An appropriate patient sample is essential to the integrity of any type of medical research. In observational studies, using an inappropriate sample can result in spurious associations unable to be replicated in prospective work. In randomized controlled trials (RCTs), testing an intervention on a non-representative group of patients can result in gaps between the efficacy of a treatment as observed in a trial and its effectiveness in clinical practice. Yet sample biases, often rooted in the historical exclusion of specific demographic groups, including older adults [1], patients with low socioeconomic status (SES) [2], racial minorities and women [3], tend to be consistent and pervasive. These exclusions are reflective of broader structural disparities in the American healthcare system, which have been highlighted by the Joint Commission as targets for assessment of quality of care [4] and by the Centers for Medicare and Medicaid Services as factors in determining reimbursement [5].

Each step of the sample selection process, from initial patient recruitment to exclusion criteria and patient attrition, holds both promise and peril for creating a representative sample. While strategies to create more inclusive, diverse and equitable clinical trial recruitment and designs have been developed in response to these failures [6], [7], [8], [9], less work has focused on how samples can be fundamentally changed by the steps that follow recruitment, and how these changes can propagate both structural and statistical bias.

As clinical artificial intelligence (AI) and machine learning (ML) applications surge in use, there is the danger that these tools will inherit these biases if trained on nonrepresentative datasets or if they fail to address these potential pitfalls of sample development. Many, for example, have highlighted instances of algorithmic discrimination [10], [11], [12], which often results from a lack of population representativeness or from the algorithm encoding structural biases already present in a dataset. An additional and under-explored dimension in AI studies is the input-level exclusion of patients due to poor-quality or missing data, which can lead to bias if there is a non-random disparity in data quality or availability among groups. Yet despite the potential for selection bias to propagate within AI algorithms, there are no standardized protocols for reporting participant characteristics and sample creation in medical AI research.

The introduction of “Data Cards” is a recent initiative aimed at changing this and enhancing transparency in AI studies by providing standardized details about a dataset's background, origin, and purpose [11], [12]. These cards offer an in-depth overview of a dataset, encapsulating 31 distinct facets, which include a variable list, descriptive statistics, and information about its intended application. Building on prior endeavors to boost AI transparency [15], [16], [17], [18], Data Cards are meant for general use in any AI project and are not tailored specifically for medicine. Thus, while they provide highly valuable insights for medical datasets, their approach could be enhanced by an additional component that tracks the evolution of patient cohort composition throughout all phases of study sample selection. Many medical studies employ tools like the CONSORT diagram for RCTs [19] and the STROBE diagram for observational studies [20] to accomplish this type of tracking. While these tools are useful, they primarily capture shifts in cohort size and do not track shifts in sample composition. Moreover, a dedicated tool that suits AI studies specifically is needed.

In this article, we advocate for the integration of a detailed participant flow diagram into the current Data Card framework for AI-based medical studies, enhancing its pursuit of transparency and promotion of health equity. In this flow diagram, we argue that it is essential not only to track the number of participants excluded at each phase of any study, but also to report changes in sociodemographic and clinical characteristics of the sample that are relevant to the study question. By doing so, we aim to mitigate and better understand potential statistical and structural biases in the application of AI. To illustrate the importance of tracking excluded participant characteristics through a flow diagram, we walk through various examples that demonstrate how biases can present in different stages of a study, from recruitment to exclusion criteria to input-level omissions to participant attrition during a study. Finally, we present a model for this updated flow diagram and an example of its implementation, envisioning this style of diagram not only as an augmentation to the Data Card but also as an element in medical AI studies generally. While we focus on the use of this flow diagram for AI-based studies here, we believe the concept of tracking cohort composition itself should be encouraged for any type of clinical study.