Hostname: page-component-77f85d65b8-bl4lz Total loading time: 0 Render date: 2026-03-26T13:10:03.774Z Has data issue: false hasContentIssue false
  • English
  • Français

Evaluating precision medicine tools in cystic fibrosis for racial and ethnic fairness

Published online by Cambridge University Press:  07 May 2024

Stephen P. Colegate Open the ORCID record for Stephen P. Colegate [Opens in a new window] ,
Anushka Palipana ,
Emrah Gecili Open the ORCID record for Emrah Gecili [Opens in a new window] ,
Rhonda D. Szczesniak  and
Cole Brokamp
Stephen P. Colegate*
Affiliation:
Division of Biostatistics and Epidemiology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA
Anushka Palipana
Affiliation:
School of Nursing, Duke University, Durham, NC, USA
Emrah Gecili
Affiliation:
Division of Biostatistics and Epidemiology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA Department of Pediatrics, University of Cincinnati, Cincinnati, OH, USA
Rhonda D. Szczesniak
Affiliation:
Division of Biostatistics and Epidemiology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA Department of Pediatrics, University of Cincinnati, Cincinnati, OH, USA
Cole Brokamp
Affiliation:
Division of Biostatistics and Epidemiology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, USA Department of Pediatrics, University of Cincinnati, Cincinnati, OH, USA
*
Corresponding author: S. P. Colegate; Email: stephen.colegate@cchmc.org
Rights & Permissions [Opens in a new window]

Abstract

Introduction:

Patients with cystic fibrosis (CF) experience frequent episodes of acute decline in lung function called pulmonary exacerbations (PEx). An existing clinical and place-based precision medicine algorithm that accurately predicts PEx could include racial and ethnic biases in clinical and geospatial training data, leading to unintentional exacerbation of health inequities.

Methods:

We estimated receiver operating characteristic curves based on predictions from a nonstationary Gaussian stochastic process model for PEx within 3, 6, and 12 months among 26,392 individuals aged 6 years and above (2003–2017) from the US CF Foundation Patient Registry. We screened predictors to identify reasons for discriminatory model performance.

Results:

The precision medicine algorithm performed worse predicting a PEx among Black patients when compared with White patients or to patients of another race for all three prediction horizons. There was little to no difference in prediction accuracies among Hispanic and non-Hispanic patients for the same prediction horizons. Differences in F508del, smoking households, secondhand smoke exposure, primary and secondary road densities, distance and drive time to the CF center, and average number of clinical evaluations were key factors associated with race.

Conclusions:

Racial differences in prediction accuracies from our PEx precision medicine algorithm exist. Misclassification of future PEx was attributable to several underlying factors that correspond to race: CF mutation, location where the patient lives, and clinical awareness. Associations of our proxies with race for CF-related health outcomes can lead to systemic racism in data collection and in prediction accuracies from precision medicine algorithms constructed from it.

Keywords

Cystic fibrosisgroup fairnesslung functionmedical monitoringprecision medicinepulmonary function tests

Information

Type
Research Article
Information
Journal of Clinical and Translational Science , Volume 8 , Issue 1 , 2024 , e94
Creative Commons
Creative Common License - CCCreative Common License - BY
This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited.
Copyright
© The Author(s), 2024. Published by Cambridge University Press on behalf of Association for Clinical and Translational Science

Introduction

Cystic fibrosis (CF) is a disease that causes the production of abnormally thick secreted fluids [1–Reference Bush, Amaral and Davies3], especially inside the lungs and the pancreas [1–Reference Rosenfeld, Emerson and Williams-Warren7]. As a result, CF lung disease progression is marked by recurring rapid declines in lung function in the form of acute respiratory events, clinically referred to as pulmonary exacerbation (PEx) events [Reference Parkins, Rendall and Elborn8–Reference Sanders, Solomon and Beckett10]. Symptoms include coughing, sputum production, wheezing, chest tightness, difficulty breathing or shortness of breath, and fever [Reference Goss, Edwards, Ramsey, Aitken and Patrick11]. Precision medicine algorithms that predict these attenuated declines in CF have been developed in recent years [Reference Yanaz, Yilmaz Yegit and Gulieva12,Reference Szczesniak, Su and Brokamp13]. While many of these algorithms entail different approaches, their primary purpose is to direct care and resources to high-need CF patients at the right time [Reference Ferryman and Pitcan14]. Development and targeted use of advanced personalized treatments such as ivacaftor and lumacaftor are highlights of the CF community embracing precision medicine [Reference Martiniano, Sagel and Zemanick15]. Precision medicine, however, permits discriminatory and harmful impacts of structural racism that could potentially impact groups that have been historically marginalized [Reference Geneviève, Martani, Shaw, Elger and Wangmo16]. Racial bias can be introduced in building and analyzing datasets, but it can also be the result of precision medicine research [Reference Ferryman and Pitcan14].

Sparking much of the latest interest in PEx prediction was the CF Foundation Learning Network’s adaptation of a data-driven definition of PEx. They considered changes in lung function, measured as a forced expiratory volume in 1 s (FEV1) of % predicted, relative to the baseline to identify PEx. The definition is known as the FEV1-indicated exacerbation signal (FIES) and is applied over time for each individual patient [Reference List, Solomon and Bichl17]. We have developed a nonstationary Gaussian stochastic process model to predict PEx using demographic, encounter-level, and hospitalization data from the US CF Foundation Patient Registry (CFF-PR) [Reference Szczesniak, Su and Brokamp13]. This precision medicine algorithm incorporates clinical measures and has been expanded to include place-based measures to forecast FIES risk [Reference Palipana, Gecili and Brokamp18] including traffic (on the basis of primary road density and secondary road density of the ZIP code tabulation area), community material deprivation, and greenspace. The algorithm then determines the probability a CF patient will experience a future PEx event from the date of an encounter by forecasting FIES risk. The model has been shown to accurately predict rapid lung function declines up to two years from a clinical evaluation (median area under the receiver operating characteristic [ROC] curve 0.817, 95% confidence interval [CI]: 0.814, 0.822), serving as a useful clinical tool to identify for whom and when a FIES-defined PEx event is imminent [Reference Gecili, Brokamp and Rasnick19]. Earlier prediction and identification of a PEx event allow for earlier preventative interventions; therefore, the accuracy of any precision health algorithm will influence an individual’s morbidity and mortality.

Concerns around racial equity in many precision health algorithms have been introduced [Reference Cheng, Goodman and Bogue20,Reference Obermeyer, Powers, Vogeli and Mullainathan21] because they include race as a predictor [Reference Vyas, Eisenstein and Jones22], but racial inequality is also present in real-world precision health algorithms that do not explicitly use race as a predictor [Reference Pennington, Rasnick and Martin23]. Group-level fairness is defined as the desired state of achieving similar model performance across subpopulations partitioned by protected attributes, such as race and ethnicity [Reference Binns24]. Most precision medicine research today, however, does not incorporate group fairness adequately into the statistical evaluation process and may not even consider group bias for aspects such as variable measurement and design selection [Reference Fleisher25–Reference McCradden, Joshi, Mazwi and Anderson27]. Rather, researchers focus on the accuracy and interpretation of the algorithm being developed and are only evaluated based on individual fairness – similar individuals within a population being treated similarly [Reference Corbett-Davies, Pierson, Feller, Goel and Huq28]. Even though symptom severity and frequency vary between individuals, precision medicine algorithms commonly use covariates associated with race and ethnicity [Reference Collaco, Blackman, McGready, Naughton and Cutting29] that could induce differences in PEx prediction accuracies between racial and ethnic groups. Using predictors associated with race or ethnicity does not necessarily imply predictions will be unfair, but we do not wish to see racial and ethnic differences in accuracies in PEx prediction from our precision medicine algorithm. We screened predictors from our PEx precision medicine algorithm to determine if they are correlated with race or ethnicity, as this is most likely how racial and ethnic differences are incorporated in prediction accuracies.

Methods

The CFF-PR collects information on CF patient encounters who receive care in CF Foundation-accredited care centers [Reference Knapp, Fink and Goss30]. Race and ethnicity information was collected using categories (White, Black or African American, Other, Asian, Native Hawaiian or other Pacific Islander, American Indian or Alaskan native) defined by the CF Foundation based on patient-level clinical records from each site [Reference Knapp, Fink and Goss30]. We then further categorized the patient’s race as either “White” if they identified as only White, “Black” if they identified as Black, and “Other” if any other race (besides White or Black) was selected. We defined a patient’s ethnicity whether they self-identified as either Hispanic or non-Hispanic. The CFF-PR cohort was primarily composed of White patients (White: n = 24,490 [92.8%], Black: n = 1,172 [4.4%], Other: n = 730 [2.8%]) and non-Hispanic patients (non-Hispanic: n = 23,392 [88.6%], Hispanic: n = 2,045 [7.7%]) with CF. There were 955 patients (3.6%) who did not report their ethnicity, due to a combination of patients refusing to report and healthcare centers not collecting this information. We did not consider patients with unknown ethnicity since the rate of missingness differs within each racial group (White: n = 691 [2.6%, Other: n = 206 [2.8%], Black: n = 58 [0.2%]).

The precision medicine algorithm models personalized thresholds of rapid lung function decline [Reference Szczesniak, Su and Brokamp13]. The algorithm is a non-stationary stochastic process model comprised of fixed effects (age, F508del, birth cohort, FEV1 at baseline, enzyme use, Pseudomonas aeruginosa, methicillin-resistant Staphylococcus aureus, Medicaid insurance use, CF-related diabetes mellitus, outpatient visits in last year, acute exacerbations in last year), between-patient heterogeneity, and a continuous-time integrated Brownian motion process to determine hyperlocal, dynamic predictive probabilities in lung function (FEV1) measurement. The algorithm was applied to data from 30,879 US CFF-PR patients, and the median (95% CI) area under the ROC curve estimates was 0.817 (0.814, 0.822). While the precision medicine algorithm has reasonable accuracy in personalized rapid lung function decline predictions, the algorithm has not been evaluated for group fairness.

We analyzed each predictor and outcome to determine if associations with race and ethnicity could be responsible for any model unfairness. Predictors considered include gender (male, female), F508del mutation (homozygous, heterozygous, neither/unknown), insurance payor status (private or non-private), smoking status (smoker, nonsmoker), smoking household (yes, no), secondhand smoke (yes, no), primary road density and secondary road density as a proxy for traffic exposure (total length of all roads in meters in the ZIP Code Tabulation Area [ZCTA] divided by the total area in square meters of the ZCTA), a community material deprivation index [Reference Brokamp, Beck, Goyal, Ryan, Greenberg and Hall31] (a census tract-level deprivation index based on five different census tract-level variables related to material deprivation, derived from the 2015 5-year American Community Survey), fraction of surrounding land characterized as greenspace (using the National Land Cover Database), straight-line distance (in meters) and drive time (in 5-minute intervals) to the healthcare center, baseline age (patient’s age at first encounter), number of encounter visits, number of PEx events, and the amount of time since baseline age at encounter.

Rapid lung function decline is defined by a decrease in FEV1 of more than 10% predicted from the maximum observed FEV1 within the past 12 months [Reference Siracusa, Weiland and Acton32,Reference Schechter, Schmidt, Williams, Norton, Taylor and Molzhon33]. We identified a PEx using predictions of rapid lung function decline based on FIES [Reference Petren, Thompson, O’Neil, Elbert and Ren34–Reference Bichl, Rangaraj, O’Malley, Rogers, Ward and Palla36]. FIES is determined by first defining the patient’s baseline FEV1 as the average of the highest two FEV1 measurements in the past 12 months when not under intravenous antibiotic treatment [Reference Bouzek, Ren, Thompson, Slaven and Sanders37]. If only one valid FEV1 was available, it was used as the baseline. A FIES-defined PEx event occurs when the FEV1% predicted declines at least 10% or more [Reference Petren, Thompson, O’Neil, Elbert and Ren34]. The FIES definition excludes any lung function measurements within 28 days of a previous PEx to ensure accurate patient baseline FEV1. We calculated prediction accuracy by comparing the PEx probability from the precision medicine algorithm with whether a PEx eventually occurred for 3-, 6-, and 12-month prediction horizons.

A ROC curve [Reference Mandrekar38] was used to determine the optimal cutoff probability for a PEx occurrence. Group-specific ROC curves were then implemented using this optimal cutoff probability to achieve group-specific sensitivity and specificity. ROC curves were formulated by contrasting a patient’s PEx probability from the precision medicine algorithm to actual PEx outcomes for the prediction horizon at each clinical visit. Actual PEx outcomes were determined by whether the patient was clinically evaluated and confirmed to have at least one PEx within the prediction horizon. The area under the ROC curve (AUC), sensitivity, and specificity were calculated overall and for each group-specific subpopulation using Youden’s J statistic [Reference Youden39]. The AUC was used as a benchmark for the PEx precision medicine algorithm performance [Reference Bradley40]. The 95% CIs for sensitivity and specificity were obtained with 2000 stratified bootstrap replicates. Statistical computing was performed in R version 4.2.3, specifically with the pROC (ver. 1.18.2) package to perform ROC curve analyses [Reference Robin, Turck and Hainard41].

Results

The CFF-PR cohort consists of patients with CF (n = 26,392) aged 6 years of age or older monitored between January 3, 2003, and December 31, 2017. The cohort was 47.9% female (n = 12,634) with a median baseline age of 11.1 years (25th percentile: 6.2, 75th percentile: 18.9). Each study participant was followed for a median of 7.8 years (25th percentile: 3.7, 75th percentile: 12.4). Overall, 92.8% (n = 24,490) of patients self-identified as White, and 4.4% (n = 1172) self-identified as Black. Furthermore, 88.6% (n = 23,392) self-identified as non-Hispanic, and 7.7% (n = 2045) self-identified as Hispanic. 2.6% (n = 691) of White patients, 4.9% (n = 58) of Black patients, and 28.2% (n = 206) of patients who self-reported as another race did not self-report their ethnicity. Since the rate of missing reported ethnicity differs by race, we did not consider those who did not report their ethnicity. Different genetic mutations for CF were considered: 47.3% (n = 12,484) were F508del homozygous, 36.9% (n = 9744) were F508del heterozygous, and the remaining 15.8% (n = 4164) were neither F508del or had an unknown mutation.

We examined predictors and the outcome to identify possible reasons for racial discriminatory performance in the PEx precision medicine algorithm. Newborn screening for CF has rapidly expanded through DNA tests for CF mutations, so seeing differences by mutation and racial identification is not necessarily novel [Reference Ross42]. From Figure 1, 49.2% (n = 12,041) and 36.7% (n = 8,988) of White patients were predominately F508del homozygous or heterozygous, respectively, whereas 41.1% (n = 482) and 40.3% (n = 472) of Black patients were mainly F508del heterozygous or neither, respectively. The distribution of F508del was more balanced in patients who self-identified as another race, with 30.8% homozygous (n = 225), 37.5% heterozygous (n = 274), and 31.6% (n = 231) neither/unknown. Differences existed by racial group in the distribution of primary road density, secondary road density, drive time (in minutes), and straight-line distance (in kilometers) to the nearest healthcare center (Table 1). On average, Black patients lived in neighborhoods with higher densities of primary and secondary roadway, with shorter distances and shorter travel times to their CF care center. We did not see any racial differences in gender (Other: 49.0% female, Black: 48.5% female, White: 47.8% female; p = 0.7183) or smoking status (p = 0.8856), but there were racial differences in smoking households (Other: 4.7%, Black: 3.9%, White: 2.5%; p < 0.0001) and secondhand smoke exposure (Black: 8.1%, Other: 6.4%, White: 5.7%; p = 0.0020). Black patients were also less likely to have private health insurance (Black: 28.2%, Other: 37.0%, White: 51.6%; p < 0.0001). White patients tended to live in ZCTAs with a higher average percentage of greenspace than both Black patients or patients who self-identified as another race (White: 82.8%, Other: 72.3%, Black: 70.7%; p < 0.0001). Black patients were more likely to live in neighborhoods that had a higher average community deprivation index (Black: 0.404, Other: 0.364, White: 0.335; p < 0.0001). The average number of CF encounters was different by racial group, with White patients having a higher number of clinical visits on average (White: 38.1, Black: 31.3, Other: 29.6; p = 0.035). Consequently, White patients had a higher average number of encounters with a PEx (White: 9.19, Black: 8.37, Other: 6.77; p < 0.0001) and a higher average number of encounters with no PEx (White: 21.9, Black: 17.1, Other: 16.9; p < 0.0001). The average number of CF encounters per year was lowest for Black patients (Other: 5.30, White: 5.25, Black: 4.95; p < 0.0001). We did not see racial differences in the average rate of encounters with a PEx per year (White: 1.19, Black: 1.13, Other: 1.10; p = 0.0154) and in the average rate of encounters with no PEx per year (Other: 2.97, White: 2.87, Black: 2.74; p = 0.0436). Black patients recorded fewer clinical visits per year when compared with other races.

Figure 1. F508del mutation by racial group in the US Cystic Fibrosis Foundation Patient Registry analysis cohort.

Table 1. Counts and averages of each predictor with 95% confidence intervals among the racial groups in the US Cystic Fibrosis Foundation Patient Registry analysis cohort

Group-specific AUC, sensitivity, and specificity were compared with the overall AUC sensitivity and specificity to evaluate model performance between each racial and ethnic group (Figs. 2 and 3 and Supplemental Table S1). Black patients had lower sensitivity (3-month: 0.596, 95% CI: 0.582, 0.608; 6-month: 0.607, 95% CI: 0.595, 0.618; 12-month: 0.608, 95% CI: 0.598, 0.619) for every prediction horizon compared with White patients (3-month: 0.627, 95% CI: 0.625, 0.630; 6-month: 0.628, 95% CI: 0.626, 0.630; 12-month: 0.620, 95% CI: 0.618, 0.622) and those who self-identified with another racial group (3-month: 0.638, 95% CI: 0.620, 0.656; 6-month: 0.672, 95% CI: 0.657, 0.686; 12-month: 0.623, 95% CI: 0.611, 0.636). PEx predictions for Black patients also had lower specificity (3-month: 0.608, 95% CI: 0.595, 0.622; 6-month: 0.615, 95% CI: 0.604, 0.625; 12-month: 0.622, 95% CI: 0.610, 0.635) for every prediction horizon compared with White patients (3-month: 0.641, 95% CI: 0.638, 0.643; 6-month: 0.653, 95% CI: 0.651, 0.656; 12-month: 0.655, 95% CI: 0.653, 0.657) and patients who self-identified with another race (3-month: 0.611, 95% CI: 0.594, 0.626; 6-month: 0.586, 95% CI: 0.572, 0.602; 12-month: 0.627, 95% CI: 0.610, 0.643). In each case, Black patients had the worst prediction performance from the PEx precision medicine algorithm in terms of AUC, sensitivity, and specificity. Actual PEx outcomes were determined using future clinical evaluations during the prediction horizon, but the results were also similar when actual PEx outcomes were instead defined only on the date of clinical evaluation during the prediction horizon (see Supplemental Table S2).

Figure 2. Area under the receiver operating characteristic (ROC) curve (AUC) for the 3-, 6-, and 12-month prediction horizons by racial and ethnic group. Overall AUC is indicated by the horizontal line. Group-specific AUC and their respective 95% confidence interval are displayed as points and vertical lines, respectively.

Figure 3. Optimal sensitivities and specificities by racial and ethnic group achieved by the precision medicine algorithm for 3-, 6-, and 12-month exacerbation prediction. The average optimal sensitivity (0.623) and average optimal specificity (0.648) are indicated by the horizontal lines.

Prior research work has shown disparities in pulmonary function exist between Hispanic and non-Hispanic patients with CF [Reference McGarry, Neuhaus, Nielson, Burchard and Ly43]. Even though non-Hispanic patients represent 88.6% of the cohort, the PEx precision medicine algorithm had similar performance between Hispanic and non-Hispanic ethnic groups. AUC values are similar for Hispanic patients (3-month: 0.668, 95% CI: 0.663, 0.673; 6-month: 0.678, 95% CI: 0.674, 0.682; 12-month: 0.678, 95% CI: 0.674, 0.683) and non-Hispanic patients (3-month: 0.672, 95% CI: 0.671, 0.674; 6-month: 0.682, 95% CI: 0.680, 0.683; 12-month: 0.679, 95% CI: 0.677, 0.680). We also allowed group-specific optimal cutoffs and verified if changing the overall optimal cutoff improves model prediction accuracy. When allowing each group to have their own optimal cutoff, we saw a similar performance from the PEx precision medicine algorithm for both racial and ethnic groups (see Supplemental Table S3). Ultimately, we found no evidence of ethnic discrimination in model prediction performance from the PEx precision medicine algorithm.

Discussion

We characterized the accuracies of a precision medicine tool for PEx prediction by race and ethnicity, which demonstrates the need to optimize an algorithm by balancing both accuracy and group fairness. Our results show that racial, but not ethnic, differences in the PEx prediction algorithm accuracies exist when applied to the CFF-PR data. We conclude the PEx precision medicine algorithm is racially biased against Black patients with worse PEx predictions than those who self-identify with another race. These discrepancies are not due to the differences in sample size but rather by ignoring group-level fairness in prediction accuracies by racial group. Even though we see differences in model accuracies between Hispanic and non-Hispanic, groups of proportionately different sample sizes, the nature of the difference in model accuracy does not lend itself an unfair advantage to either group, since Hispanic had better sensitivity while non-Hispanic had better specificity.

The same cannot be said for the discrepancy we see between the races, and we are left to wonder why the PEx prediction algorithm yields worse sensitivity and specificity to Black patients. The PEx prediction algorithm is formulated by treating each CF individual in the CFF-PR cohort equally, so the discriminatory performance of this algorithm must be caused by some underlying factors. We examined predictors and outcome variables to formulate three main reasons for discriminatory model performance in the PEx prediction algorithm: (i) CF mutation: While the prediction algorithm is treated on individuals with CF, differences in F508del mutation exist in the cohort. The severity of CF disease and the frequency of PEx events change, in part, based on the F508del mutation [Reference Stanke, Becker and Kumar2,Reference Collaco, Blackman, McGready, Naughton and Cutting29], which varies between the races. (ii) Location: Black patients tended to live in locations with higher road densities. Increased roads in these areas usually lead to increased vehicle traffic and therefore are associated with increased air pollution exposure. (iii) Accessibility: Even though Black patients tended to live closer to their nearest healthcare center on average, and the drive time to arrive at their healthcare center is also lower on average, the encounter rates are noticeably lower for Black patients.

There are several potential reasons for the discrepancy – lack of interaction or trust in the healthcare system, socioeconomic status (SES), and affordability for essential healthcare services, accessibility, and the quality of health care, all of which have been studied more generally in environmental health research [Reference Payne-Sturges, Gee and Cory-Slechta44,Reference Boyd, Lindo, Weeks and McLemore45]. More frequent clinic visits are associated with better lung function outcomes in CF [Reference Szczesniak, Su and Brokamp13], but there has been a shift in the care paradigm toward telehealth visits that were partially motivated to overcome barriers to access that were heightened during the coronavirus disease 2019 pandemic [Reference Albon, Van Citters and Ong46]. However, recent research on telehealth utilization in US CF patients during the pandemic showed that individuals who identified as Black or Hispanic/Latino and those who reported having financial constraints were less likely to have a telehealth visit [Reference Albon, Van Citters and Ong46]. Although individual-level proxies of SES (e.g., Medicaid/state insurance use) are linked to lung function and survival in CF, Albon and colleagues found no association between SES factors and telehealth utilization in the aforementioned study, and differences due to race/ethnicity persisted after SES adjustments. Coupling prior literature with our study findings, interventions to improve chronic disease management, including outcome prognostication, for Black people with CF may have suboptimal impact unless fairness is considered.

The obstacles to algorithm fairness that we identified also pose challenges for therapeutics development, which are expected to grow in light of the changing demographics of CF worldwide [Reference McGarry, Williams and McColley47,Reference Stephenson, Stanojevic, Sykes and Burgel48]. While there is generally a paucity of transparent (i.e., freely open) algorithms that are subjected to critical appraisal from and co-production by patient communities [Reference Pearson, Califf and Roper49], the CF patient subgroups identified from our study at greatest risk for algorithmic bias have also historically been underrepresented in CF clinical research (e.g., identifying as nonwhite, rural, or socioeconomically disadvantaged) [Reference Lowton50,Reference Goss, Rubenfeld, Ramsey and Aitken51]. Although research participation among CF minority populations has been a longstanding challenge in CF, it is of greater importance in the modern era of care, given the need to develop therapies for the remaining 5%–6% of the US CF population for whom highly effective modulator therapies are currently unavailable [52]. Individuals who have ultra-rare mutations that are not FDA approved for modulator treatment tend to identify as Black/African American or nonwhite Hispanic, and they have the lower average lung function, compared with their white counterparts [Reference McGarry and McColley53].

A larger issue that is raised from this research is how to address group-level fairness in CF precision medicine tools. Although unintentional, both (i) associations of model predictors with race and (ii) associations of our proxies for health outcomes with race can lead to prediction algorithms that could replicate systemic racism in the data and the system used to create it. Clinical implementation of this PEx prediction tool could possibly enhance racial disparities in access to care for CF patients. Individual-level fairness and group-level fairness cannot be maximized simultaneously in a prediction algorithm [Reference Corbett-Davies, Pierson, Feller, Goel and Huq28]. Therefore, we must examine the tradeoff in the accuracy of the PEx prediction algorithm with respect to group-level fairness [Reference Ferryman and Pitcan14]. Addressing group-level fairness in precision medicine models, particularly in respiratory diseases like CF, is essential to ensure that these tools benefit all patients equitably, irrespective of their race, ethnicity, or socioeconomic background. There have been some methods suggested to effectively address bias and factors to consider for inclusion or exclusion in these models. These suggestions include (i) diverse data representation which ensures that the data used to train precision medicine models are representative of the diverse patient population affected by the condition [Reference Vyas, Eisenstein and Jones22], (ii) bias detection and correction which may require using advanced statistical tools to detect and correct biases in the model [Reference Mehrabi, Morstatter, Saxena, Lerman and Galstyan54], and (iii) group-specific model adjustment that allows developing separate models for groups with distinct characteristics or adjusting models to account for known disparities [Reference Obermeyer, Powers, Vogeli and Mullainathan21].

There is a critical lack of evaluation of racial and ethnic fairness in precision health medicine within CF patients. Discrepancies of the PEx prediction algorithm on the CFF-PR cohort by racial and ethnic group must raise the awareness of group-level bias in precision medicine algorithm development in CF research. We hope to invite discussions on how to promote ways of addressing group-level fairness in statistical modeling research. By not addressing group fairness, researchers run the risk of developing statistical models that puts those from minority populations at a strong disadvantage when it comes to model accuracy and performance, further exacerbating racial inequalities in CF outcomes and care. Precision medicine tools can then be developed to better meet the needs of healthcare professionals and promote equitable patient care.

Supplementary material

The supplementary material for this article can be found at https://doi.org/10.1017/cts.2024.532.

Acknowledgments

The authors would like to thank the CF Foundation for the use of CF Foundation Patient Registry data to conduct this study. Additionally, we would like to thank the patients, care providers, and clinic coordinators at CF centers throughout the USA for their contributions to the CF Foundation Patient Registry.

Author contributions

All authors contributed to the study’s conception and design. The first draft of the manuscript was written by Stephen Colegate, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Each author’s contributions are provided below: Stephen Colegate: conceptualization, methodology, software, formal analysis, writing – original draft, and writing – review and editing. Emrah Gecili: writing – review and editing. Anushka Palipana: data curation and writing – review and editing. Rhonda D. Szczesniak: conceptualization, methodology, writing – original draft, writing – review and editing, and funding acquisition. Cole Brokamp: conceptualization, methodology, writing – original draft, writing – review and editing, and funding acquisition.

Funding statement

This work was supported by the National Institutes of Health (NIH) R01 Grant “Mapping environmental contributions to rapid lung disease progression in cystic fibrosis” (Grant number R01 HL141286), the NIH R01 Grant “A framework for automated and reproducible geomarker curation and computation at scale” (Grant number: R01 LM013222), and the CF Foundation Grant “Genome-sociome informed risk prediction tools for enhanced clinical management and promotion of health equity across the lifespan” (Grant number: 005412A123).

The authors declare that no other funds, grants, or other support were received during the preparation of this manuscript.

Competing interests

None.

Ethics approval

This is an observational study. Secondary data from the NIH-funded R01 award “Mapping environmental contributions to rapid lung disease progression in cystic fibrosis” was used for the development of the PEx learning algorithm. The data comes from the CFF-PR of 26,392 CF individuals aged 6 years of age or older from 2003 to 2017. Because we are not contacting patients directly and are instead using deidentified data not collected specifically for this study, we are exempt from NIH’s guidelines for human subjects (Section 46.104(d)(4) of the NIH 2018 Revised Common Rule Requirements).

References

Consortium CFGP. Correlation between genotype and phenotype in patients with cystic fibrosis. N Engl J Med. 1993;329(18):1308–1313. doi: 10.1056/nejm199310283291804.CrossRefGoogle Scholar
Stanke, F, Becker, T, Kumar, V, et al. Genes that determine immunology and inflammation modify the basic defect of impaired ion conductance in cystic fibrosis epithelia. J Med Genet. 2010;48(1):24–31. doi: 10.1136/jmg.2010.080937.CrossRefGoogle ScholarPubMed
Bush, A, Amaral, MD, Davies, JC., et al. Hodson and Geddes’ Cystic Fibrosis. London: CRC Press; 2023. doi: 10.1201/9781003262763.CrossRefGoogle Scholar
Shteinberg, M, Haq, IJ, Polineni, D, Davies, JC. Cystic fibrosis. Lancet. 2021;397(10290):2195–2211. doi: 10.1016/S0140-6736(20)32542-3.CrossRefGoogle ScholarPubMed
Bell, SC, Robinson, PJ. Exacerbations in cystic fibrosis: 2· prevention. Thorax. 2007;62(8):723–732. doi: 10.1136/thx.2006.060897.CrossRefGoogle ScholarPubMed
Ferkol, T, Rosenfeld, M, Milla, CE. Cystic fibrosis pulmonary exacerbations. J Pediatr. 2006;148(2):259–264. doi: 10.1016/j.jpeds.2005.10.019.Google ScholarPubMed
Rosenfeld, M, Emerson, J, Williams-Warren, J, et al. Defining a pulmonary exacerbation in cystic fibrosis. J Pediatr. 2001;139(3):359–365. doi: 10.1067/mpd.2001.117288.CrossRefGoogle ScholarPubMed
Parkins, MD, Rendall, JC, Elborn, JS. Incidence and risk factors for pulmonary exacerbation treatment failures in patients with cystic fibrosis chronically infected with pseudomonas aeruginosa. Chest. 2012;141(2):485–493. doi: 10.1378/chest.11-0917.CrossRefGoogle ScholarPubMed
Sanders, DB, Bittner, RCL, Rosenfeld, M, Hoffman, LR, Redding, GJ, Goss, CH. Failure to recover to baseline pulmonary function after cystic fibrosis pulmonary exacerbation. Am J Respir Crit Care Med. 2010;182(5):627–632. doi: 10.1164/rccm.200909-1421oc.CrossRefGoogle ScholarPubMed
Sanders, DB, Solomon, GM, Beckett, etal VV. Standardized treatment of pulmonary exacerbations (STOP) study: observations at the initiation of intravenous antibiotics for cystic fibrosis pulmonary exacerbations. J Cyst Fibros. 2017;16(5):592–599. doi: 10.1016/j.jcf.2017.04.005.CrossRefGoogle ScholarPubMed
Goss, CH, Edwards, TC, Ramsey, BW, Aitken, ML, Patrick, DL. Patient-reported respiratory symptoms in cystic fibrosis. J Cyst Fibros. 2009;8(4):245–252. doi: 10.1016/j.jcf.2009.04.003.CrossRefGoogle ScholarPubMed
Yanaz, M, Yilmaz Yegit, C, Gulieva, A, et al. Electronic home monitoring of children with cystic fibrosis to detect and treat acute pulmonary exacerbations and its effect on 1-year FEV1. J Cyst Fibros Off J Eur Cyst Fibros Soc. 2023;S1569-1993(23):00911–00916. doi: 10.1016/j.jcf.2023.09.007.Google Scholar
Szczesniak, RD, Su, W, Brokamp, C, et al. Dynamic predictive probabilities to monitor rapid cystic fibrosis disease progression. Stat Med. 2020;39(6):740–756. doi: 10.1002/sim.8443.CrossRefGoogle ScholarPubMed
Ferryman, K, Pitcan, M. Fairness in Precision Medicine. Data & Society; 2018. https://datasociety.net/library/fairness-in-precision-medicine/.Google Scholar
Martiniano, SL, Sagel, SD, Zemanick, ET. Cystic fibrosis: a model system for precision medicine. Curr Opin Pediatr. 2016;28(3):312–317. doi: 10.1097/MOP.0000000000000351.CrossRefGoogle Scholar
Geneviève, LD, Martani, A, Shaw, D, Elger, BS, Wangmo, T. Structural racism in precision medicine: leaving no one behind. BMC Med Ethics. 2020;21(1):17. doi: 10.1186/s12910-020-0457-8.Google ScholarPubMed
List, R, Solomon, G, Bichl, S, et al. Improved recognition of lung function decline as signal of cystic fibrosis pulmonary exacerbation: a Cystic Fibrosis Learning Network Innovation Laboratory quality improvement initiative. BMJ Open Quality. 2023;12:e002466. doi: 10.1136/bmjoq-2023-002466.Google ScholarPubMed
Palipana, A, Gecili, E, Brokamp, C, et al. 43: monitoring and phenotyping rapid cystic fibrosis disease progression using community characteristics and environmental exposures. J Cyst Fibros. 2021;20:S22. doi: 10.1016/S1569-1993(21)01468-5.CrossRefGoogle Scholar
Gecili, E, Brokamp, C, Rasnick, E, et al. Built environment factors predictive of early rapid lung function decline in cystic fibrosis. Pediatr Pulmonol. 2023;58(5):1501–1513. doi: 10.1002/ppul.26352.Google ScholarPubMed
Cheng, TL, Goodman, E, Bogue, CW, et al. Race, ethnicity, and socioeconomic status in research on child health. Pediatrics. 2015;135(1):e225–e237.CrossRefGoogle ScholarPubMed
Obermeyer, Z, Powers, B, Vogeli, C, Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447–453.CrossRefGoogle Scholar
Vyas, DA, Eisenstein, LG, Jones, DS. Hidden in plain sight—reconsidering the use of race correction in clinical algorithms. Mass Medical Soc. 2020;383:874–882.Google ScholarPubMed
Pennington, J, Rasnick, E, Martin, LJ, et al. Racial fairness in precision medicine: pediatric asthma prediction algorithms. Am J Health Promot. 2023;37(2):239–242.CrossRefGoogle ScholarPubMed
Binns, R. On the apparent conflict between individual and group fairness. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 2020:514–524.CrossRefGoogle Scholar
Fleisher, W. What’s fair about individual fairness? In: Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 2021:480–490.Google Scholar
Gianfrancesco, MA, Tamang, S, Yazdany, J, Schmajuk, G. Potential biases in machine learning algorithms using electronic health record data. JAMA Intern Med. 2018;178(11):1544–1547.CrossRefGoogle ScholarPubMed
McCradden, MD, Joshi, S, Mazwi, M, Anderson, JA. Ethical limitations of algorithmic fairness solutions in health care machine learning. Lancet Digit Health. 2020;2(5):e221–e223.Google ScholarPubMed
Corbett-Davies, S, Pierson, E, Feller, A, Goel, S, Huq, A. Algorithmic decision making and the cost of fairness. In: Proceedings of the 23rd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining, 2017:797–806.Google Scholar
Collaco, JM, Blackman, SM, McGready, J, Naughton, KM, Cutting, GR. Quantification of the relative contribution of environmental and genetic factors to variation in cystic fibrosis lung function. J Pediatr. 2010;157(5):802–807.e3. doi: 10.1016/j.jpeds.2010.05.018.CrossRefGoogle ScholarPubMed
Knapp, EA, Fink, AK, Goss, CH, et al. The cystic fibrosis foundation patient registry. Design and methods of a national observational disease registry. Ann Am Thorac Soc. 2016;13(7):1173–1179. doi: 10.1513/AnnalsATS.201511-781OC.CrossRefGoogle ScholarPubMed
Brokamp, C, Beck, AF, Goyal, NK, Ryan, P, Greenberg, JM, Hall, ES. Material community deprivation and hospital utilization during the first year of life: an urban population-based cohort study. Ann Epidemiol. 2019;30:37–43. doi: 10.1016/j.annepidem.2018.11.008.CrossRefGoogle ScholarPubMed
Siracusa, CM, Weiland, JL, Acton, JD, et al. The impact of transforming healthcare delivery on cystic fibrosis outcomes: a decade of quality improvement at Cincinnati children’s hospital. BMJ Qual Saf. 2014;23(Suppl 1):i56–i63. doi: 10.1136/bmjqs-2013-002361.Google ScholarPubMed
Schechter, MS, Schmidt, HJ, Williams, R, Norton, R, Taylor, D, Molzhon, A. Impact of a program ensuring consistent response to acute drops in lung function in children with cystic fibrosis. J Cyst Fibros Off J Eur Cyst Fibros Soc. 2018;17(6):769–778. doi: 10.1016/j.jcf.2018.06.003.CrossRefGoogle ScholarPubMed
Petren, KM, Thompson, MD, O’Neil, T, Elbert, A, Ren, CL. FEV1 indicated exacerbation signal: a new CF registry metric to detect potential pulmonary exacerbations. Pediatric Pulmonology. 2019;54:S238–S239. doi: 10.1002/ppul.22495.Google Scholar
Thompson, M, Reed, J, Moran, S, Sanders, D, Ren, C. 74: factors contributing to clinician responses to FEV1 indicated exacerbation signal (FIES) events in a pediatric CF clinic. J Cyst Fibros. 2021;20:S37. doi: 10.1016/S1569-1993(21)01499-5.CrossRefGoogle Scholar
Bichl, S, Rangaraj, V, O’Malley, C, Rogers, T, Ward, S, Palla, J. 180: racial and ethnic disparity in FEV1-indicated exacerbation signal (FIES) events in children with cystic fibrosis. J Cyst Fibros. 2021;20:S89. doi: 10.1016/S1569-1993(21)01605-2.CrossRefGoogle Scholar
Bouzek, DC, Ren, CL, Thompson, M, Slaven, JE, Sanders, DB. Evaluating FEV1 decline in diagnosis and management of pulmonary exacerbations in children with cystic fibrosis. Pediatr Pulmonol. 2022;57(7):1709–1716. doi: 10.1002/ppul.25925.CrossRefGoogle ScholarPubMed
Mandrekar, JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol. 2010;5(9):1315–1316. doi: 10.1097/JTO.0b013e3181ec173d.Google ScholarPubMed
Youden, WJ. Index for rating diagnostic tests. Cancer. 1950;3(1):32–35. doi: 10.1002/1097-0142(1950)3:.3.0.CO;2-3>CrossRefGoogle ScholarPubMed
Bradley, AP. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognit. 1997;30(7):1145–1159. doi: 10.1016/S0031-3203(96)00142-2.CrossRefGoogle Scholar
Robin, X, Turck, N, Hainard, A, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics. 2011;12(1):77.CrossRefGoogle Scholar
Ross, LF. Newborn screening for cystic fibrosis: a lesson in public health disparities. J Pediatr. 2008;153(3):308–313. doi: 10.1016/j.jpeds.2008.04.061.CrossRefGoogle ScholarPubMed
McGarry, ME, Neuhaus, JM, Nielson, DW, Burchard, E, Ly, NP. Pulmonary function disparities exist and persist in hispanic patients with cystic fibrosis: a longitudinal analysis. Pediatr Pulmonol. 2017;52(12):1550–1557. doi: 10.1002/ppul.23884.CrossRefGoogle ScholarPubMed
Payne-Sturges, DC, Gee, GC, Cory-Slechta, DA. Confronting racism in environmental health sciences: moving the science forward for eliminating racial inequities. Environ Health Perspect. 2021;129(5):055002-1–055002-7. doi: 10.1289/ehp8186.CrossRefGoogle ScholarPubMed
Boyd, RW, Lindo, EG, Weeks, LD, McLemore, MR. On racism: a new standard for publishing on racial health inequities: health affairs forefront. Health Aff (Millwood). 2020. doi: 10.1377/hblog20200630.939347.Google Scholar
Albon, D, Van Citters, AD, Ong, T, et al. Telehealth use in cystic fibrosis during COVID-19: association with race, ethnicity, and socioeconomic factors. J Cyst Fibros Off J Eur Cyst Fibros Soc. 2021;20(Suppl 3):49–54. doi: 10.1016/j.jcf.2021.09.006.CrossRefGoogle ScholarPubMed
McGarry, ME, Williams, WA, McColley, SA. The demographics of adverse outcomes in cystic fibrosis. Pediatr Pulmonol. 2019;54(Suppl 3):S74–S83. doi: 10.1002/ppul.24434.CrossRefGoogle ScholarPubMed
Stephenson, AL, Stanojevic, S, Sykes, J, Burgel, PR. The changing epidemiology and demography of cystic fibrosis. Presse Médicale. 2017;46(6):e87–e95. doi: 10.1016/j.lpm.2017.04.012.CrossRefGoogle ScholarPubMed
Pearson, TA, Califf, RM, Roper, R, et al. Precision health analytics with predictive analytics and implementation research: JACC state-of-the-art review. J Am Coll Cardiol. 2020;76(3):306–320. doi: 10.1016/j.jacc.2020.05.043.CrossRefGoogle ScholarPubMed
Lowton, K. Trials and tribulations: understanding motivations for clinical research participation amongst adults with cystic fibrosis. Soc Sci Med. 2005;61(8):1854–1865. doi: 10.1016/j.socscimed.2005.03.039.CrossRefGoogle ScholarPubMed
Goss, CH, Rubenfeld, GD, Ramsey, BW, Aitken, ML. Clinical trial participants compared with nonparticipants in cystic fibrosis. Am J Respir Crit Care Med. 2006;173(1):98–104. doi: 10.1164/rccm.200502-273OC.CrossRefGoogle ScholarPubMed
FDA Approves Trikafta for Children Ages 2 Through 5 Years With Certain CF Mutations | Cystic Fibrosis Foundation. Published April 26, 2023. https://www.cff.org/news/2023-04/trikafta-approval-ages-2-5-mutations. Accessed March 7, 2024Google Scholar
McGarry, ME, McColley, SA. Cystic fibrosis patients of minority race and ethnicity less likely eligible for CFTR modulators based on CFTR genotype. Pediatr Pulmonol. 2021;56(6):1496–1503. doi: 10.1002/ppul.25285.CrossRefGoogle ScholarPubMed
Mehrabi, N, Morstatter, F, Saxena, N, Lerman, K, Galstyan, A. A survey on bias and fairness in machine learning. ACM Comput Surv. 2021;54(6):35–35. doi: 10.1145/3457607.Google Scholar
Figure 0

Figure 1. F508del mutation by racial group in the US Cystic Fibrosis Foundation Patient Registry analysis cohort.

Figure 1

Table 1. Counts and averages of each predictor with 95% confidence intervals among the racial groups in the US Cystic Fibrosis Foundation Patient Registry analysis cohort

Figure 2

Figure 2. Area under the receiver operating characteristic (ROC) curve (AUC) for the 3-, 6-, and 12-month prediction horizons by racial and ethnic group. Overall AUC is indicated by the horizontal line. Group-specific AUC and their respective 95% confidence interval are displayed as points and vertical lines, respectively.

Figure 3

Figure 3. Optimal sensitivities and specificities by racial and ethnic group achieved by the precision medicine algorithm for 3-, 6-, and 12-month exacerbation prediction. The average optimal sensitivity (0.623) and average optimal specificity (0.648) are indicated by the horizontal lines.

Supplementary material: File

Colegate et al. supplementary material

Colegate et al. supplementary material
Download Colegate et al. supplementary material(File)
File 35.4 KB