Machine learning beats regression for complex treatment effects in large trials.

Doctors want to predict individualized treatment effects, which means guessing how well a medical treatment will work for a specific person. Researchers often combine data from multiple medical trials using a two-stage method. In the past, this usually relied on standard regression equations. The researchers created a broader framework that allows scientists to use any statistical or machine learning tool to predict continuous outcomes. First, scientists build prediction models inside each trial separately. Second, they combine the results from all the trials to predict outcomes for new patients. To see how well this works, the team ran computer simulations. They found that machine learning performed best when sample sizes were large and the relationships were complex. However, regular regression performed better when the relationships were simple and linear. Because the true pattern in real data is unknown, researchers should test their models with internal validation before deciding which one to use.

Why it matters

Doctors and researchers can now predict which patients benefit most from a treatment using any statistical or machine learning model, not just simple regression. This matters because the best method depends on the data: complex interactions need machine learning, while simple patterns are fine with regression. The framework gives clear guidance on how to combine results from multiple trials to make these predictions.

Questions answered