Abstract

Clinical algorithms have become indispensable tools in contemporary healthcare, offering advanced capabilities for accurate diagnosis, personalized treatment planning, and efficient patient management. However, their increasing integration raises profound concerns regarding the potential for perpetuating and exacerbating existing health disparities, particularly among racial and ethnic minorities, individuals of different genders, and those from diverse socioeconomic backgrounds. This comprehensive report meticulously examines the pervasive impact of inherent biases within clinical algorithms, with a particular focus on the historical and ongoing use of race-based adjustments across critical areas such as lung function assessment, kidney disease evaluation, and obstetric care. Furthermore, it delves into a broader spectrum of bias sources, encompassing subtle gender biases, the profound influence of socioeconomic status, and critical issues related to data quality, representation, and algorithmic design choices. Crucially, this report proposes a robust framework of best practices for the ethical and equitable development, rigorous auditing, and responsible implementation of clinical algorithms, aiming to foster healthcare systems that are both technologically advanced and inherently just. It emphasizes the necessity of multidisciplinary collaboration, transparent methodologies, continuous monitoring, and policy interventions to ensure these powerful tools serve to advance, rather than hinder, the pursuit of health equity.

1. Introduction

The twenty-first century has witnessed a profound transformation in medical practice, largely driven by the pervasive integration of computational technologies, most notably artificial intelligence (AI) and machine learning (ML), into various facets of healthcare. Clinical algorithms, the operational embodiment of these technologies, have emerged as powerful decision-support tools, promising to revolutionize how diseases are diagnosed, prognoses are made, treatments are tailored, and healthcare resources are allocated [Topol, 2019]. By analyzing vast and complex datasets—ranging from electronic health records (EHRs) and medical images to genomic sequences and patient-reported outcomes—these algorithms purport to derive insights that surpass human cognitive capacity, leading to more precise, efficient, and standardized clinical decision-making [Rajkomar et al., 2018].

Despite this immense potential for innovation and improvement in patient outcomes, a growing chorus of concerns has arisen regarding the ethical implications and potential pitfalls of relying on these sophisticated tools. A central apprehension revolves around the inherent biases that can be encoded within these algorithms, often inadvertently, during their development and deployment. These biases, when left unaddressed, carry the significant risk of perpetuating and even amplifying existing health inequities, particularly affecting marginalized and vulnerable populations [Char et al., 2018]. The very data upon which these algorithms are trained—historical medical records, clinical guidelines, and population health statistics—often reflect deeply entrenched societal biases and structural inequalities that have historically disadvantaged specific racial, ethnic, gender, and socioeconomic groups. Consequently, algorithms that are not meticulously designed, scrutinized, and validated for fairness can inadvertently embed and operationalize these systemic prejudices, leading to differential and suboptimal care for those who are already underserved.

This report aims to comprehensively explore the multifaceted nature of bias in clinical algorithms. It will delve into specific instances where race-based adjustments have demonstrably led to disparate treatment, such as in kidney function assessment, pulmonary diagnostics, and obstetrics. Beyond race, the analysis will extend to other critical dimensions of bias, including gender, socioeconomic status, and fundamental issues of data quality and representativeness. Ultimately, this report seeks to illuminate the profound implications of these biases for patient care, trust in healthcare systems, and legal and ethical frameworks, while simultaneously proposing actionable best practices and policy considerations. The overarching goal is to foster the development and implementation of clinical algorithms that uphold the principles of justice, equity, and patient-centered care, ensuring that technological advancements truly benefit all members of society [Emanuel & Wachter, 2019].

2. The Role of Clinical Algorithms in Healthcare

Clinical algorithms operate as sophisticated computational frameworks designed to process vast amounts of medical data, identify patterns, and generate insights that assist healthcare professionals in various decision-making processes. Their utility spans the entire clinical spectrum, from initial patient presentation to long-term disease management, fundamentally reshaping the landscape of modern medicine. The efficacy and growing adoption of these algorithms are rooted in their ability to handle complexity, enhance objectivity (theoretically), and offer predictive capabilities far beyond human capacity alone.

2.1. Diagnostic Assistance

Algorithms are increasingly pivotal in diagnostic processes, analyzing diverse patient data to aid in identifying diseases. For example, deep learning algorithms excel at image recognition, interpreting radiographic scans (X-rays, CTs, MRIs) to detect anomalies such as malignant tumors, signs of pneumonia, or retinal pathologies with accuracy sometimes comparable to, or even exceeding, human experts [Esteva et al., 2017]. In pathology, AI can analyze vast histological slides to identify cancerous cells or classify tumor subtypes, assisting pathologists in making more precise diagnoses. Similarly, algorithms trained on structured and unstructured patient data from electronic health records (EHRs)—including symptoms, lab results, medical histories, and genetic markers—can identify patterns indicative of rare diseases, infectious outbreaks, or conditions like sepsis, often earlier than traditional methods [Poon et al., 2020]. These systems can flag potential diagnoses for clinician review, reducing diagnostic errors and improving the speed of diagnosis.

2.2. Prognostic Evaluation

Beyond diagnosis, clinical algorithms are highly effective in predicting disease progression and patient outcomes, thereby guiding treatment plans and patient management strategies. Predictive analytics models can estimate a patient’s risk of developing complications (e.g., heart attack, stroke, kidney failure), readmission to the hospital, or mortality within a specific timeframe [Johnson et al., 2018]. For instance, algorithms can predict which patients are at high risk for decompensation in intensive care units, allowing for proactive interventions. In oncology, AI-powered tools can predict treatment response and recurrence risk based on tumor characteristics, genetic profiles, and treatment history, helping clinicians select the most effective therapeutic regimen for individual patients. These prognostic insights empower both clinicians and patients to make informed decisions about care pathways, end-of-life planning, and lifestyle modifications.

2.3. Treatment Recommendations

Algorithms play a crucial role in suggesting therapeutic interventions, moving beyond generalized guidelines to patient-specific recommendations. By synthesizing clinical guidelines, evidence-based medicine, and individual patient factors (e.g., comorbidities, drug allergies, genomic markers, lifestyle), these systems can propose optimal drug dosages, surgical approaches, or lifestyle interventions [Beam & Kohane, 2018]. In pharmacogenomics, algorithms can analyze a patient’s genetic profile to predict their response to certain medications, minimizing adverse drug reactions and maximizing efficacy. Similarly, in chronic disease management, AI can monitor patient data, identify deviations from target parameters, and recommend adjustments to medication or lifestyle, often delivered through patient-facing applications or direct alerts to clinicians. The goal is to move towards true personalized medicine, where treatment is precisely tailored to the individual.

2.4. Resource Allocation and Operational Efficiency

Clinical algorithms also extend their utility to operational aspects of healthcare, assisting in prioritizing healthcare resources, managing hospital workflows, and optimizing public health interventions. For instance, algorithms can predict surges in patient demand for emergency services or intensive care unit (ICU) beds, allowing hospitals to allocate staff and resources more effectively. They can optimize surgical scheduling to reduce wait times and improve operating room utilization. In organ transplantation, algorithms like the Kidney Donor Risk Index (KDRI), despite their inherent biases discussed later, have been used to evaluate donor-recipient compatibility and allocate organs, aiming to maximize graft survival. On a larger scale, predictive models assist public health agencies in forecasting disease outbreaks, identifying at-risk populations, and strategically deploying resources for vaccination campaigns or disease surveillance [Braithwaite et al., 2019].

2.5. Underlying Technologies and Data Dependence

The efficacy of these sophisticated algorithms is fundamentally reliant on two critical factors: the quality and representativeness of the data used for their development and the robustness of the underlying AI/ML technologies. Algorithms often leverage supervised learning, where models learn from labeled datasets (e.g., ‘diagnosis A’ for specific symptoms and lab results), or unsupervised learning, which identifies hidden patterns in unlabeled data. Deep learning, a subset of machine learning utilizing neural networks with multiple layers, has been particularly transformative for complex tasks like image and natural language processing. Natural Language Processing (NLP) techniques enable algorithms to extract meaningful information from unstructured clinical notes, significantly expanding the scope of data available for analysis [Shah et al., 2019]. However, this dependence underscores a crucial vulnerability: if the training data is flawed, biased, incomplete, or unrepresentative, the algorithms built upon it will inevitably inherit and amplify these deficiencies, leading to potentially harmful and inequitable outcomes in clinical practice.

3. Sources of Bias in Clinical Algorithms

Bias in clinical algorithms is a multifaceted phenomenon, not merely a technical glitch, but rather a reflection of societal inequities embedded within data, design choices, and implementation contexts. Understanding the origins of these biases is crucial for their effective mitigation. These sources can broadly be categorized from data collection to model deployment.

3.1. Racial and Ethnic Bias

Racial and ethnic biases are among the most pernicious forms of algorithmic bias in healthcare, often stemming from historical medical practices that erroneously linked race to biological differences, despite race being primarily a social construct [Roberts, 2019]. These biases have been explicitly embedded in numerous clinical algorithms, leading to disparate treatment and outcomes.

One of the most widely cited examples is the estimated Glomerular Filtration Rate (eGFR), a critical measure of kidney function. For decades, eGFR calculations included a ‘race modifier,’ specifically an adjustment factor for Black individuals, which assumed higher muscle mass and thus systematically reported a higher eGFR for Black patients compared to non-Black patients with the same creatinine levels [Ahmed et al., 2021]. This upward adjustment had severe clinical consequences: it could delay the diagnosis of chronic kidney disease (CKD) in Black patients, postpone referrals to nephrologists, delay access to life-saving kidney transplants by keeping them off waitlists longer, and hinder eligibility for kidney disease clinical trials. The physiological basis for this race modifier was largely unsubstantiated, rooted in historical misconceptions rather than robust biological evidence. The Lown Institute has highlighted how such algorithms have categorized kidneys from Black donors as ‘more likely to fail,’ contributing to longer wait times for Black patients needing transplants (lowninstitute.org). The widespread recognition of this bias has led major medical societies, like the American Society of Nephrology (ASN) and the National Kidney Foundation (NKF), to recommend the elimination of race from eGFR calculations, a significant step towards health equity [ASN-NKF Task Force, 2021].

Another example is spirometry, a test measuring lung function. Algorithms used to interpret spirometry results often include race-based adjustments, assuming Black and Asian individuals have inherently lower lung capacities than white individuals [Braun, 2014]. This adjustment can lead to underdiagnosis of respiratory conditions like asthma or chronic obstructive pulmonary disease (COPD) in these groups, as their ‘adjusted’ values might still fall within a ‘normal’ range, even if their unadjusted values indicate impairment. Such practices reinforce harmful stereotypes and delay appropriate medical intervention.

Risk prediction scores for various conditions, such as heart failure or sepsis, can also exhibit racial bias. If algorithms are trained on datasets where certain racial groups receive different standards of care or have different disease prevalence due to social determinants of health, the model might learn to associate race with different risk profiles, even if race is a proxy for underlying social or environmental factors rather than a biological determinant [Shah et al., 2020]. This can lead to Black or Hispanic patients being assigned lower risk scores for conditions they are disproportionately affected by, leading to delayed or less aggressive treatment.

3.2. Gender Bias

Gender bias in clinical algorithms can arise from historical underrepresentation of women in clinical trials, differing symptom presentations often dismissed, or societal stereotypes embedded in medical practice. This bias can manifest in various ways, leading to suboptimal care for women.

One notable example is the Vaginal Birth After Cesarean (VBAC) algorithm. This algorithm, used to assess the likelihood of successful VBAC, has been found to contribute to higher rates of cesarean sections among Black and Hispanic women (lowninstitute.org). While individual clinical factors contribute, research suggests that the algorithm might implicitly encode biased assumptions about the likelihood of successful vaginal delivery in these groups, potentially by overweighting certain risk factors or underrepresenting protective factors. This can lead to clinicians recommending repeat C-sections more readily for these women, even when a VBAC might be a safe and preferred option, further contributing to maternal health disparities.

Gender bias is also prevalent in the diagnosis and treatment of cardiovascular diseases. Women often present with atypical symptoms for myocardial infarction (heart attack) compared to men, which historically has led to delayed diagnosis and treatment. Algorithms trained predominantly on male patient data, or on data where female atypical symptoms were misclassified, may perpetuate this diagnostic gap. Similarly, in pain management, women’s pain is often underestimated or dismissed, leading to delays in appropriate analgesic prescription [Hoffman & Tarzian, 2001]. Algorithms that learn from these biased clinical records could reinforce patterns of under-treating women’s pain.

3.3. Socioeconomic Status (SES) Bias

Socioeconomic factors profoundly influence health outcomes and access to care, and when integrated into algorithms, they can create significant biases. Algorithms trained on data from higher-income populations, or that use proxies for SES, may not accurately predict health risks for individuals from lower socioeconomic backgrounds, leading to disparities in care.

Proxies for SES can include insurance status, zip code, educational attainment, or even internet access. An algorithm that uses insurance claims data might identify lower healthcare utilization among certain populations, inferring lower ‘need’ when in reality, it reflects barriers to access due to lack of insurance or financial constraints [Obermeyer et al., 2019]. A landmark study by Obermeyer et al. (2019) demonstrated this vividly: a widely used algorithm designed to predict which patients would benefit from additional care management preferentially assigned lower risk scores to Black patients, even when they were sicker. The algorithm used healthcare costs as a proxy for health needs; since Black patients historically incur lower healthcare costs due to systemic barriers to access, the algorithm erroneously concluded they were healthier, thus perpetuating disparities.

Algorithms designed for resource allocation or risk stratification can also exhibit SES bias. For instance, an algorithm designed to identify patients at high risk of readmission might disproportionately flag individuals from low-income neighborhoods simply because those areas have poorer access to follow-up care, healthy food, or safe housing – all critical social determinants of health (SDOH). While identifying these patients for intervention might seem beneficial, if the intervention only addresses clinical factors and ignores the underlying social determinants, it might be ineffective or even misdirect resources away from other high-need individuals who are clinically similar but live in more affluent areas.

3.4. Data Quality and Representativeness Bias

The fundamental building block of any algorithm is data. If this data is flawed, incomplete, or unrepresentative, the algorithm built upon it will inevitably inherit and amplify these deficiencies [Ghassemi et al., 2021]. Data quality issues can manifest in several ways:

• Sampling Bias: If the dataset used to train the algorithm does not accurately reflect the diversity of the population the algorithm will serve, it will perform poorly on underrepresented groups. For example, if a diagnostic algorithm for a skin condition is trained primarily on images of lighter skin tones, its accuracy will be significantly lower for individuals with darker skin [Adamson & Smith, 2018]. Similarly, if clinical trials historically exclude pregnant women, children, or elderly patients, algorithms trained on data from adult males may not generalize well to these populations.
• Measurement Bias: This occurs when data is collected or recorded inconsistently or inaccurately across different groups. For instance, blood pressure readings taken in a clinical setting might differ from those taken at home, and access to home monitoring devices might be correlated with SES. Furthermore, subjective measurements, such as pain scores or symptom descriptions, can be influenced by implicit biases of healthcare providers recording the data, leading to skewed representations in the dataset.
• Label Bias: This refers to inconsistencies or inaccuracies in the ‘ground truth’ labels used to train supervised learning models. If historical diagnostic labels themselves reflect societal biases (e.g., women historically being diagnosed with ‘hysteria’ for symptoms later recognized as legitimate medical conditions), an algorithm trained on such labels will perpetuate these historical inaccuracies [Buolamwini & Gebru, 2018].
• Missing Data: Incomplete records, especially for marginalized populations who may have less consistent access to care, can lead to algorithms that struggle to make accurate predictions for these groups. Algorithms often impute missing values, but if the patterns of missingness are systematic and correlated with demographic factors, imputation can further entrench bias.
• Data Sparsity for Rare Conditions/Populations: Algorithms often perform best when trained on large volumes of data. For rare diseases, specific genetic mutations, or small demographic subgroups, the available data may be insufficient, leading to poor model performance and diagnostic errors for these unique cases.

3.5. Algorithmic and Developer Bias

Beyond the inherent biases in the data, the choices made during the algorithmic design and development process can also introduce or amplify bias. These include:

• Feature Selection: The choice of which variables (features) to include in a model can be a source of bias. If a developer includes features that are proxies for protected attributes (e.g., zip code as a proxy for race or income) without critical evaluation, the algorithm will learn to associate outcomes with these proxies, even if they are not truly causal [O’Neil, 2016]. Conversely, excluding relevant features that are differentially important for minority groups can also harm performance for those groups.
• Model Architecture and Parameters: The specific algorithms chosen (e.g., linear regression vs. deep neural network), their configurations, and hyperparameter tuning can affect how biases are learned and propagated. Complex models, while powerful, can be opaque, making it difficult to trace how bias is embedded.
• Evaluation Metrics: The choice of performance metrics is crucial. An algorithm might achieve high overall accuracy but perform poorly for a specific subgroup, leading to equitable performance metrics (e.g., equalized odds, demographic parity) are not explicitly optimized for during development [Narayan et al., 2020]. If an algorithm is optimized solely for overall accuracy, it may sacrifice accuracy for minority groups where errors are less frequent but potentially more impactful.
• Developer Implicit Bias: The implicit biases of the development team can subtly influence problem formulation, data annotation, feature engineering, and model validation. A lack of diversity within AI development teams can lead to blind spots, where potential biases affecting certain populations are simply not considered or detected [Gebru et al., 2021].

3.6. Contextual and Implementation Bias

Even a well-designed, ostensibly ‘fair’ algorithm can introduce bias during its deployment and integration into clinical workflows. This includes:

• Over-reliance on Algorithmic Outputs: Clinicians, facing time pressures and information overload, might over-rely on algorithmic recommendations without applying critical clinical judgment, especially if the algorithm is perceived as infallible. This can lead to automation bias, where human decision-makers ignore conflicting information from other sources [Ghassemi et al., 2021].
• Lack of Training and Understanding: Healthcare providers may not fully understand the limitations, assumptions, or specific biases of the algorithms they are using. This lack of awareness can lead to misinterpretation of results or inappropriate application of the tool.
• Feedback Loops: If an algorithm leads to biased decisions that then become part of the training data for future iterations, it can create a ‘feedback loop’ that entrenches and amplifies the original bias over time. For example, if an algorithm recommends less aggressive treatment for a certain group, and those patients consequently have worse outcomes, this might reinforce the algorithm’s initial ‘belief’ that these patients have poorer prognoses, leading to a vicious cycle.
• Alert Fatigue: Clinical decision support systems often generate numerous alerts. If these alerts are poorly calibrated or disproportionately target certain patient groups, they can lead to ‘alert fatigue,’ where clinicians ignore important recommendations, potentially impacting care for vulnerable populations more severely.

4. Implications of Bias in Clinical Algorithms

The presence of unaddressed bias in clinical algorithms carries far-reaching and profoundly detrimental implications that extend beyond individual patient harm, impacting healthcare systems, societal trust, and the very fabric of medical ethics and law.

4.1. Exacerbation of Health Disparities

Perhaps the most direct and alarming consequence of biased algorithms is their capacity to exacerbate existing health disparities. By operationalizing and amplifying historical and systemic inequities, these tools can lead to suboptimal care, delayed diagnoses, and inappropriate treatments for marginalized groups. For instance, the race-adjusted eGFR calculation meant that Black patients often experienced delayed referrals to nephrology specialists, later initiation of dialysis, and extended waiting times for kidney transplantation, contributing to disproportionately worse kidney disease outcomes in this community [Ahmed et al., 2021]. Similarly, gender-biased algorithms in cardiovascular risk assessment might lead to women receiving less aggressive diagnostic workups or preventative care, contributing to higher mortality rates from heart disease in women. When algorithms guide resource allocation, biased risk scores can funnel resources away from communities already facing significant barriers to care, creating a ‘Matthew effect’ where ‘the rich get richer, and the poor get poorer’ in terms of healthcare access and quality. This systematic disadvantage undermines the fundamental goal of healthcare: to provide equitable care to all individuals, irrespective of their background [Obermeyer et al., 2019].

4.2. Erosion of Trust

Trust is the bedrock of the patient-provider relationship and a fundamental component of public health initiatives. When patients discover that clinical tools are implicitly or explicitly biased against them, it can severely erode their confidence in healthcare providers, institutions, and the entire medical system [Chen et al., 2021]. This erosion of trust can manifest in several ways: patients may become hesitant to seek care, adhere to treatment plans, participate in clinical research, or share sensitive health information. For marginalized communities who have historically experienced medical mistreatment, experimentation, and discrimination, biased algorithms represent a new frontier of systemic injustice, deepening existing cynicism and skepticism. This lack of trust can have cascading effects, impacting population health efforts such as vaccination campaigns or chronic disease screening programs, as community buy-in becomes difficult to secure. The perception that technology, often touted as a panacea, is actually perpetuating harm can lead to widespread distrust in technological advancements within healthcare, hindering future innovations that could genuinely improve health outcomes.

4.3. Legal and Ethical Concerns

The deployment of biased algorithms in clinical settings raises complex and challenging legal and ethical dilemmas. From a legal perspective, the use of algorithms that lead to disparate treatment could expose healthcare providers and institutions to discrimination lawsuits under civil rights laws (e.g., Title VI of the Civil Rights Act of 1964 in the U.S.). If a biased algorithm contributes to medical errors or suboptimal care, it could also fall under the purview of medical malpractice claims. Establishing accountability becomes particularly challenging in the context of AI, as it involves multiple stakeholders: data scientists, algorithm developers, healthcare providers, and health systems. The question of ‘who is responsible’ when an algorithm errs is not yet fully settled in jurisprudence [Price, 2019].

Ethically, biased algorithms violate fundamental principles of medical ethics:

• Justice: They undermine the principle of justice, which demands fair and equitable distribution of healthcare resources and equal access to quality care for all. Bias directly contravenes the idea that similar cases should be treated similarly.
• Beneficence and Non-maleficence: Algorithms are intended to ‘do good’ (beneficence) and ‘do no harm’ (non-maleficence). When biased, they actively cause harm or deny potential benefit to specific patient groups.
• Autonomy: While less direct, bias can indirectly affect patient autonomy by limiting their choices or providing them with incomplete or skewed information based on algorithmic recommendations. Patients cannot make truly informed decisions if the clinical information presented to them is based on biased assumptions.

Furthermore, the lack of transparency and explainability in complex ‘black-box’ AI models poses significant ethical challenges. Clinicians may struggle to explain algorithmic recommendations to patients, and patients may not understand why a particular decision was made, further eroding trust and making it difficult to challenge potentially biased outputs. The very definition of ‘fairness’ in AI is a contested ethical and technical challenge, with multiple, sometimes conflicting, mathematical definitions, highlighting the complexity of ensuring ethical deployment [Mehrabi et al., 2021].

5. Best Practices for Developing and Implementing Ethical Algorithms

Mitigating bias in clinical algorithms and promoting health equity requires a comprehensive, multi-faceted approach that spans the entire lifecycle of algorithm development and deployment. It necessitates a shift from purely technical optimization to a socio-technical perspective that prioritizes ethical considerations, transparency, and accountability.

5.1. Diverse and Representative Data Collection

The cornerstone of unbiased algorithms is unbiased data. This requires proactive and intentional efforts to collect diverse and representative datasets. Strategies include:

• Prospective Data Collection: Moving beyond reliance on historical, often biased, retrospective data by designing studies specifically to gather data from underrepresented populations (e.g., racial and ethnic minorities, individuals of different genders, age groups, socioeconomic strata, and geographic locations) [Ghassemi et al., 2021]. This involves dedicated funding and infrastructure.
• Data Augmentation and Balancing: For existing datasets, techniques like oversampling minority classes, synthetic data generation (with careful validation to ensure realism and avoid amplifying existing biases), and transfer learning can help create more balanced training environments. However, these methods should be used cautiously, as they can sometimes introduce new biases or make models overconfident in areas with limited real data.
• Rich Feature Sets: Collecting data on social determinants of health (SDOH)—such as housing stability, food security, education, and access to transportation—can help algorithms move beyond problematic proxies (like race) and identify underlying drivers of health outcomes, allowing for targeted and equitable interventions [Volkova et al., 2020].
• Addressing Missingness: Employing sophisticated missing data imputation techniques that are sensitive to demographic differences, and designing data collection systems that minimize missingness for vulnerable populations.

5.2. Transparent Algorithm Development and Documentation

Transparency is paramount for fostering accountability and trust. This involves clear documentation at every stage of the algorithm’s lifecycle:

• Datasheets for Datasets and Model Cards for Models: Inspired by best practices from the AI ethics community, ‘datasheets for datasets’ provide detailed metadata about the dataset’s provenance, collection methodology, potential biases, and intended use [Gebru et al., 2018]. ‘Model cards’ offer similar transparency for trained models, documenting their performance across different demographic subgroups, limitations, intended use cases, and ethical considerations [Mitchell et al., 2019].
• Explainable AI (XAI) Techniques: Utilizing methods like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) to understand why an algorithm made a particular decision. XAI can help identify features that disproportionately influence outcomes for certain groups or reveal reliance on problematic proxy variables [Rudin, 2019].
• Open-Source and Peer Review: Where appropriate and privacy-compliant, making algorithms and underlying methodologies open-source can facilitate independent scrutiny, identify vulnerabilities, and foster collaborative improvement within the research community.
• Clear Documentation: Comprehensive documentation of data sources, preprocessing steps, model architecture, evaluation metrics, limitations, and decision-making criteria is essential for internal review, regulatory compliance, and external auditing.

5.3. Regular Auditing, Validation, and Monitoring

Bias is not static; it can emerge or change over time. Therefore, continuous vigilance is required:

• Pre-deployment Auditing: Before deployment, algorithms must undergo rigorous testing to assess fairness across various demographic subgroups. This involves using diverse test datasets and employing a range of fairness metrics beyond overall accuracy, such as demographic parity, equalized odds, and predictive equality [Narayanan et2020]. These metrics help identify if an algorithm performs worse for specific groups, has different error rates, or predicts positive outcomes at different rates.
• Post-deployment Monitoring: Once an algorithm is in use, continuous monitoring is crucial to detect ‘model drift’ (where performance degrades over time due to changes in data distribution) and emergent biases. This involves tracking real-world outcomes across different patient groups and setting up alert systems for significant performance discrepancies [Veale & Binns, 2017].
• Adversarial Testing: Stress-testing algorithms by intentionally feeding them data designed to expose weaknesses or biases, similar to how cybersecurity systems are tested.
• Independent Audits: Engaging third-party auditors to conduct impartial assessments of algorithms can provide an additional layer of scrutiny and accountability.

5.4. Multidisciplinary Collaboration and Stakeholder Engagement

Addressing bias effectively requires a broad spectrum of expertise and perspectives:

• Diverse Development Teams: AI development teams should be multidisciplinary, including not only data scientists and clinicians but also ethicists, sociologists, anthropologists, legal experts, and patient advocates. This diversity helps identify potential biases that might be overlooked by a homogeneous team [Gebru et al., 2021].
• Patient and Community Engagement: Involving representatives from affected communities in the design, development, and evaluation phases ensures that algorithms are relevant, respectful, and genuinely beneficial to the populations they serve. This could involve focus groups, community advisory boards, or participatory design workshops.
• Feedback Mechanisms: Establishing clear channels for clinicians and patients to report observed biases or suboptimal performance in real-world settings, ensuring that user feedback directly informs algorithm improvements.

5.5. Patient-Centered Approach and Human Oversight

Algorithms should always serve as tools to augment, not replace, human clinical judgment and empathy:

• Algorithms as Decision Support, Not Decision Makers: Emphasizing that algorithmic recommendations are inputs to, rather than final determinants of, clinical decisions. Clinicians must be trained to critically evaluate algorithmic outputs in the context of individual patient circumstances, preferences, and social determinants of health [Kelly et al., 2019].
• Personalization and Contextualization: Designing algorithms that can incorporate individual patient values, preferences, and unique social contexts, moving beyond a ‘one-size-fits-all’ approach. This includes explicitly addressing social determinants of health within the algorithm’s logic or providing clinicians with tools to factor them in.
• Ethics-by-Design: Integrating ethical considerations from the very inception of algorithm development, rather than as an afterthought. This means anticipating potential harms and designing safeguards proactively.

5.6. Ethical Review Boards and Governance Structures

Formalizing ethical oversight is critical for systemic change:

• AI Ethics Committees: Establishing dedicated AI ethics committees within healthcare institutions, similar to Institutional Review Boards (IRBs), to review the ethical implications of proposed AI projects, data usage, and algorithm deployments [Esmaeilzadeh, 2020].
• Internal Governance Frameworks: Developing clear institutional policies, guidelines, and governance frameworks for the responsible development, procurement, and deployment of AI in healthcare, including explicit mandates for bias assessment and mitigation.
• Professional Guidelines: Encouraging professional medical organizations to develop and enforce ethical guidelines for the use of AI in their respective specialties, setting standards for responsible practice.

6. Case Studies

Examining real-world initiatives provides concrete evidence of both the challenges and the progress in addressing algorithmic bias in healthcare.

6.1. Veterans Health Administration (VHA) and eGFR

The Veterans Health Administration (VHA), one of the largest integrated healthcare systems in the United States, has been a leading institution in recognizing and actively working to eliminate race-based adjustments in clinical algorithms. Their efforts regarding the estimated Glomerular Filtration Rate (eGFR) calculation serve as a significant case study. For many years, the VHA, like most healthcare systems globally, utilized eGFR equations that included a race modifier for Black individuals. As detailed earlier, this modifier systematically inflated eGFR values for Black patients, leading to delayed diagnoses of chronic kidney disease (CKD), postponed referrals to nephrologists, and delayed access to kidney transplant waitlists [Ahmed et al., 2021].

Recognizing the ethical and clinical harms of this practice, the VHA undertook a concerted effort to remove the race modifier from eGFR calculations across its entire system. This initiative was informed by a growing body of evidence highlighting the lack of biological basis for racial differences in creatinine kinetics and the detrimental impact of the modifier on health equity. The VHA’s decision, often predating broader national recommendations, involved a multi-pronged approach: recalibrating laboratory systems, updating electronic health record (EHR) systems, and providing extensive education to clinicians about the change and its implications for patient care [PubMed.ncbi.nlm.nih.gov/38076213/]. The impact has been profound, enabling more accurate assessments of kidney function for all veterans, particularly Black veterans, and facilitating timely access to critical care and transplantation services. This systemic change within a large, integrated health system demonstrates the feasibility and ethical imperative of eliminating race-based adjustments when they lack scientific justification and contribute to health disparities.

6.2. Philadelphia Health Systems Coalition

A notable collective action against algorithmic bias emerged from a coalition of 13 prominent health systems in the greater Philadelphia area. In a landmark decision, these institutions announced their commitment to discontinue the use of race-based adjustments across a spectrum of clinical algorithms, specifically targeting those in lung, kidney, and obstetric care [Axios.com/2024/10/23/philadelphia-hospitals-drop-race-algorithms]. This collaborative effort represents a significant regional commitment to addressing systemic biases embedded in clinical tools, especially as AI models become increasingly integral to diagnosing and treating patients.

The coalition’s initiative specifically aimed to remove race modifiers from the eGFR calculation (mirroring the VHA’s efforts), spirometry interpretation (which historically assumed lower lung capacities for Black and Asian individuals), and various algorithms used in obstetric care, such as those predicting the likelihood of successful Vaginal Birth After Cesarean (VBAC). The rationale behind this concerted action was rooted in a shared understanding that using race as a biological variable in clinical decision-making often conflates social constructs with physiological realities, leading to inaccurate assessments and perpetuating health inequities. By collectively removing these adjustments, the Philadelphia health systems sought to standardize care based on physiological markers rather than racial categories, aiming to provide more equitable and evidence-based care for all patients in the region. This case exemplifies the power of institutional collaboration and the growing awareness within the medical community of the ethical imperative to scrutinize and reform biased algorithmic practices.

6.3. The Optum Algorithm Controversy

One of the most widely cited and impactful case studies illustrating algorithmic bias is the 2019 Science paper by Obermeyer, Powers, Vogeli, and Mullainathan, which investigated a commercially available algorithm used by a major U.S. healthcare provider (Optum, a subsidiary of UnitedHealth Group) to manage care for millions of patients [Obermeyer et al., 2019]. The algorithm’s purpose was to identify patients with complex health needs who would benefit from specialized care management programs.

The researchers discovered that the algorithm systematically assigned lower risk scores to Black patients compared to white patients, even when the Black patients were demonstrably sicker and had a greater need for additional medical attention. The core of the bias stemmed from the algorithm’s design: it used healthcare costs as a proxy for health needs. Since Black patients in the U.S. historically incur lower healthcare costs due to systemic barriers to access (e.g., lack of insurance, geographic barriers, distrust in the medical system), the algorithm erroneously inferred that they were healthier, thus under-prioritizing them for critical care management programs. The study estimated that if the algorithm had been race-neutral, it would have more than doubled the proportion of Black patients identified for extra care.

This case starkly highlighted several crucial points:
* Proxy Variables: How seemingly neutral variables (like cost) can serve as proxies for protected attributes (like race) and embed deep-seated societal biases into algorithms.
* Unintended Consequences: The algorithm’s developers likely had benevolent intentions, but the unexamined choice of a proxy metric led to severe discriminatory outcomes.
* The Need for Fairness Auditing: The bias was only discovered through rigorous, external auditing, underscoring the necessity of evaluating algorithms for disparate impact across demographic groups.
* Impact on Resource Allocation: The algorithm directly influenced the allocation of valuable healthcare resources, demonstrating how algorithmic bias can perpetuate structural inequities in access to care.

6.4. Maternal Mortality Risk Algorithms

Maternal mortality and morbidity rates in the United States are alarmingly high, particularly for Black and Indigenous women who face significantly greater risks compared to white women [Centers for Disease Control and Prevention, 2023]. In an effort to address these disparities and improve maternal outcomes, some healthcare systems have begun to develop and implement algorithms to predict a patient’s risk of adverse maternal events, such as severe maternal morbidity or mortality.

While well-intentioned, these algorithms face a substantial risk of encoding and amplifying existing racial biases within maternal healthcare. If such algorithms are trained on historical data where Black women’s pain was often dismissed, their symptoms were not adequately addressed, or they faced structural barriers to quality prenatal and postnatal care, the algorithm might learn to associate race with a higher baseline risk, or conversely, might miss key indicators if those indicators were historically under-documented for Black patients. For example, if an algorithm relies heavily on specific lab values or adherence to follow-up appointments, and Black women disproportionately face challenges in accessing those services due to socioeconomic factors or discrimination, the algorithm might misclassify their risk [Vedana et al., 2021].

Furthermore, if the algorithms are not carefully designed to account for social determinants of health and structural racism in their inputs and outputs, they could reinforce a narrative that attributes higher risk to race itself rather than to the systemic factors that contribute to racial disparities. This could lead to a ‘blame the victim’ approach or to differential treatment recommendations based on race, rather than focusing on individualized needs and addressing systemic issues. The development of such algorithms requires extraordinary care to ensure they become tools for equity rather than new mechanisms of discrimination in a highly sensitive area of public health.

7. Policy and Regulatory Considerations

The pervasive impact of bias in clinical algorithms necessitates robust policy and regulatory frameworks to ensure their ethical, equitable, and safe deployment. Governmental bodies, professional organizations, and international coalitions are increasingly recognizing this imperative and developing guidance.

7.1. Guiding Principles and Ethical Frameworks

Several organizations have begun to articulate guiding principles for the responsible development and deployment of AI in healthcare:

• Agency for Healthcare Research and Quality (AHRQ): AHRQ has developed guiding principles specifically aimed at helping the healthcare community address potential bias resulting from algorithms (AHRQ.gov). These principles typically emphasize the importance of transparency, accountability, fairness, safety, and equity in algorithm development and implementation. They underscore the need for continuous vigilance and proactive measures to prevent harm to vulnerable populations.
• World Health Organization (WHO): The WHO has published guiding principles for the ethics and governance of AI for health, focusing on six core principles: protecting human autonomy, promoting human well-being and safety, ensuring transparency and explainability, fostering responsibility and accountability, ensuring inclusiveness and equity, and promoting AI that is responsive and sustainable [WHO, 2021]. These international guidelines provide a broad framework for national policies.
• National Academy of Medicine (NAM): The NAM has convened initiatives to address health equity and AI, emphasizing the need for robust ethical frameworks that are integrated into the lifecycle of AI tools in medicine, advocating for principles like ‘auditable algorithms’ and ‘human in the loop’ approaches [Dzau et al., 2022].

These principles serve as foundational guidelines, but their translation into actionable policies and enforceable regulations remains a critical challenge.

7.2. Regulatory Oversight and Accountability Frameworks

Establishing clear regulatory oversight for AI-driven clinical algorithms is crucial, especially given their potential to impact patient safety and health equity. Key considerations include:

• FDA Regulation of AI/ML-based Software as a Medical Device (SaMD): In the United States, the Food and Drug Administration (FDA) is actively developing a regulatory framework for AI/ML-based Software as a Medical Device (SaMD). This includes premarket review pathways, but also considerations for ‘adaptive’ algorithms that can continuously learn and update [FDA, 2021]. The FDA emphasizes the need for ‘good machine learning practice’ (GMLP) principles, which include data quality, appropriate algorithm design, and measures to manage bias and ensure transparency.
• Accountability for Harm: A significant regulatory challenge is establishing clear accountability when an algorithm contributes to patient harm or discriminatory outcomes. Current legal frameworks are often designed for human error or device malfunction, not for complex algorithmic systems where responsibility can be distributed across developers, data providers, and deployers. New legal doctrines or regulatory bodies might be needed to address this ‘algorithmic accountability gap’ [Price, 2019].
• Algorithmic Impact Assessments (AIAs): Similar to Environmental Impact Assessments, AIAs could be mandated for high-risk clinical algorithms. These assessments would require developers and deployers to proactively evaluate the potential societal and equity impacts of their algorithms before widespread adoption, identify potential biases, and outline mitigation strategies [Crawford, 2017].
• Mandatory Auditing and Reporting: Regulations could require regular, independent auditing of clinical algorithms for bias, coupled with mandatory reporting of performance across demographic subgroups. This would ensure ongoing vigilance and provide data for public scrutiny and policy adjustments.

7.3. Ethical AI Legislation and Data Governance

Legislative efforts are emerging globally to address the broader implications of AI, including in healthcare:

• EU AI Act: The European Union’s proposed AI Act categorizes AI systems by risk level, with ‘high-risk’ systems (which would include many clinical algorithms) facing stringent requirements for data governance, documentation, transparency, human oversight, robustness, accuracy, and cybersecurity. It explicitly calls for systems to be developed to ‘minimize the risk of bias’ [European Commission, 2021]. This comprehensive legislative approach could set a global standard.
• State-Level Initiatives: Some U.S. states are exploring their own legislation. For example, laws targeting the use of AI in hiring or lending could inspire similar approaches for healthcare, requiring transparency and non-discrimination.
• Data Governance and Privacy: Strong data governance frameworks are essential. Regulations like HIPAA (Health Insurance Portability and Accountability Act) in the U.S. and GDPR (General Data Protection Regulation) in Europe are critical for protecting patient privacy. However, new considerations arise when data is used for AI training, necessitating ethical guidelines for data sharing, de-identification, and the potential for re-identification through advanced AI techniques [Vayena & Blasimme, 2018]. Policies must balance privacy protections with the need for diverse and representative datasets to train equitable algorithms.

Ultimately, a combination of voluntary ethical guidelines, robust regulatory oversight, and targeted legislation will be necessary to ensure that clinical algorithms are developed and implemented in a manner that truly enhances healthcare for all, without perpetuating or deepening existing inequalities. This requires ongoing dialogue between policymakers, regulators, industry, academia, and civil society.

8. Conclusion

Clinical algorithms represent a powerful frontier in healthcare innovation, holding immense promise to enhance diagnostic precision, optimize treatment strategies, and streamline patient management. The ability of these artificial intelligence and machine learning-powered tools to process vast datasets and discern subtle patterns can undoubtedly revolutionize medical practice, fostering a future of more personalized, efficient, and potentially life-saving care. However, the enthusiasm for technological advancement must be tempered with a critical awareness of the inherent risks, particularly the potential for these algorithms to perpetuate and even exacerbate existing health disparities.

As this report has meticulously detailed, bias in clinical algorithms is not merely a technical flaw but a pervasive reflection of historical injustices, societal inequities, and methodological shortcomings embedded within the data, design, and deployment phases. The historical use of race-based adjustments in critical calculations like eGFR and spirometry, the subtle yet detrimental gender biases in risk prediction models, the socio-economic influences captured by problematic proxy variables, and the fundamental challenges of data quality and representativeness all underscore the urgent need for systemic reform. These biases have tangible and severe consequences: delayed diagnoses, inappropriate treatments, limited access to care, erosion of patient trust, and profound legal and ethical dilemmas that undermine the very principles of justice and equity in healthcare.

Moving forward, the healthcare industry, in collaboration with technology developers, policymakers, and affected communities, must commit to a holistic and proactive approach to mitigate algorithmic bias. This commitment demands:

• Intentional Data Curation: A relentless pursuit of diverse, representative, and high-quality data, actively collected from all demographic groups, and enriched with social determinants of health to build a more accurate and equitable foundation for algorithms.
• Transparent and Accountable Development: Embracing methodologies that prioritize transparency, including comprehensive documentation through ‘datasheets for datasets’ and ‘model cards for models,’ alongside the adoption of explainable AI techniques. This fosters accountability and allows for rigorous scrutiny.
• Continuous Monitoring and Auditing: Implementing robust pre-deployment validation and ongoing post-deployment monitoring frameworks, utilizing fairness metrics to detect and correct biases as they emerge or shift over time. Independent audits are crucial for impartial assessment.
• Multidisciplinary Collaboration and Human Oversight: Ensuring that algorithm design and implementation are guided by diverse teams, including clinicians, data scientists, ethicists, legal experts, and community representatives. Crucially, algorithms must remain decision-support tools, augmenting human judgment rather than replacing it, with clinicians trained to critically evaluate and contextualize their outputs.
• Proactive Policy and Regulatory Engagement: Developing and enforcing robust ethical guidelines, regulatory frameworks (such as those from the FDA and emerging global AI acts), and accountability mechanisms to govern the development, deployment, and remediation of clinical algorithms.

The journey towards truly ethical and equitable clinical algorithms is an ongoing endeavor, requiring continuous vigilance, adaptive strategies, and an unwavering commitment to health equity. By consciously designing, validating, and deploying these powerful tools with justice at their core, healthcare systems can harness their transformative potential to genuinely improve outcomes for all patients, bridging rather than widening the chasms of health disparities. The imperative is clear: the future of AI in healthcare must be a future that serves everyone, fairly and justly.

References

• Adamson, A. S., & Smith, W. A. (2018). Dermatology and Skin of Color: An Update for the Primary Care Provider. Journal of the American Board of Family Medicine, 31(2), 346–352.
• Ahmed, S., et al. (2021). The elimination of race from eGFR: Time for action. Journal of the American Society of Nephrology, 32(5), 1033-1036.
• Agency for Healthcare Research and Quality (AHRQ). (n.d.). Guiding Principles to Help the Healthcare Community Address Potential Bias Resulting from Algorithms. Retrieved from https://www.ahrq.gov/news/newsroom/press-releases/guiding-principles.html
• ASN-NKF Task Force. (2021). Recommendation for a New eGFR Equation without a Race Variable. Retrieved from https://www.asn-online.org/policy/documents/ASN-NKF-Task-Force-Final-Report.pdf
• Axios. (2024, October 23). Philadelphia hospitals drop race algorithms. Retrieved from https://www.axios.com/2024/10/23/philadelphia-hospitals-drop-race-algorithms
• Beam, A. L., & Kohane, I. S. (2018). Big Data and Machine Learning in Health Care. JAMA, 319(13), 1317–1318.
• Braithwaite, R. S., et al. (2019). Artificial intelligence for the prevention, surveillance, and control of infectious diseases. Current Opinion in HIV and AIDS, 14(6), 461–467.
• Braun, L. (2014). Breathing Race into the Machine: The Global Extinction of an African American Lung. University of Minnesota Press.
• Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Phenotypic Demographics for Face Dataset. Proceedings of the 1st Conference on Fairness, Accountability, and Transparency, 77–91.
• Centers for Disease Control and Prevention (CDC). (2023). Maternal Mortality Rates in the United States, 2021. NCHS Data Brief, No. 468.
• Char, D. S., et al. (2018). Toward an ethical framework for artificial intelligence in psychiatry. BMC Medicine, 16(1), 1–9.
• Chen, I., et al. (2021). When algorithms learn to discriminate: a primer on the ethical and societal risks of clinical artificial intelligence. Journal of Medical Internet Research, 23(1), e21612.
• Crawford, K. (2017). The Trouble with Bias. NIPS, 1-7.
• Dzau, V. J., et al. (2022). The National Academies of Sciences, Engineering, and Medicine’s Artificial Intelligence in Medicine Initiative. NAM Perspectives.
• Emanuel, E. J., & Wachter, R. M. (2019). Artificial Intelligence in Health Care: The Hope, The Hype, The Promise, The Peril. Annals of Internal Medicine, 171(7), S1–S3.
• Esmaeilzadeh, M. (2020). AI in health care: a review of the opportunities and challenges. Health Management, 20(3), 131-137.
• Esteva, A., et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115–118.
• European Commission. (2021). Proposal for a Regulation on a European approach for Artificial Intelligence. Retrieved from https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206
• FDA. (2021). Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. Retrieved from https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-machine-learning-aiml-based-software-medical-device-samd-action-plan
• Gebru, T., et al. (2018). Datasheets for Datasets. arXiv preprint arXiv:1803.09010.
• Gebru, T., et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 61-68.
• Ghassemi, M., et al. (2021). The ethical algorithm: A framework for responsible AI in health care. Artificial Intelligence in Medicine, 116, 102070.
• Hoffman, D. E., & Tarzian, A. J. (2001). The Girl Who Cried Pain: A Bias against Women in the Treatment of Pain. Journal of Law, Medicine & Ethics, 29(1), 13-27.
• Johnson, A. E. W., et al. (2018). MIMIC-III, a freely accessible critical care database. Scientific Data, 3(1), 160035.
• Kelly, C. J., et al. (2019). Key challenges for the practical application of artificial intelligence in healthcare. Journal of Medical Internet Research, 21(1), e10480.
• Lown Institute. (n.d.). Bias in clinical algorithms make health disparities worse. Retrieved from https://lowninstitute.org/bias-in-clinical-algorithms-make-health-disparities-worse/
• Mehrabi, N., et al. (2021). A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys (CSUR), 54(3), 1-35.
• Mitchell, M., et al. (2019). Model Cards for Model Reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency, 220–229.
• Narayanan, A., et al. (2020). What are the social and ethical risks of AI in clinical healthcare? A systematic review of literature. arXiv preprint arXiv:2009.09176.
• Obermeyer, Z., et al. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6468), 447-453.
• O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown.
• OJS.stanford.edu. (n.d.). Data Quality Bias. Retrieved from https://ojs.stanford.edu/ojs/index.php/sjph/article/view/2460
• Poon, D. C., et al. (2020). Artificial intelligence and machine learning in ophthalmology: a systematic review. Ophthalmology, 127(1), 122–134.
• Price, W. N. (2019). Artificial Intelligence in Health Care: Applications, Risks, and Ethical Considerations. JAMA, 321(23), 2320–2321.
• PubMed.ncbi.nlm.nih.gov. (n.d.). Veterans Health Administration efforts to eliminate race-based adjustments in clinical algorithms. Retrieved from https://pubmed.ncbi.nlm.nih.gov/38076213/
• Rajkomar, A., et al. (2018). Scalable and accurate deep learning with electronic health records. npj Digital Medicine, 1(1), 1-10.
• Roberts, D. (2019). Fatal invention: How science, politics, and big business re-create race in the twenty-first century. New Press.
• Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206-215.
• Shah, N. H., et al. (2019). Artificial Intelligence in Medicine. New England Journal of Medicine, 381(13), 1292–1293.
• Shah, A., et al. (2020). Addressing Bias in Clinical Artificial Intelligence. NEJM Catalyst Innovations in Care Delivery, 1(1).
• Topol, E. J. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.
• Vayena, E., & Blasimme, A. (2018). Health Data and AI: The Need for an Extended Ethical and Regulatory Framework. Advanced Science, 5(5), 1700683.
• Vedana, L. A., et al. (2021). Racial and Ethnic Disparities in Maternal Morbidity and Mortality in the United States: A Systematic Review. Journal of Racial and Ethnic Health Disparities, 8(2), 297–314.
• Veale, M., & Binns, R. (2017). Fairer machine learning in the real world: Mitigating discrimination without collecting sensitive data. arXiv preprint arXiv:1710.03023.
• Volkova, N., et al. (2020). Social determinants of health in electronic health records: A systematic review. Journal of the American Medical Informatics Association, 27(11), 1782-1793.
• World Health Organization (WHO). (2021). Ethics and governance of artificial intelligence for health: WHO guidance. Retrieved from https://www.who.int/publications/i/item/9789240029200