Original title: Reproducible benchmarking of four machine-learning classifiers on a large synthetic heart-disease competition dataset
Raed M. Ennab · September 2026
Computer models can predict diseases, but tests often cheat by accidentally leaking answers during training. Researchers wanted a fair contest. They tested four machine-learning models on a public synthetic heart-disease competition dataset to see which predicted risk best. To keep the contest completely fair, they used strict controls. They split the data into five chunks, hiding test sections so models could never learn from them in advance. They also checked for duplicate records and ran random control tests to ensure no answers leaked through. CatBoost scored the highest discrimination accuracy, followed closely by LightGBM, HistGradientBoosting, and regularized logistic regression. All four models achieved high accuracy and closely matched real probabilities. The complex gradient-boosting systems beat basic logistic regression by only a tiny margin. The researchers emphasize that their main achievement is a fully rerunnable test setup, and whether these models work on real human patients remains untested.
You are choosing between four routes to school by timing them on a quiet Sunday, and they all come out almost identical. The problem is that Monday's traffic and weather are different, so the fastest route on Sunday might not be fastest then. Here, the researchers time four machine-learning models on a large artificial heart-disease dataset, and all four distinguish the disease almost equally well. But because the data is artificial, they cannot say which model will win on real patients, just like a Sunday route can mislead you on a Monday.
This matters because it shows that with plenty of clean data, different machine-learning methods perform nearly the same, so picking one may matter less than rigorous testing. It also exposes a gap: synthetic data can fairly compare algorithms, but it cannot tell doctors how a model will behave on real patients. Anyone reading model comparisons should check whether the data reflects the people they care about.
This benchmark does not establish how any of the four models will perform on real patients; the synthetic dataset cannot capture the full complexity of clinical populations. The results apply only to this dataset and this leakage-controlled framework. The study does not propose a new algorithm and does not validate any model for clinical use.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Reproducible benchmarking of four machine-learning classifiers on a large synthetic heart-disease competition dataset — Raed M. Ennab — PLoS ONE — doi:10.1371/journal.pone.0354304
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.