Artificial intelligence (AI)–enabled electrocardiogram (ECG; AI-ECG) interpretation has emerged as a transformative tool in cardiovascular diagnostics. These models, particularly those built using deep learning architectures, have demonstrated high accuracy in identifying cardiac conditions such as atrial fibrillation, heart failure, and left ventricular dysfunction [1,2]. Given the global burden of cardiovascular disease, particularly in low-resource settings, AI-ECG tools offer the potential to enhance access to timely diagnosis, especially in areas with limited cardiology expertise [3].

However, as AI adoption accelerates, concerns about algorithmic bias have gained prominence. Algorithmic bias refers to systematic disparities in model performance across subpopulations, often arising from unrepresentative training data, flawed modeling assumptions, or inappropriate deployment environments [4-6]. These biases can lead to unequal diagnostic accuracy and risk misdiagnosis, particularly in underrepresented or structurally marginalized groups. In cardiology, for instance, tools trained predominantly on datasets composed predominantly of White male individuals have underperformed in detecting myocardial infarction in women and individuals of African descent [7,8]. Physical diagnostic tools such as pulse oximeters have similarly demonstrated racial bias in measurement accuracy [9].

Bias may emerge across multiple stages of the AI development pipeline, from data collection and labeling to model training, evaluation and deployment, and clinical integration. Recent frameworks categorize these into 5 key stages: data, model, evaluation, deployment, and postdeployment bias [10]. Addressing each of these stages is critical to ensuring fairness, safety, and clinical utility. High-income countries (HICs) have started to institutionalize fairness evaluations in medical AI development, with regulatory and ethical guidelines now recommending transparency, community engagement, and subgroup performance monitoring [11-13].

Despite this progress, significant gaps persist in understanding algorithmic bias within AI-ECG models, particularly in low- and middle-income countries (LMICs). Most existing studies and validations are based on datasets from North America and Europe, often excluding the very populations that might benefit most from AI-based screening tools [14]. This mismatch, described as “health data poverty” or “digital colonization,” raises concerns that AI tools exported to LMICs may underperform or cause unintended harm [15,16].

To date, there is no comprehensive synthesis of how bias manifests in AI-powered ECG interpretation. Existing reviews focus primarily on the diagnostic accuracy of AI-ECG for specific conditions and seldom report performance stratified by demographic or geographic subgroups [17]. A few studies have started to explore fairness or subgroup performance. However, findings are scattered, and no previous scoping review has mapped the full range of bias-related evidence or mitigation efforts in this field. It remains unclear which populations are most affected, at what stages bias arises, and what strategies have been attempted to reduce disparities.

With the increasing deployment of AI-ECG tools, including in LMICs, there is an urgent need to map the evidence on algorithmic bias in this domain systematically. Doing so will support the equitable development and safe deployment of AI-based ECG technologies across diverse populations and health systems. Early identification and mitigation of such bias are essential to diagnostic equity.

This scoping review aims to comprehensively map and characterize the current evidence related to algorithmic bias in AI-powered ECG interpretation across the AI life cycle.

This review is guided by the population-concept-context (PCC) framework, which also informs the eligibility criteria:

By addressing the following research questions, this review aims to provide a comprehensive overview of the field:

This scoping review will inform future research, regulation, and implementation of AI-based cardiac diagnostics to ensure they serve all populations equitably.

This study will be conducted as a scoping review, following the framework by Arksey and O’Malley [18], with enhancements by Levac et al [19] and methodological guidance from the Joanna Briggs Institute [20]. This review will be reported following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) checklist [21].

We will undertake the five standard stages of scoping reviews: (1) identifying the research question; (2) identifying relevant studies; (3) selecting studies; (4) charting the data; and (5) collating, summarizing, and reporting results.

We will skip the optional sixth stage (stakeholder consultation) because of resource and timing constraints.

Eligibility criteria are defined using the PCC framework recommended for scoping reviews.

Consistent with Joanna Briggs Institute and PRISMA-ScR guidance, we will not conduct a formal quality appraisal. However, we will document evident methodological limitations (eg, small sample sizes or limited subgroup analysis) that may affect interpretation.

In studies reporting AI model development or validation, we will extract selected bias-related indicators aligned with the Prediction model Risk of Bias Assessment Tool–AI framework [24]. Although full risk-of-bias scoring will not be performed, this approach will support structured identification of potential bias-related to participant selection, data quality, model evaluation, and fairness reporting and will help flag limitations affecting subgroup generalizability.

This review involves secondary analysis of publicly available literature and does not include any original research involving human participants. Therefore, it does not require approval from an institutional review board or ethics committee. No individual-level, sensitive, or identifiable data will be accessed or analyzed during this study, consistent with established ethical standards for systematic reviews [23].

The results of this scoping review will be submitted for publication in a peer-reviewed journal focused on digital health, AI in medicine, or cardiovascular care. In addition to traditional academic dissemination through journal articles and conference presentations, findings will also be shared with key stakeholders, such as health care AI developers, clinicians, and policymakers, via webinars, social media, and professional networks. Special emphasis will be placed on sharing the findings in ways that are accessible to stakeholders in LMICs, where bias in AI tools can be particularly impactful.

The database search was conducted in August and September 2025. As displayed in Figure 1, study screening and data extraction took place from November 2025 to December 2025. Data analysis and manuscript submission are planned for January and February 2026.

This protocol is reported following the PRISMA-P (Preferred Reporting Items for Systematic Review and Meta-Analyses Protocols) 2015 guidelines.

The literature search strategy was finalized and executed across the specified databases in August and September 2025. Following duplicate removal, title and abstract screening was completed by two independent reviewers, leading to the identification of 148 records for full-text review. The full-text assessment concluded in November 2025, and 38 studies met the inclusion criteria for the qualitative synthesis. Data extraction was conducted using a standardized charting form between November and December 2025, with all extracted entries verified by a second reviewer. Descriptive and thematic analysis of the collated evidence is currently underway, with completion targeted for February 2026. The study selection process is summarized in a PRISMA flow diagram (Figure 1).

We plan to complete the main scoping review manuscript by the end of February 2026. This will include final synthesis of extracted data, preparation of summary evidence tables and bias classification matrices, and drafting of the Results and Discussion sections. The completed review manuscript will be submitted for peer-reviewed publication shortly thereafter, consistent with PRISMA-ScR guidance and journal requirements.