An explainable machine learning model integrating multidimensional clinical information for early identification and probability stratification of Non-ST-Segment Elevation Myocardial Infarction

Every year, millions of people rush to emergency rooms with chest pain, a symptom that signals a potential heart attack. Doctors face a difficult puzzle: distinguishing between a heart attack where the heart muscle is actively dying and a condition called unstable angina, where the heart is starving for oxygen but the muscle has not yet suffered permanent damage. The standard tool for this job is a blood test that measures a protein called high-sensitivity cardiac troponin I. When heart cells are damaged, they leak this protein into the bloodstream. While this test is excellent at spotting severe damage, it often leaves doctors in a gray area when the levels are only slightly elevated. In these uncertain moments, the protein alone cannot always tell if the patient is having a full-blown heart attack or just severe chest pain, making it hard to decide who needs immediate, aggressive treatment and who can be monitored more closely.

A team of researchers at Beijing Jishuitan Hospital has developed a new way to navigate this uncertainty. They created a computer model that acts like a highly trained assistant, looking not just at the single protein test, but at a wide array of routine blood work and patient history all at once. By feeding the computer information about inflammation, blood clotting, kidney function, and the timing of symptoms, alongside the standard protein test, the model learns to spot subtle patterns that a human eye might miss. The goal was not to replace the doctor, but to provide a clearer, more personalized probability of a heart attack, turning a vague guess into a calculated risk that can guide life-saving decisions.

The researchers began by gathering data from nearly 3,700 patients who had been admitted to their hospital between 2021 and 2025 with suspected heart issues. They carefully reviewed the medical records of each person, separating them into two groups: those who were ultimately diagnosed with a non-ST-segment elevation myocardial infarction, the specific type of heart attack they wanted to identify, and those who had unstable angina. From a massive list of 124 possible clues found in the patients' records, the team used a sophisticated selection process to find the 16 most useful pieces of information. These included the standard protein test, but also markers for inflammation, blood sugar, kidney health, and how long the patient had been feeling pain before arriving at the hospital.

They then trained a computer algorithm, a type of machine learning system known as a gradient boosting decision tree, to learn the difference between the two groups. This system works by building a series of simple decision rules that, when combined, create a powerful predictor. The computer was tested on a group of patients it had never seen before to ensure it could handle real-world situations. The results showed that the model was remarkably accurate, correctly identifying the vast majority of heart attack cases. More importantly, when the researchers compared their new model to the standard method of using only the protein test, they found that the new approach was significantly better at estimating the actual risk for each individual patient. It reduced the number of errors in prediction and was much better at reclassifying patients into the correct risk groups, ensuring that fewer people with heart attacks were missed.

One of the most valuable aspects of this work is that the model is explainable. Unlike many computer systems that act as a "black box," giving an answer without saying why, this model can show exactly which factors pushed the prediction one way or the other. The researchers used a method to visualize how the computer weighed each piece of information. They found that while the protein test remained the most important clue, the other factors played a crucial role in refining the answer. For instance, if a patient had a slightly elevated protein level, the computer would look at their inflammation markers and kidney function to decide if that slight elevation was a sign of a serious heart attack or just a minor issue. The model learned that the importance of these extra clues changed depending on how high the protein level was. In cases where the protein was only slightly high, the other blood tests became much more critical for making the right call.

To make this tool useful for doctors, the team built a simple online calculator. A physician can enter the 16 routine numbers from a patient's blood test and history, and the system instantly generates a probability score. It then breaks this score into three clear categories: low risk, intermediate risk, and high risk. For patients in the low-risk group, the model suggests a very low chance of a heart attack, potentially allowing for a quicker discharge or less intensive monitoring. For those in the high-risk group, the probability is so high that it supports immediate, aggressive treatment. The middle group, which represents a significant portion of patients, is flagged for closer observation and further testing, acknowledging that the answer is not yet clear. This approach helps doctors move away from a simple yes-or-no diagnosis and instead provides a nuanced view of risk that matches the complexity of the human body.

The study suggests that by combining the standard protein test with a broad view of the patient's overall health, doctors can make more precise decisions earlier in the emergency room. The model does not claim to be a perfect crystal ball; it is a tool designed to support clinical judgment, not replace it. The researchers emphasize that their work is based on data from a single hospital and that the model needs to be tested in other settings to confirm its reliability across different populations. However, the findings offer a promising path forward, showing that the future of heart attack diagnosis may lie in the ability to synthesize many small, routine pieces of information into a single, clear picture of a patient's risk.