EUVIMEDΑνακαλύψτε το EUVIMED
Open access · CC BY-NC-ND 4.0

Evaluating the Prediction Performance of Ensemble Super Learners

Emir Sevinc, Anders Munch, Pietro Ravani, Thomas Alexander Gerds

Statistics in Medicine · 2026-11 · 45 (25-27)

DOI: 10.1002/sim.70829 ↗

At a glance

Journal article in Statistics in Medicine, published online on 2026-11-19. Written by Emir Sevinc and 3 co-authors. 24 references. The full text is freely available (open access, CC BY-NC-ND 4.0). Funded by: Alberta Innovates.

ΠερίληψηΒιβλιογραφικά στοιχείαCiteRelated studies

Abstract (original language)

The Super Learner uses cross‐validation to evaluate a prespecified library of learners, such as regression models and random forests. The Discrete Super Learner selects the single best‐performing learner, whereas an Ensemble Super Learner combines the learners' cross‐validated predictions using estimated weights. Asymptotically, an Ensemble Super Learner whose weights minimize the cross‐validated loss performs at least as well as the Discrete Super Learner, yet selecting a single learner remains common in applied risk prediction. This article asks when the Discrete Super Learner is adequate and when combining learners offers a meaningful gain. We revisit KDpredict, a Discrete Super Learner for predicting kidney failure and death in adults with moderate or severe chronic kidney disease (CKD), and compare it with six Ensemble Super Learner weighting strategies, including nonnegative least squares, using both the full training data and random subsamples of decreasing size. A simulation illustrates a case of misspecification, through which combining learners can help. With the full training data, all strategies performed nearly identically. In subsamples of 100 and 200 patients, several ensembles achieved modestly higher accuracy, although all strategies performed poorly in absolute terms. In simulation, ensembles outperformed the Discrete Super Learner when each candidate omitted a different predictor, and nonnegative least squares performed well in all settings. The Discrete Super Learner was adequate for KDpredict, and the benefit of ensembling appears context‐dependent.

Funding

  • Alberta Innovates

Cite

APA

Sevinc, E., Munch, A., Ravani, P., & Gerds, T. A. (2026). Evaluating the Prediction Performance of Ensemble Super Learners. Statistics in Medicine, 45(25-27). https://doi.org/10.1002/sim.70829

Vancouver

Sevinc E, Munch A, Ravani P, Gerds TA. Evaluating the Prediction Performance of Ensemble Super Learners. Statistics in Medicine. 2026;45(25-27). doi:10.1002/sim.70829

BibTeX
@article{Sevinc2026,
  author = {Sevinc, Emir and Munch, Anders and Ravani, Pietro and Gerds, Thomas Alexander},
  title = {Evaluating the Prediction Performance of Ensemble Super Learners},
  journal = {Statistics in Medicine},
  year = {2026},
  volume = {45},
  number = {25-27},
  doi = {10.1002/sim.70829}
}
Start for free →