EUVIMEDEUVIMED entdecken
Crossref

Reproducible Life‐Expectancy Eligibility Criteria for Oncology Studies: Uncertainty Quantification for Machine Learning‐Derived Survival Screening Rules

Vladimir Svetnik iD, Matteo Sesia iD, Michael Johnson, Peining Tao, Mehmet Burcu iD

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

DOI: 10.1002/sim.70811 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Statistics in Medicine, online erschienen am 14.11.2026. Verfasst von Vladimir Svetnik und 4 weiteren Autor:innen. Beteiligte Einrichtung: Biostatistics and Research Decision Sciences, Merck & Co., Inc., Rahway, New Jersey, USA. Hauptthemen laut MEDLINE: Machine Learning. Die Arbeit nennt 48 Literaturangaben.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Life‐expectancy eligibility criteria appear in approximately one quarter of oncology trials but typically rely on subjective clinician judgment and cannot be replicated in retrospective electronic health record studies. Survival prediction models provide a natural basis for objective, data‐driven operationalization of such criteria, but their practical use is complicated by four challenges: Possible model miscalibration; the need to quantify uncertainty, including about the realized operating characteristics in a random future cohort; the need to support flexible selection of the model and screening threshold by a clinician after inspecting the data; and the right‐censoring intrinsic to retrospective survival data. We propose a bootstrap‐based method that addresses all four challenges using an independent, right‐censored calibration dataset from the target population. It constructs confidence and prediction bands for cohort yield and positive predictive value as functions of the screening threshold, jointly valid over any desired collection of models and thresholds, and therefore compatible with exploratory, clinician‐driven analysis. The method is model‐agnostic, treating each survival model as an arbitrary black box without requiring it to be well calibrated. We apply it to data from over 47 000 patients with advanced non‐small cell lung cancer from the Flatiron Health Research Database, across different survival models and prediction horizons, and find that it yields stable and interpretable screening rules with reliable uncertainty quantification for future cohort‐level performance.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Machine LearningHumansUncertaintySurvival AnalysisLung NeoplasmsModels, StatisticalReproducibility of Results

Stichwörter der Autor:innen

model‐agnostic inferencepredictionsimultaneous inferencesurvival analysisuncertainty estimation

Autor:innen und Institutionen

  • Biostatistics and Research Decision Sciences, Merck & Co., Inc., Rahway, New Jersey, USA.
  • Departments of Data Sciences and Operations, and of Computer Science, University of Southern California, Los Angeles, California, USA.

ORCID-Profile: Vladimir Svetnik, Matteo Sesia, Mehmet Burcu

Zitieren

APA

Svetnik, V., Sesia, M., Johnson, M., Tao, P., & Burcu, M. (2026). Reproducible Life‐Expectancy Eligibility Criteria for Oncology Studies: Uncertainty Quantification for Machine Learning‐Derived Survival Screening Rules. Statistics in Medicine, 45(25-27). https://doi.org/10.1002/sim.70811

Vancouver

Svetnik V, Sesia M, Johnson M, Tao P, Burcu M. Reproducible Life‐Expectancy Eligibility Criteria for Oncology Studies: Uncertainty Quantification for Machine Learning‐Derived Survival Screening Rules. Statistics in Medicine. 2026;45(25-27). doi:10.1002/sim.70811

BibTeX
@article{Svetnik2026,
  author = {Svetnik, Vladimir and Sesia, Matteo and Johnson, Michael and Tao, Peining and Burcu, Mehmet},
  title = {Reproducible Life‐Expectancy Eligibility Criteria for Oncology Studies: Uncertainty Quantification for Machine Learning‐Derived Survival Screening Rules},
  journal = {Statistics in Medicine},
  year = {2026},
  volume = {45},
  number = {25-27},
  doi = {10.1002/sim.70811}
}
Kostenlos starten →