Center for Artificial Intelligence, Research, and Evaluation, IBM Watson Health, Cambridge, MA 02142.
Division of Integrative Biological and Behavioral Sciences, National Institute on Minority Health and Health Disparities, National Institutes of Health, Bethesda, MD 20892.
Ethn Dis. 2023 Mar 31;33(1):33-43. doi: 10.18865/1704. eCollection 2023 Jan.
INTRODUCTION/PURPOSE: Predictive models incorporating relevant clinical and social features can provide meaningful insights into complex interrelated mechanisms of cardiovascular disease (CVD) risk and progression and the influence of environmental exposures on adverse outcomes. The purpose of this targeted review (2018-2019) was to examine the extent to which present-day advanced analytics, artificial intelligence, and machine learning models include relevant variables to address potential biases that inform care, treatment, resource allocation, and management of patients with CVD.
PubMed literature was searched using the prespecified inclusion and exclusion criteria to identify and critically evaluate primary studies published in English that reported on predictive models for CVD, associated risks, progression, and outcomes in the general adult population in North America. Studies were then assessed for inclusion of relevant social variables in the model construction. Two independent reviewers screened articles for eligibility. Primary and secondary independent reviewers extracted information from each full-text article for analysis. Disagreements were resolved with a third reviewer and iterative screening rounds to establish consensus. Cohen's kappa was used to determine interrater reliability.
The review yielded 533 unique records where 35 met the inclusion criteria. Studies used advanced statistical and machine learning methods to predict CVD risk (10, 29%), mortality (19, 54%), survival (7, 20%), complication (10, 29%), disease progression (6, 17%), functional outcomes (4, 11%), and disposition (2, 6%). Most studies incorporated age (34, 97%), sex (34, 97%), comorbid conditions (32, 91%), and behavioral risk factor (28, 80%) variables. Race or ethnicity (23, 66%) and social variables, such as education (3, 9%) were less frequently observed.
Predictive models should adjust for race and social predictor variables, where relevant, to improve model accuracy and to inform more equitable interventions and decision making.
简介/目的：纳入相关临床和社会特征的预测模型可以深入了解心血管疾病（CVD）风险和进展的复杂相互关联的机制，以及环境暴露对不良结局的影响。本次有针对性的综述（2018-2019 年）的目的是研究目前的高级分析、人工智能和机器学习模型在多大程度上纳入了相关变量，以解决可能影响患者护理、治疗、资源分配和管理的偏见。
使用预设的纳入和排除标准在 PubMed 文献中进行检索，以确定并批判性地评估发表在英语期刊上的、针对北美一般成年人群 CVD 及其相关风险、进展和结局的预测模型的主要研究。然后评估这些研究在模型构建中纳入相关社会变量的情况。两名独立的审查员对文章进行了筛选，以确定其是否符合纳入标准。主要和次要的独立审查员从每篇全文文章中提取信息进行分析。如有分歧，将由第三名审查员进行裁决，并进行迭代筛选以达成共识。采用 Cohen's kappa 来确定评分者间的可靠性。
该综述共产生了 533 条独特的记录，其中 35 条符合纳入标准。这些研究使用先进的统计和机器学习方法来预测 CVD 风险（10 项，29%）、死亡率（19 项，54%）、存活率（7 项，20%）、并发症（10 项，29%）、疾病进展（6 项，17%）、功能结局（4 项，11%）和处置（2 项，6%）。大多数研究纳入了年龄（34 项，97%）、性别（34 项，97%）、合并症（32 项，91%）和行为风险因素（28 项，80%）等变量。种族或民族（23 项，66%）和社会变量（如教育程度，3 项，9%）则较少被观察到。
预测模型应根据种族和社会预测变量进行调整，在相关情况下，以提高模型准确性，并为更公平的干预措施和决策提供信息。
