Electronic Health Record–Based Algorithms May Improve Identification of Cases of Atypical Diabetes, Including in Groups Underrepresented in Biomedical Research

Abstract

OBJECTIVE

To determine whether electronic health record (EHR)-based algorithms can identify individuals with atypical diabetes (AD) eligible for the Rare and Atypical Diabetes Network (RADIANT) study and whether these algorithms improve identification and enrollment of individuals from populations underrepresented in biomedical research (UBR).

RESEARCH DESIGN AND METHODS

Two RADIANT clinical sites independently developed EHR-based algorithms to identify AD phenotypes, including a lean type 2 diabetes–like physiology without features of metabolic syndrome (site 1) and concurrent diagnoses of type 1 and type 2 diabetes (site 2). Demographic characteristics were compared using χ² tests between algorithm-identified individuals and RADIANT participants enrolled prior to the implementation of EHR-based algorithm recruitment.

RESULTS

Deployment of EHR-based algorithms identified 550 individuals in site 1 and 620 individuals in site 2, with 108 and 147 deemed eligible for RADIANT recruitment, respectively. Compared with prior RADIANT participants at these sites, EHR-identified, RADIANT-eligible individuals were more likely to self-identify as a race or ethnicity besides non-Hispanic White and prefer a language other than English. Of the 53 and 57 individuals who were successfully contacted at sites 1 and 2, respectively, 8 and 14 enrolled in RADIANT, representing meaningful contributions to total enrollment (21% and 22% at sites 1 and 2, respectively) compared with pre–EHR-based recruitment. However, demographics of newly enrolled individuals did not differ from traditionally recruited participants.

CONCLUSIONS

EHR-based algorithms successfully identified rare diabetes phenotypes and yielded greater representation of study-eligible individuals from UBR populations than traditional participant identification methods. However, additional strategies are needed to address persistent barriers to participation of UBR populations in AD research.

Graphical Abstract

Introduction

Diversity in research is critical to promote health equity, ensure generalizability, document the full range of disease pathophysiology, and explore heterogeneity of treatment effects. However, clinical research recruitment is challenging, particularly when the target disease is rare, inclusion and exclusion criteria are complex, or the research study is time-consuming and burdensome to participants. Recruitment of populations underrepresented in biomedical research (UBR) can be even more challenging (1–5).

Multiple systemic factors contribute to underrecruitment of UBR populations, including differences in resources (time, transportation, or financial security) and mistrust of the medical system (6–8). Lack of awareness of and access to clinical studies due to health care context can also prevent interested individuals from participating in medical research. Black and Hispanic Americans are less likely to be referred to medical subspecialists and more likely to be cared for in safety-net hospitals rather than academic medical centers (9–15). Recruitment methods such as physician referral and in-clinic advertising (e.g., flyers) preferentially target populations with access to research-focused specialists and institutions (12,15,16). These access barriers are compounded for individuals with limited literacy and/or English proficiency (17,18) who may not be able to understand research advertisements even when presented.

The Rare and Atypical Diabetes Network (RADIANT) is a multisite study seeking to identify and recruit individuals across the U.S. with nonclassical forms of diabetes, with the aim of improving clinical and genetic characterization of atypical diabetes (AD) (19). Given that AD is uncommon, heterogeneous, and difficult to define and often requires expert evaluation, RADIANT faces unique challenges to recruitment, particularly among individuals from UBR populations (20). To date, most participants have been referred either by an expert clinician (usually by a diabetes specialist at an academic medical center and often by study investigators) or research groups or registries in which the participant had previously enrolled. As individuals from UBR populations are less likely to receive subspecialty care (9–11) and less likely to participate in research (1–5), many potential candidates may be missed by these mechanisms. Alternative recruitment strategies, such as outreach to primary care providers and social media campaigns, have resulted in fewer eligible referrals, as criteria for AD are complex and not widely understood. Therefore, new recruitment approaches addressing barriers to research engagement faced by individuals from UBR populations are vital for increasing UBR participant enrollment in RADIANT.

Aiming to improve both overall and UBR population–specific recruitment, two RADIANT sites leveraged electronic health record (EHR)-based algorithms to identify individuals with specific AD phenotypes, marked either by type 2 diabetes–like physiology in lean individuals without other evidence of metabolic syndrome or mixed features of type 1 and type 2 diabetes. We performed a cross-sectional analysis of records from these EHR-based algorithm recruitment efforts and compared them with recruitment records from the prospectively maintained RADIANT study, hypothesizing that EHR-based algorithms designed for this purpose could identify previously unrecognized candidates for the RADIANT study, including individuals from UBR populations, compared with the current standard of recruitment, which is driven predominantly by specialist access.

Research Design and Methods

Setting and EHR-Identified Participants

The study was performed in two large health systems with large catchment areas and at least one quaternary referral center. RADIANT defines AD broadly as diabetes that does not conform to known forms of diabetes, including type 1, type 2, gestational, lipodystrophic, secondary, and monogenic forms (19), and each site identified a particular phenotype of AD to query in the EHR.

Site 1

Site 1 leveraged previously validated algorithms found to identify individuals with AD in a hospital-based biobank (21). Although many features may contribute to a diagnosis of AD, we selected the phenotype of lean type 2–like diabetes (not entirely insulin deficient) without metabolic syndrome. Briefly, we developed and tested these algorithms in the Mass General Brigham Biobank (∼120,000 people at the time of study design), including all individuals meeting a system-specific algorithm for type 2 diabetes (based on ICD codes, laboratory values, medication use, anthropometrics, and natural language processing) (22). We then excluded those with evidence of metabolic disease (requirement for documentation of normal BMI on at least three occasions and exclusion for any documentation of obesity by BMI, HDL levels ever <50 mg/dL, or triglyceride levels ever >150 mg/dL) or other forms of diabetes (positive for type 1 diabetes autoantibodies or diagnosis codes for cystic fibrosis). We then tested six different methods for excluding type 1 diabetes, identifying two (requirement to have received antibody testing with negative results or requirement to have never taken outpatient insulin) as the most specific for identifying our phenotype (type 2–like diabetes without features of metabolic syndrome) compared with expert chart review. Full details of this algorithm validation, including a flowchart used for systematic characterization of diabetes type, have been previously published (21).

We then deployed these two most successful algorithms in the full EHR (∼12 million individuals). Each case then underwent chart review by a trained research coordinator (V.C., C.I.F.H., M.K., M.F., or W.M.), noting personal diabetes history, family history, comorbidities, diabetes-related and other medications, anthropometrics, laboratory results, and other factors with potential relevance to diabetes classification. The case was then presented to a RADIANT site investigator (S.J.C.) who designated each case as likely AD, likely a known form of diabetes, or unable to be classified (need more information) based on the available information, including consideration of racial- and ethnic-specific BMI cutoffs (23,24). Individuals with likely AD were invited to enroll in RADIANT stage 1, and individuals needing more information were invited to enroll if the needed information was likely to be obtainable (e.g., islet autoantibody status, age at diabetes diagnosis) during the enrollment process.

Site 2

Site 2 adapted three published algorithms to identify individuals with diabetes in EHR data within the University of Colorado Health (UCHealth) system, as previously described (25). These algorithms were applied to EHR data through August 2020 to identify individuals with any form of diabetes (algorithm 1, positive predictive value [PPV 99%]) (26), type 2 diabetes (algorithm 2, PPV 98%) (27), and type 1 diabetes (algorithm 3, PPV 94.5%) (28,29), each relying on ICD codes, laboratory values, medication use, and clinical notes. Among those identified as having diabetes by algorithm 1, we then applied algorithm 2 and algorithm 3. We used iterative manual chart review of individuals identified as having any diabetes (algorithm 1 only), type 2 diabetes (meeting criteria of algorithm 1 and algorithm 2), and type 1 diabetes (meeting criteria of algorithm 1 and algorithm 3) to optimize the algorithms’ performance in our health system (25). During this process, we prioritized sensitivity for all-cause diabetes and specificity for type 1 and type 2 diabetes computable phenotypes. For example, we found that relaxing two of the published criteria for type 2 diabetes (27) improved sensitivity of identifying individuals with type 2 diabetes without compromising specificity. First, we permitted equal minimum dates for ordering of insulin and a noninsulin glucose-lowering medication rather than requiring that insulin use occur subsequent to noninsulin glucose-lowering medication (25). Second, individuals with nonmajority use of diagnosis codes specific for type 1 diabetes remained eligible for classification as type 2 diabetes, whereas the published algorithm for type 2 diabetes disqualified individuals with any use of a type 1 diabetes diagnosis code (25). Final manual chart review validation found that the locally optimized algorithms for type 1 and type 2 diabetes had 100% and 99% specificity, respectively, and a PPV of 100% and 96%, respectively (30).

Despite this algorithm optimization, we found 620 individuals who met the highly specific algorithm criteria for both type 1 and type 2 diabetes. We considered these individuals as potentially having AD based on their indistinguishable computable phenotyping. Each case underwent manual chart review by a board-certified endocrinologist (M.G.) with attention to personal and family medical history, medication use (including diabetes medications and medications known to promote the development of diabetes), substance use, comorbidities, weight trajectory, laboratory results, and any other data pertinent to their diabetes history. Charts were determined to be likely AD, likely known type of diabetes, or unable to be classified based on available data (needed more information). When diagnostic uncertainty remained, the case was reviewed with the RADIANT site investigator (N.R.), and final classification was determined by consensus.

Recruitment Strategy

The RADIANT study (ClinicalTrials.gov identifier NCT05544266) includes the following three stages of participation: stage 1 comprises an online questionnaire regarding diabetes-related personal and family history and a blood collection for islet autoantibodies; stage 2 includes a blood collection for whole-genome sequencing; and stage 3 comprises a battery of physiologic testing, including an oral glucose tolerance test, physical function measures, health-related questionnaires, an in-depth history and physical examination, and photography (19). Because RADIANT referrals may have complex phenotypic presentations with some elements that are or are not consistent with known forms of diabetes, an adjudication committee of clinical and genetics experts reviews each RADIANT referral prior to progression to stage 2 of the study. Stages 1 and 2 may be completed entirely remotely, with virtual consent processes and sample collection by a commercial laboratory, or may be completed in person at a RADIANT study site. Stage 2 may alternatively be completed concurrently with stage 3, which requires an in-person, all-day visit to a RADIANT site. Participants are compensated for their time and travel expenses as needed for this stage. All informed consent and questionnaires are available in English and Spanish, and interpretation services are available as needed. Because improving representation of participants from UBR populations in the RADIANT cohort was a key goal of this study, candidates identified through the EHR received enhanced support for enrollment compared with standard recruitment procedures.

At site 1, recruitment materials were redesigned to support individuals with low health literacy and to emphasize participant safety and autonomy. This site implemented more intensive contact strategies, including outreach via patient portal messaging, phone calls, and mailed letters, and partnered with trusted physicians to engage participants with unstable housing, inconsistent phone or internet access, or limited prior engagement with the health center. Additional measures included extended coordinator availability (including evenings and weekends) to facilitate completion of stages 1 and 2, as well as home visits and compensation for time and travel at all stages.

At site 2, study coordinators contacted eligible individuals who had received care at the UCHealth endocrinology clinic using the EHR patient portal, followed by phone calls to assess interest and provide a comprehensive explanation of study procedures. The corresponding endocrinology providers were notified of their patients’ eligibility to help facilitate recruitment and address any clinical questions. For eligible individuals who were not receiving care in the endocrinology clinic, the corresponding primary care provider was contacted first to obtain permission to reach out. If no objection was received, study coordinators then contacted the candidate via the patient portal and, subsequently, by phone to share study information. Additional support was provided throughout the enrollment process, including assistance with electronic consenting, help completing online questionnaires, and, when needed, options for onsite enrollment to facilitate participation.

Statistical Analysis

We report baseline characteristics using means and SDs for continuous measures and number and proportion for categorical measures. We used χ² tests to examine differences in sex, race or ethnicity, and preferred language between individuals at subsequent stages of recruitment (e.g., those identified by the EHR-based algorithm and those deemed eligible by the investigator) and between individuals enrolled at the same site who were identified by the EHR-based algorithm or who enrolled following standard recruitment methods prior to the initiation of the EHR-based algorithms, which began in November 2023 at both sites. All statistical analyses were performed using R 4.4.2 (R Foundation for Statistical Computing). Figure 1 was generated using Python (version 3.12.12) code adapted from published methods (31). The candidate identification and recruitment strategies were approved by the institutional review boards of Mass General Brigham (Boston, MA) and the University of Utah (Salt Lake City, UT), respectively.

Figure 1.

Flow diagram of participant recruitment following identification by EHR-based algorithm vs. standard recruitment procedures at site 1 (A) and site 2 (B). Adapted with permission from Ellen et al. (31).

Data and Resource Availability

Data from RADIANT (version 1, https://doi.org/10.58020/1wxp-dw53) reported here are available by request at the National Institute of Diabetes and Digestive and Kidney Diseases Central Repository Resources for Research website (https://repository.niddk.nih.gov/).

Results

Site 1 Recruitment Results

Prior to the initiation of EHR-based recruitment, site 1 had enrolled 39 individuals (31% female; 69% non-Hispanic White, 5.1% non-Hispanic Black, and 20.5% non-Hispanic Asian; 95% English-speaking individuals), of whom 31 had scheduled or completed stage 2 procedures. Physician referral was the most common recruitment method, followed by public advertisement of the study (e.g., flyers, social media) (Table 1 and Fig. 1A).

Table 1.

Characteristics of recruited sample prior to algorithm-based recruitment and potentially eligible sample identified through EHR-based algorithm

	Massachusetts General Hospital (lean type 2 diabetes algorithm)		UCHealth (type 1 and type 2 diabetes overlap algorithm)
	Recruited prior to algorithm (n = 39)	Potential candidates identified by EHR-based algorithm (n = 108)	Recruited prior to algorithm (n = 64)	Potential candidates identified by EHR-based algorithm (n = 147)
Age (years), mean (SD)	49 (18)	69 (9)	46 (13)	61 (14)
Female sex	12 (31.0)	69 (63.9)	31 (48.0)	59 (40.1)
Race and ethnicity
Non-Hispanic White	27 (69.0)	50 (46.3)	47 (73.0)	64 (43.5)
Non-Hispanic Black	2 (5.1)	22 (20.4)	5 (7.8)	26 (17.7)
Hispanic or Latino	2 (5.1)	9 (8.3)	0 (0)	4 (2.7)
Non-Hispanic Asian	8 (21.0)	20 (18.5)	6 (9.4)	42 (28.6)
American Indian or Alaska Native	0 (0)	0 (0)	1 (1.6)	0 (0)
Native Hawaiian or Pacific Islander	0 (0)	0 (0)	1 (1.6)	1 (0.7)
Other or more than one race	0 (0)	1 (0.9)	2 (3.1)	8 (5.4)
Unknown or prefer not to answer	0 (0)	6 (5.5)	2 (3.1)	2 (1.4)
Preferred language
English	37 (95)	87 (80.6)	64 (100)	123 (83.7)
Spanish	2 (5.1)	12 (11.1)	0 (0)	21 (14.3)
Other	0 (0)	9 (8.3)	0 (0)	3 (2.0)
Referral source
Physician referral	25 (64.1)	0 (0)	49 (76.6)	0 (0)
Studies/registries	1 (2.6)	0 (0)	2 (3.1)	0 (0)
Flyer, letter, news report, social media, or family/friend	8 (20.5)	0 (0)	5 (7.8)	0 (0)
Other/unknown	5 (12.8)	0 (0)	8 (13.0)	0 (0)
EHR-based algorithm	0 (0)	108 (100)	0 (0)	147 (100)

Data are n (%) unless otherwise indicated.

At site 1, 550 individuals were identified by the EHR-based algorithms, with 108 (19.6%) deemed eligible based on likely diagnosis of AD, and 86 (15.6%) deemed possible to recruit (e.g., excluding those who were deceased or lost to follow-up) (Fig. 1A). In step-by-step comparisons, there were no significant differences in race and ethnicity, sex, or preferred language at any recruitment step following identification of potentially eligible individuals (prior to exclusion 1 in Fig. 1A); however, there was a nonsignificant trend toward decreased enrollment of individuals who expressed interest in the study but who identified as a race or ethnicity other than non-Hispanic White (vs. non-Hispanic White; exclusion 4, P = 0.21). Compared with participants who enrolled in stage 1 prior to the adoption of algorithm-based recruitment, those deemed eligible and those deemed possible to recruit were more likely to be female, identify as a race or ethnicity other than non-Hispanic White (with the greatest increase in non-Hispanic Black individuals), and prefer a language other than English.

Of the 86 individuals deemed possible to recruit, only 53 (61.6%) were able to be contacted by any means (EHR-based messaging, phone calls, or mailed letters). Of these individuals, 27 (50.9%) expressed interest in enrollment, but only 8 (15.1%) completed stage 1 procedures. Ultimately, 4 individuals (0.7% of those reviewed, 4.7% of those for whom contact was attempted, 14.8% of those who expressed interest in enrollment, and 50% of those who completed stage 1 procedures) were deemed eligible for and completed stage 2 procedures as of 24 months following initial contact attempts (Fig. 1A).

Among participants who completed stage 1 and stage 2 procedures, respectively, recruitment via the EHR-based algorithms yielded a numerically higher percentage of female participants compared with recruitment through traditional means but with otherwise similar demographics. At site 1, these individuals represented a 20.5% and 12.9% increase in stage 1 and stage 2 enrollees, respectively, compared with prior to the initiation of algorithm-based recruitment and 25.8% and 25.0% of stage 1 and stage 2 enrollees, respectively, during the study period.

Although some participants expressed interest in additional support for participation, no participants elected to receive off-hours coordinator support, home visits, or financial support for travel for stage 1 or 2 procedures. Anecdotally, initial unsuccessful efforts to reach some individuals from UBR populations for first contact improved when site 1 implemented communication strategies focusing on synchronous person-to-person contact (usually phone calls) after standard working hours with bilingual staff. With these efforts, individuals from UBR populations were equally represented among those for whom contact was attempted, those who were reached, and those who expressed interest in joining the study (Fig. 1A).

Site 2 Recruitment Results

Prior to the initiation of EHR-based recruitment, site 2 had enrolled 64 individuals (49% female; 73% non-Hispanic White, 7.8% non-Hispanic Black, 9.4% non-Hispanic Asian, and 9.4% other race and ethnicity; 100% English-speaking individuals), of whom 45 had scheduled or completed stage 2 procedures. Physician referral was the most common recruitment method, resulting in more than three-fourths of total enrollment prior to EHR-based recruitment (Table 1 and Fig. 1B).

At site 2, 620 individuals were identified by the algorithms, with 147 (23.7%) found through expert chart review to have likely AD and 94 (15.2%) eligible for recruitment (Fig. 1B). In step-by-step comparisons, potentially eligible female candidates were less likely to be eligible for contact (i.e., more likely to be deceased or out of contact with the health system; exclusion 1), and those who preferred a language other than English were less likely to be reached due to systemic barriers (i.e., consent forms not available in target language; exclusion 2). Candidates who identified as a race or ethnicity other than non-Hispanic White were less likely to be reached (among those eligible for contact; exclusion 2, P < 0.001) or enrolled (among those who expressed interest in enrollment; exclusion 4, P = 0.03) but more likely to advance to stage 2 if enrolled (less likely to be excluded via exclusion 5, P = 0.02). Compared with participants who enrolled in stage 1 prior to the adoption of algorithm-based recruitment, those deemed eligible and those deemed possible to recruit were less likely to be female but more likely to identify as a race or ethnicity other than non-Hispanic White (with the greatest increase in non-Hispanic Black and Hispanic or Latino individuals) or prefer a language other than English.

Of the 94 individuals found eligible, 57 (60.6%) were successfully contacted (EHR-based messaging, phone calls, or mailed letters). Of these individuals, 27 (47.4%) expressed interest in enrollment, but only 14 (24.6%) completed stage 1. Despite extensive recruitment efforts, only nine participants (1.5% of those reviewed, 9.6% of those for whom contact was attempted, 33.3% of those who expressed interest in enrollment, and 64.3% of those who completed stage 1 procedures) completed or scheduled stage 2 procedures as of 24 months following initial contact attempts (Fig. 1B), with demographic characteristics similar to participants recruited through traditional means.

At site 2, these individuals represented a 21.9% and 20.0% increase in stage 1 and stage 2 enrollees, respectively, compared with prior to the initiation of algorithm-based recruitment and 32.6% and 36.0% of stage 1 and stage 2 enrollees, respectively, during the study period. No differences in demographics were observed between individuals recruited via the EHR-based algorithms and those recruited through traditional means at either stage 1 or stage 2.

In addition to individuals identified by the EHR-based algorithm, recruitment-related procedures indirectly resulted in an increase in physician referrals to RADIANT. Particularly, increased communication between the RADIANT study teams and local endocrinologists and generalists seemed to improve both awareness of and interest in the study, with these direct contacts resulting in up to 10 additional referrals at each site.

Conclusions

In this study, we applied EHR-based algorithms to identify individuals with AD, specifically a lean type 2 diabetes–like phenotype without features of metabolic syndrome or a mixed diabetes phenotype, as a recruitment strategy for the RADIANT study. This approach increased the number and improved the racial, ethnic, and linguistic diversity of identified candidates, particularly by increasing the representation of non-Hispanic Black and Hispanic or Spanish-speaking individuals. Although AD is a specific use case, we present proof of concept that finely curated EHR algorithms may improve identification of individuals eligible for clinical research both overall and among UBR populations, particularly in studies of rare or difficult-to-define disease. However, even the provision of additional support and resources for participation did not significantly increase UBR enrollment.

Uncommon or difficult-to-identify phenotypes may pose challenges for clinical trial recruitment. The granularity of EHR data may allow increased identification of eligible individuals but often at the cost of extensive chart review burden. Our EHR-based algorithms build upon established efforts, leveraging data to augment and streamline clinical study recruitment by prioritizing individuals with a higher likelihood of a true rare phenotype. Similarly, more complex machine learning algorithms based on both structured and unstructured data successfully identified patients eligible for a rheumatoid arthritis study, decreasing chart review burden while minimizing exclusion of eligible individuals (32). An algorithm deployed in the Veterans Affairs system successfully supported recruitment for a diuretic-related study, with high rates of enrollment among UBR groups, including rural and socioeconomically disadvantaged groups (33).

Several preliminary studies used EHR data to identify patients with AD, primarily for the RADIANT study. Two of these, including the site 1 algorithm validation, focused on a lean type 2 diabetes–like phenotype with either insulin insufficiency or evidence of insulin resistance without other features of metabolic syndrome (21,34); both found that EHR-based algorithms could successfully identify a population enriched for AD. Another trained a clustering algorithm among expert- classified individuals from a well-phenotyped Hispanic cohort, then applied the algorithm in a pediatric cohort, identifying two phenotypes of AD in children (antibody negative, lean, and with low β-cell function; antibody positive, higher BMI, and at risk for diabetic ketoacidosis) (35). Finally, another study identified pediatric AD case individuals using both clinician-completed questionnaires and EHR-based queries to identify children fitting three atypical pediatric phenotypes (unknown diabetes type, type 2 diabetes onset before age 10 years, and antibody-negative type 1 diabetes) (36), identifying AD in 1–6.6% of the cohort, with high representation of UBR individuals. Although these EHR-based algorithms have successfully identified patients with AD, including in UBR populations, previous studies have not quantified the effect of these strategies on clinical study enrollment or directly compared UBR participant representation between EHR algorithm–based methods and standard recruitment strategies.

The burden of diabetes falls disproportionately on minoritized communities, including non-Hispanic Black, Hispanic or Latino, non-Hispanic Asian, and American Indian or Alaska Native communities (37). Despite this, minoritized groups are underrepresented in clinical diabetes research (5). A constellation of factors may contribute to this, including reduced access to health care specialists, lack of awareness of research, limited time and resources to facilitate participation, and mistrust of the medical and research establishment (6–8). For rare diseases studies such as RADIANT, expert referrals play a crucial recruitment role due to complex eligibility criteria and awareness of clinical study availability, leading to underrepresentation of populations without access to specialist care. Developing EHR-based algorithms may help to overcome this barrier to identification of UBR research participants, particularly in studies of rare diseases in which retrospective identification of individuals may be more powerful than a prospective search. However, even these more objective measures for identifying candidates may still be susceptible to bias by excluding those who are less represented in EHR data, such as individuals with unstable insurance coverage or housing who may seek care sporadically and through different health systems.

Resource limitations represent another critical barrier to research participation. Lack of transportation, competing vocational and caregiving responsibilities, and financial costs associated with participating in research often prevent interested participants from enrolling in clinical studies (6,7). Although we implemented multiple measures to support participants with resource constraints, our candidate participants did not ultimately engage with these measures. By contrast, flexibility of timing (especially after-hours) and means of initial communication about the study (primarily multiple phone numbers) allowed us to reach more candidates than standard recruitment procedures (EHR-based messaging, letters, and daytime phone calls). Finally, mistrust of medical and research institutions remains a barrier to participation in research, particularly in the African American community in which numerous egregious abuses have occurred. Limited English proficiency and literacy also restrict participation. We were unable to assess the effect of our efforts in this space (language-concordant communication with an emphasis on participant autonomy) in the scope of this study.

This study had several strengths, including the application of previously validated algorithms and follow-up from identification through enrollment in a clinical study. However, it was also limited in its generalizability and scalability, having been performed at only two academic centers and relying on center-specific, and in one case proprietary, diabetes algorithms. Furthermore, our study focused on only two presentations of AD, and as such, our findings related to diversity of recruitment may not apply to studies involving other rare phenotypes. Individuals with limited EHR information (presumably including many with limited access to health care) were more likely to have missing information, such that health care access continued to limit trial accessibility to some degree. Finally, data were not available on the number of individuals who were referred to RADIANT through traditional channels but did not enroll; thus, we were only able to compare the number of algorithm-identified individuals who enrolled with those who enrolled through traditional means rather than comparing all who were referred.

In conclusion, EHR-based algorithms identified individuals with a rare phenotype—specifically, atypical forms of diabetes—and increased identification of potential candidates for the RADIANT study, providing a proof of concept that EHR-based algorithms may aid recruitment to studies of rare diseases. Although this method improves the efficiency of candidate identification compared with EHR-wide chart review, it remains labor-intensive, may not be practical for all studies, and resulted in small absolute increases in recruitment. However, this method improved racial and ethnic diversity of identified candidates compared with standard identification and recruitment tools, thus improving equity in eligibility for research participation and our understanding of AD presentations in diverse populations.

Funding Statement

The RADIANT study is funded by National Institute of Diabetes and Digestive and Kidney Diseases grants U54 DK118638 and U54 DK118612. J.G. was supported by a U.S. Department of Veterans Affairs Health Systems Research Postdoctoral Fellowship. S.R. was supported by National Institutes of Health award P30 DK116073 and U.S. Department of Veterans Affairs award I01-BX006417. S.J.C. was supported by American Diabetes Association grant 7-21-JDFM-005.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.