{"id":21589,"date":"2022-05-13T05:04:00","date_gmt":"2022-05-12T21:04:00","guid":{"rendered":"http:\/\/csccm.org.cn\/?p=21589"},"modified":"2022-05-13T05:26:06","modified_gmt":"2022-05-12T21:26:06","slug":"lancet-respir-med%e5%8f%91%e8%a1%a8%e8%bf%b0%e8%af%84%ef%bc%9a%e9%87%87%e7%94%a8%e5%9f%ba%e4%ba%8e%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0%e7%9a%84%e5%88%86%e7%b1%bb%e6%b3%95%e5%b0%86ards%e7%94%9f","status":"publish","type":"post","link":"https:\/\/csccm.org.cn\/?p=21589","title":{"rendered":"[Lancet Respir Med\u53d1\u8868\u8ff0\u8bc4]\uff1a\u91c7\u7528\u57fa\u4e8e\u673a\u5668\u5b66\u4e60\u7684\u5206\u7c7b\u6cd5\u5c06ARDS\u751f\u7269\u5b66\u8868\u578b\u5e26\u5230\u5e8a\u65c1"},"content":{"rendered":"\n

COMMENT| VOLUME 10, ISSUE 4<\/a>, P319-320, APRIL 01, 2022<\/p>\n\n\n\n

Bringing biological ARDS phenotypes to the bedside with machine-learning-based classifiers<\/h1>\n\n\n\n

Stephen Whebell, <\/a>J Zhang<\/h3>\n\n\n\n

Lancet Respir Med Published:January 10, 2022<\/h3>\n\n\n\n

DOI:https:\/\/doi.org\/10.1016\/S2213-2600(21)00492-6<\/a><\/h3>\n\n\n\n

The identification of distinct phenotypes within heterogeneous disease states is a key component of personalised medicine, enabling enrichment of clinical trials, better prognostication, and delivery of tailored treatments to well defined, homogeneous patient subgroups. In acute respiratory distress syndrome (ARDS), a clinically defined syndrome characterised by diffuse inflammatory alveolar insult and rapid onset hypoxaemia, subphenotypes have been identified with distinct clinical, radiological,1<\/a><\/sup>\u00a0genomic,2<\/a><\/sup>\u00a0and biomarker profiles.3<\/a><\/sup>\u00a0Despite demonstrating differing outcomes and treatment responses across independent cohorts, use of subphenotypes in clinical practice has remained limited. Phenotyping methods in research might not translate well to clinical implementation. For example, expense or unavailability of biomarkers typically used in research, and variable interpretations of radiological and clinical findings described in the literature, can hamper or prevent attempts at real-world validation.4<\/a><\/sup>\u00a0Intensive care units (ICUs) have differences in resource availability and local practices, and phenotyping approaches must have potential for implementation in a standard clinical workstream if they are to demonstrate widespread clinical utility.<\/p>\n\n\n\n

In\u00a0The Lancet Respiratory Medicine<\/em>, the study by Manoj Maddali and colleagues5<\/a><\/sup>\u00a0represents the latest piece in a cohesive body of research that identifies and validates biological subphenotypes of ARDS. The transfer of phenotype prediction from laboratory assay to the use of machine-learning techniques on clinical features available from most electronic health records (EHRs), brings us a step closer to personalisation of treatment in both clinical trials and in practice. The authors create multiple machine-learning models to perform a binary classification task in distinguishing hypoinflammatory and hyperinflammatory subphenotypes, with training, testing, and validation performed in independent and well separated prospective cohorts.<\/p>\n\n\n\n

Although robust and well reported, several considerations might limit real-world utility. A supervised learning approach was adopted, where ground truth labels were established using biological models previously tested using data from prospective observational cohorts.6<\/a><\/sup>\u00a0While supplying a validated target for prediction, the requirement of available biomarkers for generating labels also necessitates the use of curated research cohorts for model training, resulting in preselection of a study population that is unlikely to be representative of a diverse hospital cohort. Additionally, the generation of three models that make use of different data subsets is understandable, but does not result in classifiers that are robust to issues with data quality and variable missingness inherent to EHRs. Although tested on limited observational and EHR-derived cohorts, this approach falls short of developing a single model for use in prospective clinical evaluation, that is well calibrated to diverse data and populations. Finally, although data for training from most research cohorts are based on point-in-time measurements, clinical phenotypes have an unavoidable longitudinal dimension, where clinical attributes might fulfill ARDS criteria or cross phenotypes at different points in time. There is limited evidence of class stability,7<\/a><\/sup>with suggestion that temporal phenotypes provide most discrimination,8<\/a><\/sup>\u00a0but the best methods for mapping and detecting phenotypes over time are not well established.<\/p>\n\n\n\n

Alternative approaches to ARDS phenotyping use unsupervised learning methods, where data is presented agnostically without previous labelling of truth, and patients are algorithmically divided into distinct groups.9<\/a>,\u00a010<\/a><\/sup>\u00a0Phenotypes are not predefined, potentially reducing selection bias and inaccurate classification from inadequate biological signals. However, unsupervised methods are entirely dependent on quality and scope of available data to define potential latent phenotypes. Neither supervised nor unsupervised approaches for ARDS phenotyping have been tested in real-world clinical validation, and as such, are only promising proofs of concept.<\/p>\n\n\n\n

The existence of ARDS subphenotypes characterised by differing inflammatory characteristics is now well supported by available evidence.11<\/a><\/sup>\u00a0Maddali and colleagues see future utility in combining informatics and machine learning to identify these subphenotypes for trial recruitment, and in stratification for different treatment approaches. EHR availability is not ubiquitous, but given the accelerating digitisation of hospitals and ICUs, future widespread use of data-driven approaches is clearly foreseeable. However, several challenges bear consideration. First, any data used for model training must reflect real-world availability and circumstance, and must be closely examined for quality and bias, including representation of diverse patient groups. Machine-learning runs by association, and data bias will be reflected in algorithmic bias. Second, barriers to model deployment should be considered as part of any machine-learning research roadmap. Real-world deployment of models is reliant on EHR infrastructure capabilities, availability of requisite data streams, and feedback to allow continued model re-evaluation (a key safety mechanism when used to support clinical decision making). While EHR use is proliferating, proprietary software implementations and a lack of interoperability remain substantial barriers. Finally, adequate representation of phenotypes might require the integration of rich data from multiomic and unstructured narrative text sources into readily available EHR data.<\/p>\n\n\n\n

Overall, we applaud recent developments in this field, but highlight the need to consider robust machine-learning development and data practices that will support translation of ARDS subphenotype detection to clinical practice.<\/p>\n\n\n\n

\"\"\/<\/figure>\n","protected":false},"excerpt":{"rendered":"

COMMENT| VOLUME 10, ISSUE 4, P319-320, A […]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[24,23],"tags":[],"_links":{"self":[{"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/posts\/21589"}],"collection":[{"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=21589"}],"version-history":[{"count":1,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/posts\/21589\/revisions"}],"predecessor-version":[{"id":21590,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=\/wp\/v2\/posts\/21589\/revisions\/21590"}],"wp:attachment":[{"href":"https:\/\/csccm.org.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=21589"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=21589"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/csccm.org.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=21589"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}