Clinical prediction with localized modeling using similarity-based cohorts: A scoping review

Abstract

AbstractBackgroundApplications of classical case-based reasoning (CBR) have given rise to a family of techniques we call “localized models”, in which a statistical model is fitted to a neighborhood of labeled cases matched by similarity to a target case. We aim to describe clinical and health applications of localized models to date and propose a general framework for their design and evaluation.MethodsWe searched four bibliographic platforms during 2021 July 19–22, updated 2024 January 24. We set four eligibility criteria to identify applications of localized models to clinical and health tasks. Two authors divided title/abstract screening and reviewed screened entries for inclusion. We discussed settings, tasks, and tools; identified and tabulated themes; and synthesized the methods into a general framework.ResultsOf 1,657 search results, 360 were reviewed, then combined with 43 publications that seeded the review and 1 obtained by citation tracking. 27 were included, published 1997–2022. The specificity of search terms was poor, and inter-rater reliability was low. Almost all models were predictive, the most common tasks being prognosis and diagnosis. Most studies used clinical, occasionally laboratory and image, data. Several addressed memory and runtime costs. A general technique that specializes to most of those reviewed involved matching, retrieval, fitting, and evaluation steps that could optionally be supervised, optimized, or recursively performed.ConclusionsLocalized models have potential to improve the performance of clinical decision support tools while maintaining interpretability, but rigorous comparisons to competing methods must be conducted and computational hurdles must be overcome. We hope that our review will spur future work on efficiency, reproducibility, and user needs.