Program in Child Health Evaluative Sciences, The Hospital for Sick Children, Toronto, ON, Canada; Institute of Health Policy, Management and Evaluation, University of Toronto, Toronto, ON, Canada.
Institute of Health Policy, Management and Evaluation, University of Toronto, Toronto, ON, Canada; KITE-Toronto Rehabilitation Institute, University Health Network, Toronto, ON, Canada.
Int J Med Inform. 2022 Apr;160:104706. doi: 10.1016/j.ijmedinf.2022.104706. Epub 2022 Jan 31.
Machine learning (ML) has been increasingly used in clinical medicine including studies focused on Clostridioides difficile infection (CDI) to inform to clinical decision making. We aimed to summarize ML choices in studies that used ML to predict CDI or CDI outcomes.
We searched Ovid MEDLINE, Ovid EMBASE, Web of Science, medRxiv, bioRxiv and arXiv from inception to March 18, 2021. We included fully published studies that used ML where CDI constituted the study population, exposure or outcome. Two reviewers independently identified studies and abstracted outcomes. We summarized study characteristics and approaches to CDI definition and ML-specific modelling.
Forty-three studies of prediction (n = 21), classification (n = 17) or inference (n = 5) were included. Approaches to defining CDI were labelling during a clinical study or chart review (n = 21), electronic phenotyping (n = 13) or not specified (n = 9). None of the studies using an electronic phenotype described phenotype validation. Almost all studies (n = 41, 95%) conducted supervised ML and the most common ML algorithms were penalized logistic regression (n = 20, 47%) and classification tree (n = 17, 40%). Approaches to feature selection and dimension reduction were heterogeneous. Metrics were evaluated in a held-out test set in 16 (37%) studies; only seven used a time-based split. In terms of reporting quality assessment, the most poorly reported items were data leakage prevention (n = 0, 0%), code availability (n = 8, 19%) and class imbalance management (n = 12, 43%).
While many studies have used ML to investigate CDI or CDI outcomes, electronic phenotyping of CDI was uncommon and phenotype validation was not reported in any study. Methodological approaches were heterogeneous. Validating CDI electronic phenotypes, evaluating performances of CDI models during a silent trial and deploying a CDI classifier to guide clinical practice are important future goals.
机器学习（ML）已越来越多地应用于临床医学，包括针对艰难梭菌感染（CDI）的研究，以辅助临床决策。本研究旨在总结使用 ML 预测 CDI 或 CDI 结局的研究中的 ML 选择。
我们检索了 Ovid MEDLINE、Ovid EMBASE、Web of Science、medRxiv、bioRxiv 和 arXiv 从建库到 2021 年 3 月 18 日的数据。我们纳入了使用 ML 且 CDI 作为研究人群、暴露或结局的完全发表的研究。两名审查员独立识别研究并提取结局。我们总结了研究特征以及 CDI 定义和 ML 特定建模的方法。
共纳入 43 项关于预测（n=21）、分类（n=17）或推断（n=5）的研究。CDI 的定义方法包括临床研究或病历回顾时的标记（n=21）、电子表型（n=13）或未说明（n=9）。没有一项使用电子表型的研究描述了表型验证。几乎所有研究（n=41，95%）都进行了有监督的 ML，最常见的 ML 算法是惩罚逻辑回归（n=20，47%）和分类树（n=17，40%）。特征选择和降维方法存在差异。16 项研究（37%）在独立测试集上评估了指标，只有 7 项研究使用了基于时间的分割。在报告质量评估方面，数据泄露预防（n=0，0%）、代码可用性（n=8，19%）和类别不平衡管理（n=12，43%）的报告最差。
尽管许多研究已经使用 ML 研究 CDI 或 CDI 结局，但 CDI 的电子表型并不常见，并且在任何研究中都没有报告表型验证。方法学方法存在差异。验证 CDI 电子表型、在静默试验期间评估 CDI 模型的性能以及部署 CDI 分类器以指导临床实践是未来的重要目标。
