A Multisite Super Learner for Survival Prediction with Limited Target-Site Data

Survival prediction models developed in multi-site studies may transport poorly to a new site because of between-site heterogeneity in covariate effects and baseline hazards. Pooled models may fail to capture site-specific risk patterns, while site-specific models can be unstable in small cohorts. We propose a Multi-Site Super Learner (MSSL), a domain-adaptive ensemble for survival prediction that combines diverse pooled and site-specific base learners. A meta-learner optimizes the ensemble weights using source data alone or, if available, a small target-site calibration sample. In simulations, when even a small target-site sample was available, MSSL substantially reduced prediction error relative to a pooled super learner under moderate to large between-site heterogeneity and generally matched or improved upon models fit to the target-site data alone. It did so by downweighting dissimilar or noisy models and upweighting models from structurally similar sites. Without target-site data, MSSL did not improve upon the pooled super learner. In an application to predicting dementia progression using the National Alzheimer's Coordinating Center (NACC) Uniform Data Set across 10 sites, MSSL matched or outperformed standard super learning using pooled data in 9 of 10 sites. MSSL offers a practical framework for adapting multi-site survival prediction models to a new site when a small calibration sample is available.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-11-18
DOI
https://doi.org/10.48550/arxiv.2611.23675
Primary Topic
Statistical Methods and Inference
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Multisite Super Learner for Survival Prediction with Limited Target-Site Data

Prasad Patil, Henry Bayly, Yorghos Tripodis, Haolin Li
arXiv (Cornell University)
Statistical Methods and Inference
preprint

A Multisite Super Learner for Survival Prediction with Limited Target-Site Data

Prasad Patil, Henry Bayly, Yorghos Tripodis, Haolin Li
preprint en

Abstract

Survival prediction models developed in multi-site studies may transport poorly to a new site because of between-site heterogeneity in covariate effects and baseline hazards. Pooled models may fail to capture site-specific risk patterns, while site-specific models can be unstable in small cohorts. We propose a Multi-Site Super Learner (MSSL), a domain-adaptive ensemble for survival prediction that combines diverse pooled and site-specific base learners. A meta-learner optimizes the ensemble weights using source data alone or, if available, a small target-site calibration sample. In simulations, when even a small target-site sample was available, MSSL substantially reduced prediction error relative to a pooled super learner under moderate to large between-site heterogeneity and generally matched or improved upon models fit to the target-site data alone. It did so by downweighting dissimilar or noisy models and upweighting models from structurally similar sites. Without target-site data, MSSL did not improve upon the pooled super learner. In an application to predicting dementia progression using the National Alzheimer's Coordinating Center (NACC) Uniform Data Set across 10 sites, MSSL matched or outperformed standard super learning using pooled data in 9 of 10 sites. MSSL offers a practical framework for adapting multi-site survival prediction models to a new site when a small calibration sample is available.

arXiv (Cornell University)
Statistical Methods and Inference
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon