What Is Fair? Defining Fairness in Machine Learning for Health

ABSTRACT

Ensuring that machine‐learning (ML) models are safe, effective, and equitable across all patients is critical for clinical decision‐making and for preventing the amplification of existing health disparities. In this work, we examine how fairness is conceptualized in ML for health, including why ML models may lead to unfair decisions and how fairness has been measured in diverse real‐world applications. We review commonly used fairness notions within group, individual, and causal‐based frameworks. We also discuss the outlook for future research and highlight opportunities and challenges in operationalizing fairness in health‐focused applications.

1. Introduction

There are myriad potential applications of machine learning (ML) for health, ¹ including automated disease detection, computer‐aided diagnosis, and personalized treatment planning [3]. However, there is substantial evidence that, without appropriate forethought and planning, ML models can introduce or exacerbate health inequities by making less accurate decisions for certain groups or individuals [4]. Within medical imaging, state‐of‐the‐art ML models used for disease diagnosis, risk prediction, and triage management are known to underperform within minority groups defined by protected attributes, including sex, race, and ethnicity [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]. For example, deep‐learning models used to detect 14 common diseases from chest X‐rays were found to under‐diagnose under‐served subgroups situated at intersections of oppression, such as Hispanic female patients, potentially resulting in treatment delays if deployed in practice [11]. Similarly, ML models trained on electronic health records (EHRs), administrative claims, and genomic data have shown biased performance, often making less accurate predictions for certain subgroups [16, 17]. A landmark study revealed that a widely used commercial risk prediction tool for identifying patients with complex health needs exhibited significantly lower accuracy for Black patients compared to white patients. As a result, Black patients with similar levels of illness were less likely to be recommended for essential care services [18]. In genomics, polygenic risk scores frequently perform less accurately for individuals of non‐European ancestry due to their historical underrepresentation in genomic data sets, which can contribute to unequal access to preventive care [19, 20, 21].

As applications of ML in health become commonplace, it is crucial to recognize, account for, and mitigate such disparities in model performance to support health equity. Broadly, an ML model is said to be fair if it does not discriminate against an individual or group [22]. Concepts of fairness have been extensively studied across various disciplines, including social choice theory, game theory, economics, philosophy, and law [23, 24, 25, 26, 27, 28, 29]. Building on these principles, the subfield of fairness within ML provides a framework for evaluating and mitigating bias throughout the model development process.

Although fairness has been an extremely active area of research over the past decade, implementing fairness within ML for health is relatively nascent [30, 31, 32]. A systematic review of articles utilizing ML for EHR‐based phenotyping found that only 5% of studies assessed fairness [33]. Similarly, a review of EHR‐based prediction models found that most studies investigate the overall performance of ML models, but do not interrogate potential biases [34]. Beyond EHR applications, several scoping reviews of clinical ML models developed with diverse data sources found that the adoption of fairness remains inconsistent, partly due to a knowledge gap between ML and clinical researchers [32, 35, 36, 37, 38]. Moreover, over two dozen definitions of fairness have been proposed, most of which originate outside health‐focused literature, making it a particularly challenging domain to navigate.

To review this fundamental topic, we examine key notions of fairness and their use in ML for health. ² We first introduce biases that can emerge throughout the model development process using examples from the literature to illustrate why ML models can be unfair. We then review what it means for a model to be fair, beginning with an overview of the most common fairness framework, group fairness, and moving to the emerging frameworks of individual and causal fairness [39]. Our discussion includes the mathematical formulation of various fairness criteria as well as numerous real‐world examples. We conclude by highlighting the limitations of current approaches and outlining opportunities for future research. Our work highlights that, in light of the large number of fairness notions, operationalizing fairness in health‐focused applications remains an open problem. We therefore aim to provide a concise overview and organization of fairness definitions used within ML in health that can be used as a foundational resource for researchers navigating and contributing to the rich and nuanced fairness literature (see Supporting Information Table S1 for a detailed comparison with existing references).

2. Why ML Models Are Unfair

Even with a well‐defined and well‐intentioned research question, ML models can be unfair due to biases in the data, in the model, and/or in the deployment of the model [30, 40, 41, 42, 43, 44, 45, 46]. Figure 1 provides a visualization of sources of bias that can arise throughout model development and Table 1 details common biases using examples from the literature.

FIGURE 1.

Sources of bias. Bias can arise at each stage of the model development process, from data collection to deployment. Each stage influences the next, with bias being potentially perpetuated and compounded throughout the development process. After deployment, an unfair model can also introduce or reinforce societal bias. This figure is adapted from existing literature [47, 48].

TABLE 1.

Common biases that arise in machine‐learning (ML) applications in health and examples from the literature.

Type of bias	Definition	Example
(a) Bias in the data
Societal (or historical) bias	The data reflects long‐standing societal disparities encoded within the data over time.	Clinical word embeddings trained on large corpora of text, such as clinical notes from healthcare systems, reflect biases about ethnic minorities [71].
Selection bias	The data is not representative of the population of interest.	Data collected from wearable devices does not reflect the general population, as usage rates are higher among younger individuals and those with higher socioeconomic status [72].
Measurement Bias	The data contains variables that are collected or measured inaccurately.	When documented in EHRs, gender identity is often recorded without adequate provider training, which can lead to misrepresentation of an individual's identity [73].
Temporal bias	The data captures a specific time period that may not reflect current or future conditions.	Administrative health data showed that pediatric mental health visits were lower than expected among individuals with lower socioeconomic status during the first year of the COVID‐19 pandemic [74].
Minority bias	The data lacks adequate representation from the minority group for the model to accurately learn about them.	Most of the data for genetic studies is from European ancestry populations [20].
Missing data bias	The data has variables that are incompletely measured.	Patients from low‐income backgrounds have higher rates of missing medical measurements in their medical records, which can be partially attributed to inequities in access to health care [75].
(b) Bias in the model
Label bias	An imperfect proxy is selected to train a model instead of the outcome of interest.	Healthcare cost was used as a proxy for healthcare need in a commercial algorithm used to identify patients for high‐risk care management programs. The algorithm significantly under‐identified Black patients for care services as less money is spent on Black patients relative to similarly healthy white patients [18].
Algorithmic bias	Properties of a model and/or its training algorithm create or amplify bias in the data.	In clinical prediction tasks, differentially private models ensure privacy through the addition of calibrated noise. The noise can reduce the model's ability to learn from the tails of the data distribution, leading to accuracy loss that disproportionately affects minority groups [76].
Evaluation bias	An inappropriate choice of a benchmark data set or performance metric(s) is used for evaluation.	Existing public skin disease AI benchmarks do not have images of biopsy‐proven malignancy, the gold standard for disease annotation, on dark skin [77, 78].
(c) Bias in deployment
Automation bias	Model users overly trust model outputs, sometimes even against their own knowledge.	A computer‐aided diagnosis system that scans mammograms and marks suspicious areas of potential cancer features had lower sensitivity for women aged 40‐49 compared to older age groups. A clinician using this system can disproportionately miss cancers for this age group [79].
Dismissal bias	Model users ignore model recommendations, often due to frequent false alerts.	If the same computer‐aided diagnosis system also has a lower positive predictive value for women aged 40‐49, then clinicians may disregard its recommendations, as they are more likely to be false positives [80].
Privilege bias	Models are not available or cannot be deployed in all settings, such as parts of the world where medical technology is unavailable.	Less than 20% of rural emergency departments in Canada have in‐house access to computed tomography (CT) scanners required for ML‐assisted diagnostic imaging [58].

Briefly expanding on Table 1, bias in the data arises from (i) societal and/or (ii) statistical bias [47]. With respect to the former, the data used in health applications most often measures and categorizes people and therefore encodes societal structures, injustices, and stereotypes, such as gender, racial, and age bias [49, 50]. For example, underlying social inequalities in healthcare access can limit the amount of data available in EHRs of certain subgroups, such as Black patients [51]. In terms of statistical bias, the data can also fail to represent the population of interest due to the sampling method (selection bias), time of collection (temporal bias), or data quality issues such as missing, mismeasured, or insufficient data (missing data, measurement, and minority bias). In the fairness literature, data with any of these undesirable properties are informally referred to as biased [22, 52]. Observational clinical data (e.g., claims data, medical images) inherently contains various biases and researchers must carefully consider sources of bias in any application [53, 54]. ³

Additionally, choices made during model training and evaluation can further amplify biases in the data or incorporate new bias, including selecting an inappropriate label (label bias), model (algorithmic bias), and/or evaluation metric or data set (evaluation bias). Table 1 details several health examples subject to model bias. Lastly, during deployment, bias in the data and/or model can be reinforced or introduced when users selectively disregard (dismissal bias) or overly trust (automation bias) a model's outputs, or if models are inaccessible to certain subpopulations (privilege bias). For instance, ML‐assisted diagnostic imaging requires that hospitals have the necessary equipment, yet fewer than 20% of rural emergency departments in Canada have in‐house computed tomography (CT) scanners to utilize these advances [58]. A review of biases outside of health‐focused applications can be found in several recent works [41, 50].

Dr. Leo Celi, a health AI expert, emphasizes that “data bias is the waterloo of health AI.” Bias must be a primary consideration throughout the ML pipeline‐from problem formulation to model deployment [59]. Building on existing AI reporting guidelines [60, 61], a bias evaluation checklist was introduced to enable practitioners to systematically and holistically address bias in clinical predictive models [30]. Bias mitigation strategies and common pitfalls in applying ML in health care are further discussed in the clinical literature [62, 63]. The fairness criteria we introduce in the subsequent sections play an important role in this process. The criteria are typically integrated directly into model training or used as evaluative metrics to identify disparities after the model has been trained [48, 64, 65, 66, 67, 68, 69]. While this paper focuses on how to define fairness, strategies for mitigating bias are an equally important area of research and have been reviewed previously [70].

3. What It Means for an ML Model to Be Fair

Existing definitions of fairness primarily fall into three categories: group fairness, individual fairness, and causal fairness. Group fairness criteria are commonly used in health and deem a model as fair if its predictions are similarly accurate or calibrated across a predefined set of groups. These groups are most often defined by a protected attribute, ⁴ such as age or race. Other attributes used in the health‐focused literature include disability, marital status, national origin, sex, and socioeconomic status. We summarize these attributes in Table S2, along with representative examples from the literature. In contrast, individual fairness is a less frequently used framework and requires that the model provide similar predictions to similar individuals based on user‐defined similarity metrics [29, 38]. Lastly, causal fairness criteria utilize causal estimands to quantify unfairness and link disparities in model performance to their underlying cause [37, 83]. Causality‐based fairness notions are particularly attractive for health‐focused applications as they enable practitioners to interrogate biases [84].

Group fairness criteria are referred to as “oblivious” as they equate fairness with parities in model performance across groups based solely on the distribution of the data (i.e., the predictions, outcome, and protected attribute), while individual and causal fairness criteria are “non‐oblivious” as they require additional context in the form of user‐defined similarity metrics and causal models, respectively [47]. Importantly, many fairness criteria are incompatible, in that they cannot be simultaneously satisfied. A taxonomy of fairness, including different notions of group, individual, and causal fairness and their incompatibilities which are introduced in subsequent sections, is presented in Figure 2.

FIGURE 2.

Taxonomy of fairness. Fairness criteria primarily fall into three categories: group, individual, and causal fairness. Group fairness criteria are oblivious, in the sense that they can be entirely inferred from the distribution of the data. Individual and causal fairness criteria are non‐oblivious in that they require specification of similarity metrics or a causal model, respectively. Many fairness criteria cannot be simultaneously satisfied and incompatibilities between different notions of fairness are depicted with a red X. This taxonomy is not exhaustive and includes the notions of fairness and incompatibilities introduced in this review.

3.1. Notation

Throughout, we denote the outcome of interest as $Y$, the features used for model training as $\mathit{X}$, ⁵ and the variable for the protected attribute as $A$. $\mathit{X}$ may or may not contain $A$ and we discuss this issue when we introduce individual fairness. We let $S=\widehat{f}(\mathit{X})$ be the output from the ML model, where $\widehat{f}$ is learned from a set of training data. For example, in classification, $Y$ is a binary label for membership in the positive or negative class and $S$ is the predicted probability of being in the positive class, referred to as the score. In this work, we primarily focus on binary classification settings, where the final decision $D$ is classified as positive if the score $S$ exceeds a predetermined threshold and negative otherwise. Binary classification tasks are the most widely used and well‐studied in the fairness literature [41, 47, 85].

3.2. Group Fairness

Group fairness criteria require ML models to perform similarly across groups defined by $A$ and are the most popular fairness framework in health‐focused applications [39]. The criteria primarily fall into three categories: independence, separation, and sufficiency [22, 86]. In the subsequent sections, we provide descriptions of common metrics as well as examples of their usage within health‐focused applications. Table 2 summarizes mathematical definitions together with an interpretation in the context of a real‐world example. We provide a brief discussion of approaches for continuous outcomes, continuous protected attributes, categorical protected attributes that define more than two groups, and multiple protected attributes (i.e., subgroup fairness) in Supporting Information Section S2.

TABLE 2.

Common group fairness criteria.

Metric	Definition	Interpretation	Relaxation
Independence‐based criteria
Statistical Parity	$P(D=1|A=a_0)=P(D=1|A=a_1)\forall a_0,a_1\in 𝒜$	In predicting hospitalization and emergency department (ED) visits in heart failure, the probability of a predicted hospitalization or an ED visit is the same for males and females [96].
Conditional Statistical Parity	$P(D=1|\mathit{Z}=\mathit{z},A=a_0)=P(D=1|\mathit{Z}\mathbf{=}z,A=a_1)\forall a_0,a_1\in 𝒜$	In predicting hospitalization and ED visits in heart failure, the probability of a predicted hospitalization or an ED visit is the same for males and females after adjusting for age and pre‐existing conditions [96].	$✓$
Separation‐based criteria
Equalized Odds	$P(D=1|Y=0,A=a_0)=P(D=1|Y=0,A=a_1)\text{and}P(D=0|Y=1,A=a_0)=P(D=0|Y=1,A=a_1)\forall a_0,a_1\in 𝒜$	In predicting asthma exacerbation in children, the rates of false positives (children without exacerbation incorrectly identified as having exacerbation) and false negatives (children with exacerbation incorrectly identified as not having exacerbation) are the same for children from low and high socio‐economic classes [97].
Predictive Equality	$P(D=1|Y=0,A=a_0)=P(D=1|Y=0,A=a_1)\forall a_0,a_1\in 𝒜$	In a model predicting cardiovascular disease (CVD), the rate of false positives (people without CVD incorrectly identified as having CVD) is the same for males and females [98].	$✓$
Equal Opportunity	$P(D=0|Y=1,A=a_0)=P(D=0|Y=1,A=a_1)\forall a_0,a_1\in 𝒜$	In a model predicting suspicious findings from chest X‐rays, the rate of false negatives (patients with true findings incorrectly classified as normal) is the same for males and females [99].	$✓$
Balance for Positive Class	$𝔼(S|Y=1,A=a_0)=𝔼(S|Y=1,A=a_1)\forall a_0,a_1\in 𝒜$	In a model used to screen for lung cancer from chest X‐rays, the average score of people with lung cancer is equal for males and females [99].	$✓$
Balance for Negative Class	$𝔼(S|Y=0,A=a_0)=𝔼(S|Y=0,A=a_1)\forall a_0,a_1\in 𝒜$	In a model used to screen for lung cancer from chest X‐rays, the average score among males and females without lung cancer are equal [99].	$✓$
Sufficiency‐based criteria
Conditional Use Accuracy Equality	$P(Y=1|D=1,A=a_0)=P(Y=1|D=1,A=a_1)$ and $P(Y=0|D=0,A=a_0)=P(Y=0|D=0,A=a_1)\forall a_0,a_1\in 𝒜$	In predicting hospital readmission, the probability of being readmitted given the model makes that decision and the probability of not being readmitted given the model makes that decision is the same for Black and white patients [100].
Predictive Parity	$P(Y=1|D=1,A=a_0)=P(Y=1|D=1,A=a_1)\forall a_0,a_1\in 𝒜$	In predicting hospital readmission, the probability of being readmitted given the model makes that decision is the same for Black and white patients [100].	$✓$
Well Calibration	$P(Y=1|S=s,A=a)=s\forall a\in 𝒜,s\in [0,1]$	In a model used to predict in‐hospital mortality, the predicted event rates match the observed event rates at all values of the score for males and females [101].
Test Fairness	$P(Y=1|S=s,A=a_0)=P(Y=1|S=s,A=a_1)\forall a_0,a_1\in 𝒜,s\in [0,1]$	In a model used to predict in‐hospital mortality, the predicted event rates are the same for males and females at all values of the score [101].	$✓$
Other criteria
Brier Score Parity	$𝔼[{(Y-S)}^2|A=a_0]=𝔼[{(Y-S)}^2|A=a_1],\forall a_0,a_1\in 𝒜$	In a model used to enroll patients into high‐risk care management programs, the mean squared error between the score and the label is the same for Black and white patients [18].
Overall Accuracy Equality	$P(D=Y|A=a_0)=P(D=Y|A=a_1)\forall a_0,a_1\in A$	In a model used to classify sex based on gait data collected from wearable sensors, the probability of correct classification (i.e., correctly identifying an individual's true sex) is the same for younger and older individuals [102].
Treatment Equality	$\frac{\#\{D=0,Y=1,A=a_0\}}{\#\{D=1,Y=0,A=a_0\}}=\frac{\#\{D=0,Y=1,A=a_1\}}{\#\{D=1,Y=0,A=a_1\}}\forall a_0,a_1\in A$	In a model used to screen for lung cancer from chest X‐rays, the ratio of the false negatives (patients with lung cancer incorrectly classified as not having lung cancer) and false positives (patients without lung cancer incorrectly classified as having lung cancer) is the same for both males and females [99].

Note: Mathematical definitions of group fairness criteria, their interpretation in the context of an example from the literature, and an indication of whether the criterion is a relaxation of independence ($D\perp A$), separation ($D\perp A|Y$), or sufficiency ($Y\perp A|D$). Symbols: $\#=$ number of, $P=$ probability, $E=$ expected value. Notations: $Y$: outcome, $\mathit{X}$: features used for model training, $A$: protected attribute that takes value in the set $𝒜$, $S$: model score, $D$: model classification based on thresholding $S$, $\mathit{Z}$: additional set of features.

3.2.1. Independence

3.2.1.1. Definitions

Under independence, an ML model is said to be fair if its decisions do not depend on the protected attribute (i.e., $D\perp A$). Statistical (or demographic) parity is a common measure of independence that requires that the model classify individuals into the positive class at the same rate in each group [29, 87]. Conditional statistical parity relaxes this concept by requiring the rate of positive classifications to be the same within more granular groups defined by the protected attribute and other relevant factors.

3.2.1.2. Usage Within Health‐Focused Applications

Independence‐based fairness metrics, such as statistical parity, are infrequently used in health‐focused applications as the prevalence of clinical outcomes often differs across groups defined by protected attributes (e.g., multiple sclerosis is more common in females than males). Enforcing independence may also prevent a model from learning a genuine association between the protected attribute and the outcome, potentially leading to an overall reduction in performance [88, 89]. However, independence‐based metrics may still be informative when the goal is to assess whether a model disproportionately assigns high‐risk predictions to specific groups. For example, Li et al. used statistical parity to evaluate a model predicting heart failure length of stay and in‐hospital mortality based on clinical and social determinants of health data [90]. This metric assessed whether patients from different ethnoracial groups were equally likely to be classified as high risk for the outcomes of interest. The study found disparities in prediction rates, particularly in models that omitted social determinants, and showed that including these variables can enhance fairness without sacrificing overall accuracy.

3.2.2. Separation

3.2.2.1. Definitions

Separation requires that the model's decisions do not depend on the protected attribute within the positive and negative classes (i.e., $D\perp A|Y$). This implies that, among individuals in the positive (or negative) class, the rate of making a positive (or negative) decision is consistent across groups. Common separation‐based metrics therefore aim to equalize error rates across the groups, including the false negative rate (FNR, known as equal opportunity), false positive rate (FPR, known as predictive equality), or both (known as equalized odds) [91]. Additional separation‐based metrics are detailed in Table 2, including balance for the positive class and balance for the negative class [91].

3.2.2.2. Usage Within Health‐Focused Applications

Separation‐based metrics have been widely used in health‐focused applications, although the specific choice of metric depends on the context. When false negatives have the most severe consequences, equal opportunity may be preferred. For instance, this metric was used in a study of state‐of‐the‐art computer vision models used to detect common diseases from chest X‐rays, where a false negative corresponded to incorrectly identifying a patient as not having “no finding” on their X‐rays [11]. The models had higher FNRs in several under‐served subpopulations situated at intersections of oppression, such as Hispanic female patients, potentially resulting in delayed access to care. There are also situations where balancing both the FNR and FPR is more appropriate and equalized odds should be prioritized [64]. For instance, Yang et al. [92] employed equalized odds to assess the fairness of ML‐based rapid COVID‐19 screening tools used in emergency departments as disparities in FNRs can lead to inadequate monitoring in certain groups, while imbalanced FPRs may result in disproportionate unnecessary testing.

3.2.3. Sufficiency

3.2.3.1. Definitions

Sufficiency aims to equalize error rates among individuals with similar decisions [85]. Formally, sufficiency requires that the label does not depend on the protected attribute given the model's decision (i.e., $Y\perp A|D$). The decision is therefore “sufficient” for predicting the outcome, in the sense that it subsumes the protected attributes [22]. Common sufficiency‐based metrics focus on equalizing the positive predictive value (PPV, known as predictive parity or positive predictive parity), both the PPV and negative predictive value (NPV, known as conditional use accuracy equality), and calibration (known as well‐calibration ⁶ ) [52, 86, 91, 94].

3.2.3.2. Usage Within Health‐Focused Applications

Sufficiency‐based metrics have been used to evaluate models for disease screening, management, and triage. For example, Raza et al. considered evaluating a model for predicting 30‐day hospital readmission using predictive parity across age groups [95]. In this setting, disparities in the PPV can cause the model to disproportionately flag elderly patients as being at high risk of readmission, even when their baseline health status is similar to that of younger patients. This type of bias may lead to unnecessary interventions and undue stress for elderly patients.

Calibration‐based metrics are also frequently applied in health‐focused examples. Test fairness was evaluated in our introductory example of evaluating the fairness of a commercial risk prediction algorithm used to enroll patients in a high‐risk care management program. In this example, healthcare costs were used as a proxy for health needs in training the model. As a result, at equivalent model scores, Black patients were in significantly poorer health than white patients as less money had historically been spent on their healthcare, meaning that they had to be sicker to qualify for the program [18].

3.2.4. Incompatibilities

Independence, sufficiency, and separation provide different perspectives on what it means for a model to be fair. Except under highly restrictive conditions, it is not possible for an algorithm to fulfill all criteria simultaneously [52, 64, 87]. It is therefore critical for researchers to choose which group fairness considerations are most relevant to their context. More specifically, the following pairs of criteria are incompatible, in the sense that they cannot generally be simultaneously satisfied: independence and sufficiency, independence and separation, and separation and sufficiency. A basic requirement for any of these pairs to hold is that the outcome and the protected attribute are marginally independent (i.e., $Y\perp A$). In classification problems, this means the probability of being in the positive class is the same across groups. This condition is violated in many clinical contexts, such as when a disease or outcome is more common among certain subpopulations. Supporting Information Section S3 provides additional mathematical details related to the incompatibilities and also introduces approximate (or $ϵ$) fairness, which allows for small deviations in group fairness metrics in order to address these incompatibilities [103]. Under approximate fairness, it becomes possible to satisfy multiple fairness criteria across the three categories [104].

3.3. Individual Fairness

3.3.1. Definitions

In contrast to group fairness which targets the average performance of a model across groups, individual fairness ensures that “like cases are treated alike,” an idea grounded in Aristotle's conception of justice [29, 105]. Here, we introduce a foundational concept in individual fairness known as fairness through awareness (FTA). FTA relies on similarity metrics to quantify the distance between individuals and ensure that similar individuals receive similar predictions. More formally, fairness is achieved if for any two individuals with features ${\mathit{X}}_i$ and ${\mathit{X}}_j$,

$$d_S(S_i,S_j)\le d_{\mathit{X}}({\mathit{X}}_i,{\mathit{X}}_j)$$

where $d_S$ and $d_{\mathit{X}}$ denote distance metrics defined on the score and feature spaces, respectively [29]. ⁷ A counter approach to FTA, though not an individual fairness criterion, is fairness through unawareness (FTU). FTU is intended to be a catch‐all solution to prevent bias by not explicitly including protected attributes into modeling.

3.3.2. Usage Within Health‐Focused Applications

The implementation of FTA‐based approaches critically depends on the metrics used to define similarity between individuals and their predictions. For predictions, the most commonly used metric is the absolute difference between predicted probabilities [29, 38, 107, 108, 109]. However, the metric used to measure the similarity between individuals has been a focus of ongoing research as it determines how individuals are comparable, which relies on an “awareness” of the context. For example, Zemel et al. proposed to identify the $k$ closest individuals in the feature space based on a chosen distance metric, such as the Euclidean distance, and then assess similarity by comparing the predictions across these neighborhoods [107]. In the original FTA paper, the authors suggest leaving the choice of metric to domain experts. More recently, various strategies have been proposed to learn the metric from available data [110, 111, 112].

A scoping review [38] on individual fairness in health care provides a comprehensive list of individual fairness methods and corresponding software. The review highlights that individual fairness is just emerging within health‐focused applications. For example, a small number of studies have developed models aimed at achieving individual fairness in the context of survival analysis. Motivated by an AI system used to perform needs‐based prioritization of the Medicaid waitlists, Keya et al. [108] introduced an individual fairness constraint for the Cox proportional hazards model, using Euclidean distance in the feature space to define similarity across individuals and penalizing the absolute difference in predicted hazard scores. Subsequently, Rahman and Purushotham [109] extended this approach to general survival models (e.g., non‐hazard‐based) by reformulating the fairness constraint based on the absolute difference in predicted survival probabilities. The proposed method significantly reduced disparities in predicted survival probabilities across individuals while maintaining competitive overall predictive accuracy with standard survival models in analyses of three diverse real‐world healthcare data sets.

It is important to note that individual fairness is motivated by the inherent weakness in group fairness criteria that only consider average model performance within groups. That is, there are situations in which group fairness can be satisfied, but individuals within a group can be discriminated against [29]. The incompatibilities among individual and group fairness have been examined from a theoretical perspective in several recent studies [85, 113, 114, 115]. However, under certain conditions, individual fairness can imply statistical parity [29, 116].

We close our discussion of individual fairness by briefly commenting on FTU as it is a counter‐approach to FTA. FTU warrants careful consideration in health‐focused applications as it removes protected attributes from the model. First, protected attributes may serve as critical predictors for the outcome of interest, such as age in sepsis or cardiovascular disease prediction and race in cancer screening models [117, 118, 119, 120]. Excluding these predictors can diminish overall predictive accuracy and impact all individuals adversely or lead to bias against the majority group. Second, protected attributes are often highly correlated with non‐protected attributes. Simply removing protected attributes from the model does not prevent the model from inferring them from other attributes. This is evident in a recent study [121] that found ML models can infer race from medical images. Finally, the quality of predictors included in the model may vary across the protected groups (e.g., rates of missing data, measurement error). For instance, family history is an important predictor for cancer risk prediction, but has been shown to be less reliably documented for Black participants in self‐reported family history data and therefore less useful for models developed with these data [122, 123]. To address racial disparities in data quality, Zink et al. [124] showed that including race as a predictor significantly improved model performance compared to race‐blind algorithms for colorectal cancer.

FTU is therefore not always ethical, achievable, or desirable [125] and has been a point of debate within ML, policy, and related fields [124, 126, 127, 128, 129, 130, 131, 132, 133, 134]. For example, Rajkomar et al. explicitly reject FTU in their recommendations for advancing health equity through fair ML, stating that they “purposefully do not recommend the commonly discussed fairness principle of unawareness” [40]. The Fairness and Machine Learning textbook [22] similarly critiques FTU as “ineffective and even harmful.” Within health‐focused applications, several studies argue that incorporating protected attributes can improve predictive performance in specific contexts [124, 135, 136] and reduce disparities in model accuracy [137]. However, the inclusion of these variables can also reinforce harmful assumptions, such as biological determinism in the context of race, and lead to greater stigmatization of marginalized groups [126]. To navigate these tensions, Liu et al. [138] advocate for the use of causal and interventional analysis to assist researchers in discerning whether certain variables should be included in modeling. For clinical risk prediction, Coots et al. [137] recently proposed a decision‐analytic framework for accounting for race and ethnicity. We also highlight complementary work that interrogates the broader question of how certain protected attributes, such as race [131] and gender [139], are conceptualized and operationalized within algorithmic systems [124, 140].

3.4. Causal Fairness

We next provide a brief overview of causal fairness, an increasingly popular topic in ML that warrants a dedicated review within health applications [38, 84, 141, 142, 143, 144, 145]. As the name suggests, causal fairness generally focuses on understanding the causal relationships between protected attributes and a model's decisions [37]. Although causality‐based approaches are especially valuable in health applications for disentangling mechanisms of bias, the use of causal fairness in health is in its early stages. As such, we use toy examples to introduce most of the concepts and reference existing real‐world examples whenever possible.

We focus our discussion on several common fairness notions based on counterfactuals, many of which build upon the previously introduced notions of group and individual fairness. Additional definitions that we do not cover in detail and that require background knowledge in causal inference, such as interventional fairness, path‐specific fairness, and principal fairness, are detailed in the broader ML literature [84, 146, 147, 148, 149, 150, 151, 152, 153, 154]. In the context of defining what it means to be fair, the counterfactual, or more simply the “what‐if” statement, is most often the unobserved model's decision that would have happened if the protected attribute had been different [142, 155, 156]. One of the earliest fairness criteria based on counterfactual reasoning is (individual) counterfactual fairness, which considers a model to be fair if, for a given individual, the distribution of predicted outcomes remains the same when the protected attribute is counterfactually altered while holding all other variables constant [155]. Counterfactual fairness may also be viewed as an individual fairness criterion as it is grounded in achieving similar treatment for similar individuals [141].

To further explain the idea behind counterfactual fairness, we use a simplified example adapted from the original work [155] on the topic and provide the technical definition in Table 3. Figure 3 presents the directed acyclic graph (DAG) for this example. Suppose that $Y$ is an indicator of 30‐day hospital readmission, $X$ a variable for emergency healthcare service utilization in the past 6 months that is used to predict $Y$, $A$ an indicator of membership in a socioeconomic group, and $U$ a latent variable measuring frailty that is not observed. In this example, frailty causes people to be more likely to be readmitted and to visit the emergency department. Furthermore, individuals of a particular socioeconomic group are more likely to use the emergency department, not because they are more frail (i.e., $A$ does not impact $U$), but potentially due to structural barriers such as lack of primary care access or insurance coverage. These individuals, however, are no more likely to be readmitted than anyone else with the same level of frailty. In this scenario, if a model uses $X$ to predict $Y$, it may assign higher predicted probabilities of readmission to individuals from that socioeconomic group, even though their true risk of readmission given $U$ is no higher. In terms of counterfactual fairness, this prediction is considered unfair; changing $A$ while holding $U$ constant would change $X$ and consequently the prediction. In other words, the model's prediction is sensitive to $A$ in a way that is not justified by underlying differences in health.

TABLE 3.

Common notions of causal fairness.

Metric	Definition
(Individual) Counterfactual Fairness [155]	$P(D_{A\leftarrow a}(U)=d|\mathit{X}=\mathit{x},A=a)=P(D_{A\leftarrow a^{\prime }}(U)=d|\mathit{X}=\mathit{x},A=a)\text{for any}\mathit{x},d\in \{0,1\},\text{and}\forall a,a^{\prime }\in \mathcal{A}$
Individualized Equalized Counterfactual Odds [157]	$P(D_{A\leftarrow a}(U)=d|\mathit{X}=\mathit{x},A=a,Y_{A\leftarrow a}=y)=P(D_{A\leftarrow a^{\prime }}(U)=d|\mathit{X}=\mathit{x},A=a,Y_{A\leftarrow a^{\prime }}=y)\text{for any }\mathit{x},d\in \{0,1\},\text{and }\forall a,a^{\prime }\in \mathcal{A}$
Counterfactual Equalized Odds [158]	$P(D=d|Y^0=y,A=a)=P(D=d|Y^0=y)$ for $d\in \{0,1\}$, $y\in \{0,1\}$, and $\forall a\in \mathcal{A}$
Counterfactual Predictive Parity [157]	$P(Y^0=1|D=d,A=a)=P(Y^0=1|D=d)$ for $d\in \{0,1\}$ and $\forall a\in \mathcal{A}$

Note: Symbols: $P=$ probability. Notations: $Y$: outcome, $D$: model classification, $\mathit{X}$: features used for model training, $A$: protected attribute that takes value in a set $𝒜$, $D_{A\leftarrow a}$: classification when $A=a$, $D_{A\leftarrow a^{\prime }}$: classification when the protected attribute is counterfactually altered to $A=a^{\prime }$, $D_{A\leftarrow a}(U)$: classification when $A=a$ holding $U$ constant, $D_{A\leftarrow a^{\prime }}(U)$: classification when the protected attribute is counterfactually altered to $A=a^{\prime }$ holding $U$ constant, $Y_{A\leftarrow a}$: outcome when $A=a$, $Y_{A\leftarrow a^{\prime }}$: outcome when the protected attribute is counterfactually altered to $A=a^{\prime }$, $Y^0$: potential outcome with $D=0$.

FIGURE 3.

Directed acyclic graph (DAG) for the counterfactual fairness example. The DAG illustrates the relationships among the outcome, $Y$, the feature $X$, the protected attribute $A$, and the unobserved factor $U$. Gray nodes denote variables that are observed and white nodes denote variables that are unobserved.

While an oversimplified view of the many factors that impact hospital readmission, this example illustrates both the value of causal thinking in understanding the contributors to unfairness as well as the potential difficulties. Namely, causal fairness criteria typically rely on a well‐defined causal model, which can be difficult to specify in more complex situations, and whose assumptions cannot generally be verified using observational data [142, 153, 155]. Moreover, issues of identifiability can also arise (i.e., situations where causality‐based fairness notions cannot be measured uniquely from the data) and we refer readers to the work of Makhlouf et al. for further discussion of this topic [143]. That said, numerous other metrics based on counterfactuals have been proposed to quantify fairness [157].

For example, there are a number of counterfactual extensions of group fairness metrics. Pfohl et al. introduced an individual‐level extension of equalized odds in order to evaluate prediction models for prolonged inpatient length of stay and mortality across groups determined by gender, race, and age [158]. The criterion, individual equalized counterfactual odds, is satisfied if the distribution of observed and counterfactual predictions align for an individual when their protected attribute is changed (i) given all other variables are held constant and (ii) conditioned on the observed outcome matching the counterfactual outcome. The condition in (ii) distinguishes individual equalized counterfactual odds from counterfactual fairness. The purpose of adding this condition is to ensure that the fairness comparison is made only when the observed and counterfactual outcomes align  . In our running example, individualized counterfactual equalized odds means that for a given patient, the distribution of predictions of 30‐day readmission will stay the same if that patient were counterfactually assigned to a different socioeconomic group, provided their readmission status itself would also remain unchanged and holding other variables constant. This notion of fairness helps identify whether a model's use of socioeconomic group reflects unjustified bias, as opposed to reflecting legitimate differences in outcomes due to other factors. Additional extensions of group fairness criteria, including counterfactual equalized odds and counterfactual predictive parity, are formally defined in Table 3.

4. Outlook

Our work explores why models are unfair and the various ways fairness has been defined in ML for health. Despite substantial progress in recognizing the importance of bias and in applying group, individual, and causal‐based criteria in health applications, there is a lack of consensus on how to appropriately quantify fairness [159, 160]. Each framework comes with unique benefits and challenges. Group fairness criteria are relatively easy to implement and interpret, but limited by potential individual fairness violations and by incompatibilities that exist among various criteria. Likewise, individual fairness criteria prioritize similar outcomes for similar individuals to enhance equity in decision‐making, but can mask disparities at the group level and leave certain populations underserved. Additionally, individual fairness requires researchers to define appropriate similarity metrics tailored to their specific context. Causal fairness, which is gaining traction in health, enables researchers to investigate sources of unfairness to inform fair decision‐making. However, specifying reliable causal models can be challenging in health contexts. Compared to group and individual fairness, causal fairness remains particularly underexplored in ML for health [84, 158]. This gap presents opportunities for future work in evaluating the feasibility and utility of causal‐based criteria in diverse applications.

In practice, balancing different fairness frameworks requires weighing the benefits of group‐level equity, individual‐level considerations, and causal understanding [161]. For instance, in large‐scale lung cancer screening, equal opportunity may be prioritized as missing a diagnosis can lead to delayed treatment and poor clinical outcomes for certain subgroups [40]. Conversely, in predicting organ compatibility after transplant, individual fairness may take precedence to ensure that patients with similar clinical profiles and predicted outcomes are treated similarly irrespective of protected attributes given the scarcity of donor organs [162]. Causal fairness may be relevant in either setting if the goal is to identify and interrogate sources of bias [37]. We therefore recommend that researchers stay informed about the full range of available fairness methods and collaborate within interdisciplinary teams (e.g., data scientists, clinicians, ethicists, social scientists) to make context‐appropriate decisions. Significant effort has been made to provide practical recommendations for incorporating fairness into ML, including the development of a bias evaluation checklist, a framework for integrating health equity into model development, and recommendations for identifying ethical concerns [30, 163, 164]. Importantly, these resources emphasize that a quantitative understanding of bias is just one component of operationalizing fairness in ML for health. Establishing common principles and standards that prioritize fairness throughout the entire model development pipeline, from data collection and problem formulation to deployment, is essential, as post hoc fairness assessments alone are insufficient [22, 31, 165, 166].

While we focus on how to define fairness, strategies for mitigating bias within ML applications are an equally active and important area of research. We refer interested readers to a recent systematic review that presents various bias mitigation strategies and their use in practice [70]. Importantly, a necessary step in any bias mitigation strategy is to select the fairness definition that the mitigation strategy will attempt to enforce. Group fairness metrics are the most common targets, though alternative strategies have been proposed [68]. Broadly, bias mitigation can be performed by de‐biasing the training data (pre‐processing), during model training (in‐processing), and after model training (post‐processing). Pre‐processing techniques include resampling [167] or reweighting samples [168] to correct imbalances among groups. One approach to in‐processing is to include additional regularization terms to penalize a model's deviation from a pre‐specified fairness metric [169]. In post‐processing, for example, a technique to achieve equalized odds is to set distinct decision thresholds for different groups [64]. However, a practical consideration in any bias mitigation strategy is the well‐known trade‐off between achieving fairness and high overall accuracy and calibration [48, 65, 67]. This tension stems from the fact that enforcing fairness constraints at any stage of the ML pipeline can result in a decrease in overall predictive performance to ensure more equitable performance across specific groups or individuals. For a comprehensive examination of how bias mitigation strategies can be implemented throughout the AI development process within health care, Gichoya et al. [62] systematically identify common pitfalls and propose practical solutions at each stage of the process.

Lastly, health data present inherent and unique challenges, making it difficult or sometimes impossible to measure fairness accurately [170, 171, 172, 173]. Most existing fairness definitions were developed outside the healthcare context and may not adequately capture the forms of bias that arise in health‐related settings [54]. For instance, most fairness approaches focus on a single protected attribute, which limits their applicability in complex, real‐world scenarios that involve imperfect models, multiple intersecting attributes, and numerous practical considerations across the ML pipeline, such as data limitations, resource constraints, and policy or operational impacts [174]. Moreover, attributes like religion, gender identity, and socioeconomic status are often misreported, incompletely recorded, or entirely absent due to sociocultural issues and inconsistent data collection practices [73, 175]. For example, socioeconomic status is frequently missing in observational health data, though proxy variables such as insurance type may be available [176]. To address such limitations, the concept of proxy fairness has been introduced to enable fairness assessments using variables that approximate the protected attributes of interest [177]. When neither protected attributes nor reliable proxies are available, fairness metrics that account for missing data become essential. While by no means a remedy for inadequate data collection, there has been progress in fairness approaches that operate with limited data on protected attributes and that may prove useful for health applications [178, 179]. To date, however, many protected attributes have been omitted from fairness considerations, which underscores the necessity of future methodological work to acknowledge “a multiplicity of considerations, from privacy preservation, context sensitivity and process fairness, to an awareness of sociotechnical impact and the increasingly important role of inclusive and participatory research processes” [180]. We anticipate future work in these aforementioned directions, particularly in strategies focused on data collection and data equity to establish a strong foundation for fair decision‐making.

Author Contributions

Je.G. conceived and supervised the study. Je.G. and Ji.G. drafted the manuscript. P.V. provided clinical oversight, Z.R.M. and C.H. provided guidance on machine learning, and H.T. provided guidance on ethics. All authors provided valuable feedback on the manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supplementary materials: Supporting Information.

Gao J., Chou B., McCaw Z. R., et al., “What Is Fair? Defining Fairness in Machine Learning for Health,” Statistics in Medicine 44, no. 20‐22 (2025): e70234, 10.1002/sim.70234.

Data Availability Statement

The authors have nothing to report.