Bootstrapping-based Regularisation for Reducing Individual Prediction Instability in Clinical Risk Prediction Models

Researchers have developed a new bootstrapping-based regularization method that dramatically stabilizes AI predictions in clinical risk models. The technique embeds bootstrapping directly into neural network training, reducing prediction variability across different data samples. In tests on major datasets like GUSTO-I and Framingham, it slashed mean absolute differences by up to 68% (e.g., from 0.059 to 0.019) while maintaining high interpretability—a key weakness of traditional ensemble methods.