Establishing a Data Science Unit in an Academic Medical Center: An Illustrative Model

Abstract

The field of data science has great potential to address critical questions relevant for academic medical centers. Data science initiatives are consequently being established within academic medicine. At the cornerstone of such initiatives are scientists who practice data science. These scientists include biostatisticians, clinical informaticians, database and software developers, computational scientists, and computational biologists. Too often, however, those involved in the practice of data science are viewed by academic leadership as providing a noncomplex service to facilitate research and further the careers of other academic faculty. The authors contend that the success of data science initiatives relies heavily on the understanding that the practice of data science is a critical intellectual contribution to the overall science conducted at an academic medical center. Further, careful thought by academic leadership is needed for allocation of resources devoted to the practice of data science. At the Stanford University School of Medicine, the authors have developed an innovative model for a data science collaboratory based on 4 fundamental elements: an emphasis on collaboration over consultation; a subscription-based funding mechanism that reflects commitment by key partners; an investment in the career development of faculty who practice data science; and a strong educational component for data science members in team science and for clinical and translational investigators in data science. As data science becomes increasingly essential to learning health systems, centers that specialize in the practice of data science are a critical component of the research infrastructure and intellectual environment of academic medical centers.

The potential of data science to address essential questions is being recognized by academic institutions, prompting the investment of resources into establishing numerous data science initiatives.¹ For example, in 2015, the University of Michigan announced its investment of $100 million in a data science initiative because “progress in … fields ranging from medicine to transportation relies critically on the ability to gather, store, search, and analyze ‘big data’—collections of information so vast and complex that they challenge traditional approaches to data processing and analysis.”² In 2017, Harvard University established the Harvard Data Science Initiative with the intention that the investment would lead to “advances that will create solutions to the world’s most important challenges.” Accompanying these initiatives is an increasing demand for data scientists to be integrated into academic campuses.³

While such initiatives have renewed ongoing debates about how the field of data science is evolving (see, for example, work by Hulsen et al,⁴ Donoho,⁵ and Tukey⁶), there is consensus within academic medicine on the critical need to design and conduct studies that are robust to the replication crisis and scientifically rigorous. This imperative further necessitates the collection, storage, management, processing, analysis, and interpretation of data. In this article, we refer to all of these activities as falling under the umbrella of the practice of data science and consider them essential to addressing questions of critical import for an academic medical center. For example, nephrologists may wish to know whether findings from a recent clinical trial apply to their patients. Infectious disease physicians may want to know the ideal therapy for patients newly diagnosed with HIV who have a history of cardiovascular disease. Pediatricians may wish to develop lasting interventions that reduce obesity in children at high risk for type 2 diabetes. Such gaps in knowledge and practice may be informed by a variety of data resources. These resources include data generated by the academic medical center’s own health care system; publicly available data such as the National Health and Nutrition Examination Survey (NHANES) made accessible by the Centers for Disease Control and Prevention⁷; privately obtained data resources such as OptumInsight or IBM MarketScan^8,9; new clinical trials that recruit patients from multiple health care systems; or national registries.

While biomedical questions are typically identified by clinical and translational investigators (and increasingly posed by data scientists themselves), scientists who practice data science are at the cornerstone of addressing such questions. These scientists include biostatisticians, clinical informaticians, database and software developers, computational scientists, and computational biologists. Too often, however, those involved in the practice of data science are viewed by academic leadership as providing a noncomplex service to facilitate research and further the careers of other academic faculty. For academic medical centers to thrive in their missions of providing outstanding care, they need to rely on the collection and dissemination of the highest quality of evidence. Producing such evidence requires that the practice of data science be taken seriously and its involving activities treated as intellectual academic pursuits. This requirement includes investing in faculty who practice data science and who can train the next generation of practicing data scientists.

Investment in data science faculty and resources requires prioritization and strategic thinking. We are not the first to discuss models or guidelines on how to establish data science entities within academic medical centers. For example, Welty et al¹⁰ touch upon the importance of centralization of quantitative resources within academic medical centers. Perkins et al¹¹ stress the importance of a diverse funding portfolio for the sustained financial health of data science units and discuss how to prioritize effort across varied incoming projects. Mazumdar et al¹² provide a set of guidelines to measure intellectual contributions of data scientists to address the retention and promotion of academic team scientists. We build upon this excellent body of literature. In this article, we describe a novel model for building essential research infrastructure for the academic medical center through the practice of data science as an academic field and an investment in the careers of faculty who practice data science.

Four Elements of the Stanford Quantitative Sciences Unit Model

The Stanford Quantitative Sciences Unit (QSU) was founded in 2009 in the Department of Medicine—one of 30 departments within the School of Medicine—because of an unmet need identified by the chair of the Department of Medicine at the time. This data science collaboratory grew to meet the needs of the Department of Medicine by diversifying the type of quantitative staff and skill sets to include classical biostatisticians, clinical informaticians, epidemiologists, and software and database developers. Through continued support and advocacy by the current chair of the Department of Medicine, the QSU has expanded its engagement to other departments within the School of Medicine through partnerships described in greater detail below.

The QSU has relied on 4 key elements: a collaborative philosophy; an annual subscription funding mechanism that develops partnerships between the QSU and a clinical or basic science entity; a joint investment in the career development of faculty who practice data science; and training of faculty and staff in the practice of data science and the science behind team science.

Collaborative philosophy

While consultation has an important place in academic medical centers, many have long recognized the potentially larger impact of a deeper level of engagement through collaboration between data scientists and clinical and basic scientists.¹¹ The fundamental difference between consultation and collaboration is that the former implies that one party (consultant) aids in achieving the other party’s (consultee’s) goals, whereas the latter involves a synergistic effort to achieve jointly defined goals.¹¹ We assert that faculty in academic medicine should rely heavily on the collaboration framework. Such a model benefits the quality of the science and the career development of both the data and nondata scientists.

The QSU model is therefore one of collaboration. QSU members actively participate on teams with clinical and basic science investigators to gain a strong understanding of the biomedical context so they can be most impactful. Table 1 presents an example of a full cycle of a project in which QSU investigators are integrated into the research team as peer investigators and involved from project inception through dissemination of findings.

Table 1.

Illustration of the Cycle of a Project From Design to Dissemination Through Collaborative Engagement, With Data Scientists From the Quantitative Sciences Unit, Stanford University School of Medicine, Involved in All Steps

1 Step	Description	Illustrative example
1. Refine scope	Refine study question Identify data sources Identify expertise needed	Original question: Does early vs late adoption of continuous glucose monitors improve management of T1D? Refined question: Is early vs late adoption of monitors more likely to achieve sustained control of HbA1c starting at 6 months post diagnosis? Data source: EHRs Expertise identified by QSU: EHR informatics expert
2. Design study	Define outcome Define target population Set accrual and follow-up Identify study type Derive analytic methods and develop statistical analysis plan	Outcome: HbA1c repeatedly measured Target population: Children newly diagnosed from 2017 to 2019 and treated at Stanford Health Care Accrual period: 2017–2019 Follow-up period: minimum 1 year Study type: observational Analytic methods: regression techniques with parameters to represent sustained control that incorporate propensity score-based weighting to account for confounders
3. Analyze data	Extract data Perform analysis Interpret findings	Extract data: QSU data scientists and EHR informatics experts derive algorithm for capturing newly diagnosed T1D patients from 2017 to 2019 with International Classification of Diseases codes and apply exclusion criteria; HbA1c data are captured for eligible patients from 2017 to 2020 Perform analysis: QSU data scientists and EHR informatics experts format the data to fit the proposed analysis Interpret findings: QSU data scientists fit proposed model and perform data visualization in preparation for presentation and interpretation
4. Disseminate findings	Disseminate findings and code	The entire team prepares the manuscript for submission to peer-reviewed journal; QSU data scientists archive code from data creation and analysis for purposes of sharing and reproducibility

Abbreviations: T1D, type 1 diabetes; HbA_1c, hemoglobin A_1c; EHR, electronic health record; QSU, Quantitative Sciences Unit.

As consultants in the preaward stage, data scientists often limit their efforts to developing the statistical section for the proposal with a focus on justifying the sample size. In contrast, as collaborators, data scientists are involved in all aspects of the proposal: to frame attainable questions, to provide input on enrollment and recruitment strategies, to direct the flow of data, and to design the data monitoring and statistical analysis plans. This level of participation yields higher-quality proposals and ensures successful study execution.

Similarly, in the execution stage, consulting data scientists may get involved when data are ready for analysis. Our experience points to serious flaws with this approach. For example, understanding how recruitment and retention approaches have evolved in response to issues that arise throughout the study period can prompt new plans in the analysis to ensure a desirable interpretation of findings. To illustrate, numerous challenges were encountered while monitoring recruitment for the Apple Heart Study, including one that involved linking the data elements to the same individual, requiring QSU data scientists to develop and validate a novel algorithm to maintain integrity of the data while the study was ongoing.^13,14 Identification of these issues would not have occurred without this collaborative approach. Moreover, the solution involved continuous iterative engagement with the entire team before the formal analysis took place, demonstrating the importance of including data scientists beyond study design and data analysis.

Partnership subscription

One of the biggest challenges to implementing a collaborative model involves resources. Whereas a consultative model may enable a broader outreach to a larger number of investigators (albeit limited to a certain number of hours per investigator or project), the deeper engagement of the collaborative model likely requires more resources. This requirement is particularly true in the presence of postaward funds, where the commitment of effort required to address the specific aims of projects decreases capacity to devote toward nonsponsored activities, especially preaward exploratory work and the initiation of new collaborations. The collaborative framework therefore requires growth of the unit to address both postaward and continued preaward needs. When the Department of Medicine established the QSU, it prioritized a return on investment in the form of awards sponsored by the National Institutes of Health (NIH) and consequently provided continued resources (i.e., discretionary funds) earmarked for preaward activity. Thus, the direction of a data science unit is driven by the amount of resources available, allocation of resources, and how and whether it can grow to continue to accommodate needs.

Expanding beyond the Department of Medicine has been enabled through partnerships with leaders willing to commit resources for similar arrangements, which then allows members of the leader’s group to collaborate with the QSU. The arrangement involves an annual subscription fund, adoption of the collaborative philosophy, and commitment to the career development of a practicing data science faculty member.

The subscription covers basic infrastructure costs, the career development of a faculty member, and preaward and nonsponsored activities that address priorities identified by the chair. The latter includes partial salary of staff scientists to engage in initiatives such as establishing a data source for investigators, creating a short course for fellows or residents, or preliminary analyses leading to a published manuscript or grant proposal. A start-up package for the new data science faculty member is typically included. Prioritization of the funds is tailored to the needs described by the partnering chair.

Adoption of the collaborative philosophy is critical because in most cases, successful partnership involves a shift in culture to engage data scientists as peer collaborators. The funding model enables this thinking because there is no request for a fee for “services.” The leader of the partnering entity can further facilitate this cultural change by joining QSU members in educating their faculty members on how to engage data scientists effectively. In addition, a partnership involves a commitment by both the partnering chair and QSU director to invest in the hiring and career development of a data science faculty member who will spearhead the partnership.

Investment in faculty who practice data science

In addition to outstanding staff, a key element of the QSU model has been the inclusion of faculty members. Such faculty are charged with (1) establishing their own research programs and (2) enhancing the research infrastructure of the partnering entity. The latter is partially achieved through hands-on deep collaboration with partner-affiliated investigators and by tracking and delegating collaborations so that expertise needs are fully understood. While a balance between the 2 charges is required, in practice, the collaborative work that builds research infrastructure adds to and shapes how the faculty member’s lab evolves. Importantly, the data science faculty members serve as leaders in the practice of data science and contribute to the training of the next generation of practicing data scientists. To that end, QSU faculty mentor scientific staff and trainees, lead development of new methods, and identify optimal methods to address research questions and are therefore essential to the mission of the academic medical center. QSU faculty are also charged with putting the data science for the partnering department (e.g., neuro-analytics) on the map, an appealing idea for many department heads.

Most departments in a school of medicine do not have the appropriate infrastructure to nurture the career of a data scientist. Trajectories for promotion of team scientists and practicing data scientists are often distinct from those of clinical and basic scientists. At Stanford, the professorial line chosen for the data science faculty is one that requires demonstration of excellence in scholarship that includes fulfillment of programmatic need.¹² The partnership involves a commitment by the QSU director and partnering lead to jointly devise a strategic plan for the faculty member’s career development that includes assembling an appropriate mentoring team. The mentorship team includes a senior practicing data scientist faculty member to advise on strategies for building promotable portfolios. Criteria for promotion of faculty who practice data science—a topic discussed at great length in the team science literature¹²—typically requires a mix of (1) data science investigator-led initiatives to demonstrate leadership and (2) fulfillment of programmatic need by directing the data science component within collaborations. Thus, mentorship should involve direction on which collaborative projects may be particularly career building, fulfilling the demonstration of both the data science faculty as leaders and the collaborative needs of the partnering entity. Collaborative projects may naturally present opportunities to the data science faculty to identify gaps in methodology and to fill those gaps. Mentoring on appropriate conferences to attend and networking opportunities within and outside the institution also facilitate career advancement. A unique advantage for QSU faculty is access to the existing infrastructure of the QSU, which allows the faculty members to initiate their own labs by engaging QSU staff members before obtaining external funds. Once a faculty member’s lab is established with external support, postdoctoral fellows and graduate students can join the QSU cohort as part of the lab, an addition that is beneficial to both the new trainees and the QSU.

Education in team science methods and in the practice of data science

Building a successful team and engaging team members across different disciplines has led to an entire field unto itself, referred to as the science behind team science.¹⁵ Most investigators do not receive formal training in the practice of or the science behind team science. Along these lines, there may be a misperception on how or what the data scientist should contribute. The burden may therefore fall on the data scientist to educate other team members on the data scientist’s role. For example, an investigator may approach a data scientist with background on the goals and design of a trial and request a power calculation. In this case, the science has already been developed: The research question has been formed, the study has been designed, and the analysis plan has been developed. The investigator is asking for a stamp of approval from the data scientist through a sample size justification. The role of the data scientist, however, should be to ensure that the data generated from the trial will indeed address the set of questions of interest. This effort may require reformulating the questions. Further, the data scientist may derive a more optimal design or analytic plan that more efficiently models the data. The power calculation must reflect these scientific decisions, in which the data scientist should be involved. Moreover, educating nondata science investigators requires training for the data scientists in the science behind team science.

Additionally, education of the data scientists on new state-of-the-art methods for the design and analysis of complex studies is critical for their growth and effectiveness, especially in the ever-evolving landscape of biomedical research.

To address such needs, the QSU provides a forum for ongoing learning in both arenas of team and data science in weekly 2-hour brown-bag meetings. In this forum, QSU members meet to troubleshoot how to best communicate technical ideas and influence the research team, to discuss approaches to complex study designs and analyses, and to learn new data science methods in a journal club-style format. Although the primary purpose is to ensure scientific integrity, part of the training includes reviewing each project involving a QSU member. Reviewing members’ work can occur in 1 of 2 ways. Members can either present the statistical analysis plan before the study is conducted or the manuscript prepared for submission can be critiqued independently by rotating reviewers, who present their assessment to the larger group for discussion. The primary reviewer is a junior member who leads the presentation, with the senior member serving as a secondary reviewer to provide guidance for the junior. Thus, juniors learn how to critically assess a study and QSU community members learn new approaches from one another. Reviews are also one of several ways in which senior QSU members receive training in how to mentor junior QSU members.

Mentorship of collaborators in research methods is often a priority of the partnership and conducted through one-on-one meetings and short courses tailored for the needs of the entity. The QSU Research Methods Seminar Series is a monthly forum developed to interactively present research methods to collaborators for this purpose as well.

Illustrations of Typical Projects

The QSU has participated in addressing questions that involve evaluation of the quality of care delivered within the Stanford Health Care system. One example involved investigating whether a new protocol could reduce the incidence of delirium in hospitalized patients at Stanford. The protocol was developed by a task force established at Stanford and standardized for implementation in 18 medical-surgical units to prevent delirium, a condition that affects 10% to 64% of hospitalized patients, with a large proportion of these cases considered preventable. Through assessment of a large number of hospital encounters within Stanford Health Care between November 2013 and January 2018 (N = 105,455), the authors concluded that the protocol demonstrated desired and sustained effects.¹⁶ Such studies that evaluate the impact of delivery of care involve not only skills in handling large volumes of data extracted from a combination of sources generated within the health care system not typically designed for research but also careful thought in the study design. Often, a design that includes randomization is not feasible. In this example, the protocol was rolled out throughout the entire hospital and randomization at either the patient or provider level was not considered because of anticipated contamination (i.e., providers randomized would influence those not randomized). As is common in quality improvement evaluations, assessment of the effect of the protocol had to rely on quasi-experimental methods, which can be imperfect because issues of bias and confounding are likely to be present.¹⁷ Careful thought, therefore, needs to be devoted to both the design and analysis of quality improvement evaluations to mitigate such issues.

Other examples of studies that rely heavily on the QSU infrastructure include the previously mentioned Apple Heart Study,^13,14 for which the QSU served as the data coordinating center. The work involved designing a novel pragmatic site-less study that assessed the performance of a new application on the Apple Watch to notify participants of possible atrial fibrillation. With more than 400,000 participants enrolled, the study involved tackling profound data science challenges. These challenges included (1) uniquely identifying participants for the purposes of linking data elements in the presence of issues that led to the generation of multiple identifiers for a participant (e.g., updates to the app) and (2) appropriately syncing the timing of data generated from the gold standard electrocardiogram monitor to that generated from the app on the Apple Watch. The study required a data science team of 7 from the QSU to manage, process, and analyze the data, in addition to a core team of 4 from Stanford Research Information Technology to lay down the pipelines through which data could flow. The study demonstrated that Stanford has the data science capabilities to design complex and novel studies, to handle large volumes of diverse data streams, and to analyze noisy data and interpret findings in an expedited manner that may be expected in Silicon Valley and similar environments. Thus, a resource such as the QSU can position an academic medical center to collaborate with industry, enabling innovative research that neither academia nor industry could do alone.

QSU in Numbers

Since its founding in the Department of Medicine in 2009, the QSU has established partnerships with 10 other entities (6 departments and 4 institutes) across the School of Medicine. While still growing to meet demands, the QSU currently houses 5 faculty members hired specifically to spearhead partnerships, approximately 18 scientific staff (PhD level and master’s degree level), and 4 administrative staff, in addition to postdoctoral fellows and graduate students of the respective faculty. In 2019, the QSU engaged investigators on 367 projects and 151 grant proposal submissions, with 47 awards established, bringing in $4.6 million specifically for the QSU (corresponding to $3.3 million in direct costs and $1.3 million in indirect costs).

Partnerships with departments and institutions have led to engagement of investigators with primary appointments that span 23 basic and clinical science departments across the School of Medicine.

Concluding Remarks

Data science initiatives are being established on academic campuses to solve important problems ranging from sustainable transportation to precision medicine. We focus on addressing data science needs relevant to an academic medical center. The QSU was developed on 4 fundamental elements to accommodate the current environment of the institution. Underpinning this model is the philosophy that faculty who practice data science are needed as leaders and that the practice should be viewed as intellectual and is deserving of its own field. As such, practicing faculty need to be recruited, nurtured, and promoted with career portfolios that reflect an emphasis on team science. These may look different from those belonging to data science faculty in basic science departments, calling for a diversity in career pathways—between those who primarily practice and those who primarily develop methods—that is critical to achieving the mission of an academic medical center.

The QSU model encompasses several novelties. Its infrastructure allows QSU faculty members to initiate their careers by engaging scientific staff to launch their own research lab before obtaining external funds. Once the faculty member is established with external support, additional postdoctoral fellows and graduate students can join the QSU cohort as part of the faculty member’s lab. Thus, the QSU can be a resource not just for the partnering entity but also for the new faculty recruit, expediting the lab’s establishment. Key to the QSU’s productivity has been the novel funding mechanism of the partnership model. Consequently, the QSU does not enter partnerships with individual investigators because the QSU’s resources are committed to its existing partners. Engaging an individual outside of the partnerships, even if the individual has postaward funds to cover personnel effort, can compromise achieving the goals specified in the partnership agreement. Importantly, the partnership represents a shared investment in a data science faculty member and recognizes the need for data science leadership at the faculty level and the practice as an intellectual field. Finally, the prioritization scheme for each partnership is tailored to the needs specified by the partnering leader so that the allocation of funds is tied to the leader’s priorities. An added benefit of the model is that the diversity of partners enables the QSU to serve as a connector of investigators from different departments to join in solving complex problems by bridging gaps in expertise.

The QSU model has allowed for a unique centralization of data science resources that can span the academic medical campus. Department leaders can leverage the existing infrastructure of the QSU by investing in a partnership as an alternative to building an in-house unit that would require considerable financial resources to launch. Mentoring of data scientists by nonquantitative leaders may not be ideal for developing promotable career paths, and variation in the quality of resources across units is highly likely with in-house units. In contrast, centralization provides a solution for streamlining high-quality engagement, mentoring data scientists, creating a rich learning environment, and establishing and implementing best practices.

We recognize there may be barriers to adopting the model described here. In particular, the model relies on adoption of a collaborative approach by the academic medical community. The current culture at many institutions, however, may view practicing data scientists as consultants only. We believe that this is an addressable barrier. To that end, data scientists need to demonstrate what can be done outside of the consulting arena by strategically engaging the research team. Training in this engagement and the science behind team science is provided by the QSU. Changes to the culture should also be accomplished with the partnering leader, who can set a collaborative tone. Another barrier to adopting the model may be hiring quantitative scientists because of the lack of competitive salaries in an academic setting. Although industry is a competitor in hiring quantitative scientists, academia can offer appealing aspects, which include diversity of projects, career development provided in a training environment, and more formal continued learning. Another limitation may be one of resources. In a smaller institution with similar data science needs, there may not be discretionary funds readily available to enable partnerships. Although it may take more time to scale up to meet the needs of the entire institution, principles can be borrowed. Key is an agreement with leadership on (1) the collaborative philosophy and (2) continued effort for preaward and nonsponsored activities. Growth in the data science unit can then occur with the granting of postaward funds. The existing infrastructure and resources of the institution will necessarily drive which gaps are unmet by the data science unit and may serve as limitations to what can be accomplished. For example, our unit is able to focus on data science expertise that includes classical biostatistics (study design and modern data analysis), clinical informatics (data management and software application development), and bioinformatics, without the need to expand into areas of accessing data within the institution’s health care system or providing the means to capture such data. For example, many of the QSU projects that rely on Stanford Health Care-generated data are further enabled through other data science resources on campus with missions that involve the access to and housing of data from Stanford Health Care. In particular, there are 2 that the QSU engages when studies involve data from Stanford Health Care: (1) the Stanford Research Informatics Center, a fee-for-service data vendor (https://med.stanford.edu/ric.html) that provides cohort capture from Stanford Health Care and (2) Stanford Research Information Technology, a fee-for-service consulting center that provides the infrastructure to enable investigators to collect and combine data (http://med.stanford.edu/researchit/about-us.html). We recognize that some institutions do not have such infrastructure, which can limit some of the activities of an academic medical center. Alternatively, data science units should be tailored to address the unmet needs of their institutions, where the lack of such resources can drive the mission of the particular data science unit.

We recognize that the QSU model is not the only answer. Models that are adopted by academic medical centers should efficiently accommodate the available resources of the institution. As such, data science resources should complement and leverage existing resources. Another example is the Duke University Biostatistics, Epidemiology, and Research Design (BERD) Methods Core (https://biostat.duke.edu/berd-methods-core). While also adopting a collaborative philosophy, the Duke BERD Methods Core builds on the university’s Clinical and Translational Science Award (CTSA) from the NIH. The NIH established the CTSA program in 2006 to provide academic medical centers the ability to establish robust research infrastructure for their faculty. Many institutions leverage their own BERD programs to enable access to data science resources for the larger academic community.¹⁸ The CTSA BERD programs vary widely in structure and missions, with some practicing a consultative model where fees are exchanged for service and others adopting a collaborative approach.¹⁹ The mission of the Duke BERD Methods Core is to meet 2 equally important goals: (1) link investigators with methodologists across Duke University and affiliated institutions, and (2) collaborate with an interdisciplinary network of clinical and translational investigators by providing expertise in study design, real-world evidence, analysis, and interpretation of results.

In contrast to the QSU’s partnership model, the financial approach of the Duke BERD Methods Core relies on funding from the following mechanisms: (1) Duke University’s Department of Biostatistics and Bioinformatics discretionary funds and the CTSA, to cover efforts involved in building teams, pursuing educational initiatives, and managing agreements with collaborative groups, and (2) other funding streams identified by collaborative groups to cover the proportion of full-time equivalent effort (i.e., not in the form of fee-for-service) to meet the research and educational needs. The latter often involves the development of agreements between the Duke BERD Methods Core and the collaborator(s). Thus, in contrast to the QSU model, collaborators of the Duke BERD Methods Core are not required to be part of a larger partnership. The Duke BERD Methods Core includes affiliated faculty from the Duke University School of Medicine Department of Biostatistics and Bioinformatics, offering an alternative approach to that of the QSU for engaging faculty; the QSU model relies on faculty hired specifically for the purpose of spearheading partnerships. Similar to the QSU model, the Duke BERD program applies a team science-based approach. For other institutions that have a collaborative mission, both the Duke and QSU models can serve as a resource for how to establish collaborative units in academic medical centers, where the former demonstrates how institutions that are CTSA recipients can build on the awards.

The real success of a data science unit in an academic medical center should be measured through its impact on public health. While we have yet to determine that impact, we instead note the number of awards granted, the number of investigators engaged, and the demand for partnerships from additional departments. Additionally, to date all QSU faculty have demonstrated the ability to obtain external awards as principal investigators, reflecting their ability to build independent labs while fulfilling programmatic need. The training environment at the QSU has also helped to elevate staff members, as evidenced by several staff members who have received training at the QSU and graduated to faculty positions at outstanding institutions. While we acknowledge there may be barriers to incorporating the model presented here, principles behind the establishment of the QSU could easily be adopted by other centers that build on its key elements. As data science becomes increasingly essential to learning health systems, centers that specialize in the practice of data science are a critical component of the research infrastructure and intellectual environment of academic medical centers.