements, to see how well different computer programs could predict heart disease when forced to deal with missing information. They tested eleven different types of learning algorithms, ranging from simple statistical methods to complex, modern systems that build many small models to work together. Crucially, they did not just pick one way to fix the missing numbers. They tested four completely different strategies. One strategy filled in the blanks with estimated numbers. Another left the blanks empty to see if the computer could handle them on its own. A third strategy treated the missing values as a specific category of their own, essentially telling the computer, "We don't know this, but the fact that we don't know it might be important." The fourth strategy, which is common in many studies, simply deleted every single record that had even one missing number.
The results showed that how the researchers handled the missing data mattered far more than which computer program they chose. When they used the strategy of deleting incomplete records, the performance of every single computer model dropped significantly. The most accurate models fell from correctly identifying about 77 percent of cases to just 69 percent. This loss happened because deleting records removed so much valuable information and left the computer with too few examples to learn from. In contrast, the strategies that kept the incomplete records—either by filling in the gaps or by treating the missingness as a known fact—preserved the models' ability to predict heart disease accurately. The computer programs that performed best were those that could handle the messy data without throwing it away. Among the different types of algorithms tested, a family of systems known as boosting ensembles proved to be the most resilient. These systems, which combine many simple decision-makers to form a strong prediction, maintained high accuracy even when the data was incomplete. Simpler methods, like those based on finding the closest neighbors in a dataset, struggled the most when faced with missing values.
Perhaps the most surprising discovery came when the researchers looked closely at how the computer programs made their decisions. They used a special analysis tool to see which features the computer considered most important. They found that the computer was paying close attention to the fact that certain measurements were missing. Specifically, the indicator that a blood pressure reading was missing ranked as one of the most influential factors in predicting heart disease. This does not mean that having no blood pressure reading causes heart disease. Instead, it suggests that the pattern of missingness carries its own information. In this hospital, the fact that a blood pressure reading was not recorded might indicate something about the patient's situation, such as a specific type of visit or a socioeconomic factor that influenced which tests were ordered. The computer learned that this absence was a signal, just as real as a high cholesterol number. This finding challenges the idea that missing data is just a defect to be fixed or ignored. It shows that in real-world clinical settings, the way information is recorded—or not recorded—is part of the story the data tells.
The study concludes that for computer programs to be trusted in real hospitals, scientists must stop treating data quality as a minor detail to be cleaned up before the main work begins. The way missing information is handled is a fundamental part of the science. If researchers simply delete incomplete records, they risk building models that work well on paper but fail in practice. The most reliable approach is to use methods that respect the incompleteness of the data, either by filling in gaps carefully or by acknowledging that the absence of a measurement is itself a piece of information. By using a locally sourced dataset from Türkiye, the researchers demonstrated that models built on real, messy data are more representative of the actual conditions doctors face. They found that while the models were not perfect, they were honest and robust. The study argues that the future of medical prediction lies not in finding a slightly better algorithm to run on perfect data, but in building systems that can understand and learn from the imperfect reality of human health records.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.