An Explainable, Robust, and Empirically-Validated Stacked Ensemble (RXTG-Stack) for Cardiovascular Risk Prediction Across Heterogeneous Datasets

Imagine you are a detective trying to solve a mystery, but instead of a crime scene, you are looking at a person's body to figure out if they are at risk of a heart attack. For a long time, doctors have been the detectives, using tools like blood pressure cuffs and heart monitors to gather clues. But sometimes, the clues are tricky, and the human brain can get tired or miss a subtle pattern hidden in the data. This is where a new kind of detective enters the scene: Artificial Intelligence (AI). Think of AI as a super-smart assistant that can read thousands of medical records in a second. However, there's a catch. Most AI assistants are like "black boxes"—they give you an answer, but they won't tell you why they think that way. If a doctor can't understand the AI's reasoning, they can't trust it with a patient's life. So, the big challenge isn't just making an AI that is smart; it's making one that is smart and honest about how it thinks.

This paper introduces a new AI detective team called RXTG-Stack. The researchers built this team to predict heart disease by combining the strengths of four different AI "specialists" (Random Forest, XGBoost, TabNet, and Gradient Boosting) and then having a wise "team leader" (Logistic Regression) make the final call based on what the specialists say. The team is designed to be "leakage-free," meaning it doesn't use information from the test set during training, and it uses a special tool called SHAP to explain its decisions in plain English. The researchers tested this team on two different groups of patients: a small, carefully checked group of 1,000 people and a massive group of nearly 70,000 people. They found that RXTG-Stack was incredibly accurate, especially on the smaller group where it got it right 98.5% of the time. But more importantly, they proved that the team is fair, doesn't get confused by small changes in the data, and can explain exactly which clues (like chest pain type or blood pressure) led to its conclusion.

The Team of Detectives

Imagine you are trying to predict if a car will break down. You could ask one mechanic, but what if that mechanic is great at engines but bad at electronics? The RXTG-Stack team solves this by hiring four different experts.

• Random Forest is like a group of friends who each look at the car from a different angle and vote.
• XGBoost and Gradient Boosting are like a coach who learns from every mistake the previous coach made, getting smarter with every step.
• TabNet is a high-tech expert that pays attention to the most important parts of the car, ignoring the noise.

Instead of just letting these experts shout their opinions, the team uses a "meta-learner" (the team leader) to listen to all of them. The leader doesn't just pick the loudest voice; it learns how much to trust each expert based on their past performance. This happens in a special way called "Out-of-Fold" stacking. Think of it like a practice exam: the experts take turns testing on a group of students they haven't seen before, so they can't use information from the test set during training. This ensures the final team leader gets a truly honest report on how well each expert actually performs.

The Results: How Good is the Team?

The researchers put this team to the test on two very different datasets.

• The "Clinically Curated" Group (CVD-1K): This was a smaller group of 1,000 patients with very detailed, high-quality medical records. Here, the RXTG-Stack team was a superstar. It achieved 98.5% accuracy, meaning it was wrong only about 15 times out of 1,000. It also scored a 0.999 on a scale called ROC-AUC (which measures how good the model is at telling sick people from healthy ones), which is practically perfect.
• The "Population" Group (Cardio-68K): This was a massive group of nearly 70,000 people, but their data came from surveys and self-reports, which can be messier and less precise. Even with this "noisier" data, the team performed well, reaching 77.0% accuracy and an ROC-AUC of 0.803. While this isn't as high as the first group, it was still better than any single expert model the researchers tried on its own.

The paper suggests that the difference in scores isn't because the AI is bad, but because the data itself is different. The small group had very clear, clinical clues (like specific heart test results), while the large group relied on people remembering their own habits, which is harder to predict.

Why Trust the AI? (The "Explainable" Part)

The most exciting part of this paper isn't just that the AI is accurate, but that it is explainable. In the past, AI might say, "This patient is at risk," but couldn't say why. RXTG-Stack uses a tool called SHAP to break down the decision.

• Global View: It showed that for the small dataset, the most important clues were the slope of the peak exercise ST-segment (a specific heart test result), chest pain type, resting blood pressure, and serum cholesterol.
• Local View: For individual patients, it can draw a "force plot" showing exactly which factors pushed the prediction toward "sick" and which pushed it toward "healthy." For example, for one patient, high blood pressure and a specific chest pain type were the main reasons for a "high risk" prediction, while their active lifestyle helped lower the risk slightly.

The "Stress Test" (Empirical Verification)

The researchers didn't just stop at accuracy; they put the AI through a rigorous "stress test" to see if it was robust and fair.

• Robustness: They slightly tweaked the data (like adding a tiny bit of noise to a blood pressure reading) to see if the AI would panic and change its answer. The team stayed stable 99.3% of the time on the small dataset and 94.8% of the time on the large one, even when the data was nudged by ±3%.
• Fairness: They checked if the AI treated men and women differently. The difference in how often it predicted risk for each group was tiny (at most 0.026), which is well below the threshold for concern.
• Confidence: They used a method called "split-conformal prediction" to ensure that when the AI says it's 90% sure, it actually is. The results matched the target perfectly (0.900 and 0.899 coverage).

What This Means

The paper concludes that RXTG-Stack is a powerful, reliable tool for predicting heart disease. It suggests that by combining different types of AI and forcing them to explain their work, we can build systems that doctors might actually trust. However, the authors are careful to note that this was tested on two specific datasets. They suggest that before this becomes a standard tool in hospitals, it needs to be tested on even bigger, real-world medical records from many different hospitals to make sure it works everywhere. For now, it stands as a strong proof that we can build AI that is not only smart but also transparent and fair.