Hostname: page-component-8448b6f56d-jr42d Total loading time: 0 Render date: 2024-04-17T02:29:23.258Z Has data issue: false hasContentIssue false

Lessons learned: Development of COVID-19 clinical staging models at a large urban research institution

Published online by Cambridge University Press:  27 March 2023

Sean S. Huang
Department of Medicine, Division of Academic Internal Medicine and Geriatrics, University of Illinois Chicago, Chicago, IL, USA
Lelia H. Chaisson
Department of Medicine, Division of Infectious Diseases, University of Illinois Chicago, Chicago, IL, USA Center for Global Health, University of Illinois Chicago, Chicago, IL, USA
William Galanter
Department of Medicine, Division of Academic Internal Medicine and Geriatrics, University of Illinois Chicago, Chicago, IL, USA Department of Pharmacy Systems, Outcomes, and Policy, University of Illinois Chicago, Chicago, IL, USA
Arash Jalali
School of Public Health, University of Illinois Chicago, Chicago, IL, USA
Martha Menchaca
Department of Radiology, University of Illinois Chicago, Chicago, IL, USA
Natalie Parde
Department of Computer Science, University of Illinois Chicago, Chicago, IL, USA
Jorge M. Rodríguez-Fernández
Department of Neurology and Rehabilitation, University of Illinois Chicago, Chicago, IL, USA
Andrew Trotter
Department of Medicine, Division of Infectious Diseases, University of Illinois Chicago, Chicago, IL, USA
Karl M. Kochendorfer*
Department of Family and Community Medicine, University of Illinois Chicago, Chicago, IL, USA
Corresponding author: K. M. Kochendorfer, Email:
Rights & Permissions [Opens in a new window]



The University of Illinois at Chicago (UIC), along with many academic institutions worldwide, made significant efforts to address the many challenges presented during the COVID-19 pandemic by developing clinical staging and predictive models. Data from patients with a clinical encounter at UIC from July 1, 2019 to March 30, 2022 were abstracted from the electronic health record and stored in the UIC Center for Clinical and Translational Science Clinical Research Data Warehouse, prior to data analysis. While we saw some success, there were many failures along the way. For this paper, we wanted to discuss some of these obstacles and many of the lessons learned from the journey.


Principle investigators, research staff, and other project team members were invited to complete an anonymous Qualtrics survey to reflect on the project. The survey included open-ended questions centering on participants’ opinions about the project, including whether project goals were met, project successes, project failures, and areas that could have been improved. We then identified themes among the results.


Nine project team members (out of 30 members contacted) completed the survey. The responders were anonymous. The survey responses were grouped into four key themes: Collaboration, Infrastructure, Data Acquisition/Validation, and Model Building.


Through our COVID-19 research efforts, the team learned about our strengths and deficiencies. We continue to work to improve our research and data translation capabilities.

Research Article
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 (, which permits unrestricted re-use, distribution and reproduction, provided the original article is properly cited.
© The Author(s), 2023. Published by Cambridge University Press on behalf of The Association for Clinical and Translational Science


The University of Illinois at Chicago (UIC), along with many academic institutions worldwide, made significant efforts to address the many challenges presented by the COVID-19 pandemic. Many institutions leveraged electronic medical information [Reference Mamidi, Tran-Nguyen, Melvin and Worthey1Reference Vaid, Somani and Russak3] to develop COVID-19 risk models. Unfortunately, few models demonstrated clinical utility [Reference Syrowatka, Kuznetsova and Alsubai4Reference Wynants, Van Calster and Collins6]. Deploying novel models in hospital settings is often complicated and requires robust infrastructure and innovation [Reference Zhang, Xing, Zou and Wu7] as machine learning shifts from development to deployment. Our goal was to define a COVID-19 clinical staging system to stratify patient risk and provide an accurate prognosis based on demographics, medical history, medications, symptoms, physical exam findings, laboratory results, and imaging findings. The rationale was that this system would help our institution improve patient care and hospital preparedness.

Data from patients with a clinical encounter at UIC from July 1, 2019 to March 30, 2022 were abstracted from the University of Illinois electronic health record (EHR) and stored in the UIC Center for Clinical and Translational Science (CCTS) Clinical Research Data Warehouse (CRDW). We included patients with confirmed COVID-19, and compared them to patients without COVID-19 admitted within the same time frame. Radiology images were collected by UI Health Picture Archiving and Communication System (PACS). Analyses were conducted within a secure computing environment (SCE). A CCTS COVID-19 Rapid Response Pilot grant was obtained to support this project.

This project provided an opportunity for collaboration between physicians, researchers, informaticists, and engineers across the University of Illinois System, including the Chicago and Urbana-Champaign campuses. This represented a novel initiative. We hoped to strengthen both institutions’ informatics and data science capabilities with this collaboration. In addition, we aimed to improve our computing and data infrastructure for performing research. We also hoped to improve the agility and timeliness of our university and health system’s response to urgent clinical research questions and their direct application to a novel and evolving health problem. This capability would be critically important in a time when globalization and other factors have made emerging infectious diseases an ongoing public health issue [Reference Baker, Mahmud and Miller8].

During this project, our institution employed techniques to extract relevant information from unstructured UI Health data and gleaned many insights about COVID-19 patients hospitalized at UIC. In addition, we generated multiple predictive models, incorporated machine learning, leveraged natural language processing (NLP), and performed imaging analytics. Our work has been published in several peer-reviewed journals [Reference Pyrros, Flanders and Rodríguez-Fernández9Reference Pishgar, Harford and Theis13]. New goals and aims were developed throughout the project, including examining social determinants of health and comparing our findings with larger national and regional COVID-19 patient datasets (e.g., NC3 [Reference Sharafeldin, Bates and Song14]).

However, there were many challenges and failures along the way. For instance, we had a few rejected grants, including an NIH NLM R21 grant and at least three others centered around artificial intelligence. When we did receive funding, we often had difficulty finding approval to utilize those funds. The collaboration between UIC and UIUC could have been more seamless and required the melding of disparate systems. There were bandwidth challenges among many team members, leading to slowdowns. We also had descriptive analyses that were not able to be completed as key members transitioned to operations. An expensive SCE led to distress and unforeseen technical challenges for our research staff. There were many scope challenges as well. This paper discusses many of these obstacles and more while reflecting on the lessons we have learned.


We qualitatively explored the context, facilitators, and barriers to implementing our COVID-19 clinical staging project. Principle investigators, research staff, and other project team members were invited to complete an anonymous Qualtrics survey (Fig. 1) to reflect on the project. The survey included open-ended questions centering on participants’ opinions about the project, including whether project goals were met, project successes, project failures, and areas that could have been improved. Email invitations to participate were sent in two separate rounds over a month apart. Participation was voluntary and not compensated. Participants provided informed consent prior to completing the survey. In addition, we discussed findings with team members and key representatives of the CCTS. Finally, we conducted a content analysis of surveys and interviews and identified and extracted prominent themes.

Figure 1. Survey questions.


Nine project team members (out of 30 members contacted) completed the survey. While the responses were anonymous, the responders included principal investigators, physicians, engineering faculty, and research staff. The survey responses identified four key themes: Collaboration, Infrastructure, Data Acquisition/Validation, and Model Building.

Theme 1: Collaboration

One of the primary benefits of this project was the opportunity to foster collaboration and build connections with others within the University of Illinois System, both within and across campuses. The University of Illinois System comprises researchers at different sites throughout the state, including an R1 research university and the primary medical campus in Chicago and an R1 land grant university in Urbana-Champaign. Collaboration was iterative toward shared goals and involved significant amounts of coordination. Typically, such collaboration happens not only between those who represent the same discipline but across disciplines. Throughout the project, we met on a weekly or biweekly basis via Zoom or Webex; these meetings began on 3/20/2020 and provided an opportunity for research groups to share their progress, establish deadlines, identify existing challenges quickly, and ensure that projects remained on track.

Many survey respondents felt satisfied with the amount of collaboration and networking with those at different sites. Our research investigators were generally happy with the progress made and noted their enjoyment of being part of an interprofessional project. However, for some, the collaboration did not go far enough. Weekly meetings often were insufficient to address minor, day-to-day issues. Sometimes there were difficulties within the groups themselves, as is often the case within any larger group project. More frequent technical meetings were needed to address questions regarding implementation-level details not covered at larger, high-level project meetings. How the meetings are conducted is essential, as they must be designed to allow team members to excel, foster a spirit of creativity, and build cohesion toward a mission [Reference McGeorge15,Reference Petz16]. This process helped the team directly communicate their expectations to one another and ensure the compatibility of their technical approaches in a more up-to-date manner. Some project members felt that the various teams were siloed from one another, with busy clinical content experts not always accessible and technical experts unsure how to best translate between clinical expectations and the practical aspects of their processing and modeling approaches.

With larger teams, it was occasionally challenging to collaborate well with poorly defined goals and consensus aims. Some subgroups felt difficulties in collaboration. Data analysts felt overutilized and left mid-way to work on other operational efforts, partly due to changing priorities and increased responsibilities from the COVID pandemic. As a result, the researchers, physicians, and other informaticists found it challenging to perform adequate data processing. Similar to general healthcare settings, having more precise goals and objectives would have helped provide direction to answer many of our questions [Reference Xyrichis and Lowton17]. As a result, accountability and responsibility can be promoted among researchers and institutional officers [18].

Having members of two separate University of Illinois System locations (Chicago and Urbana-Champaign) led to a need to work within separate administrative systems. In regards to the regulation of research conduct and ethical approval, administrative operations, and finances, the campuses function as distinct entities, introducing barriers to collaboration and research across campuses.

Communication is essential and notoriously challenging when working between clinicians, researchers, epidemiologists, and data scientists. Available grants often did not engage these groups similarly. NIH grants are often crucial for producing impactful health research [Reference Ardehali19]. However, attracting funding from the NIH requires understanding the various funding programs [Reference Jacob and Lefgren20]. For instance, we noticed separate unique processes to apply for clinical NIH and engineering NSF grants. Salary support for project involvement may also provide motivation and feasibility for greater participation and engagement.

The NLP group required clinician annotators to provide supervised labels for a subset of their data, facilitating automated techniques for extracting relevant clinical variables. Many factors and guidelines exist to develop proper annotation procedures [Reference Wei, Franklin, Cohen and Xu21]. However, the competing demands on time from other aspects of the project and the ongoing pandemic left the clinicians in the group with little time to provide annotations. As a result, tasks were often deferred, requiring the addition of new team members to IRBs, familiarizing them with project details, and other delays. Thus, requested annotations were rarely completed on time, hindering progress in developing and evaluating NLP models and approaches. A way to ameliorate issues arising from data annotation bottlenecks in future work would be to include room in the budget to pay trained annotators and clinical trainees to complete these tasks, thus incentivizing and protecting the effort and allowing the team to identify annotators able to provide dedicated or protected time for this work [Reference Barnard22,Reference Young, Dehaven, Passmore and Baumer23]. Likewise, establishing clear annotation guidelines early in the project, in concert with clinicians, will promote high annotation quality and reproducibility, thereby reducing the need for later potential revision and reannotation. Nondomain and automated annotators supported by machine learning have been considered [Reference Krenzer, Makowski and Hekalo24]. Leveraging tools like NLM Scrubber [25] were only more recently utilized on 16 GB of textual clinical note datasets to remove identifiers.

Lessons Learned

  • Good coordination and iteration toward shared goals are important.

  • Frequent meetings help ensure projects stay on track. Even weekly meetings are often insufficient for day-to-day and technical issues that frequently arise.

  • Research analysts should be supported in their work, but accountability and responsibility should also be promoted.

  • Disparate administrative systems may lead to unforeseen barriers.

  • Communication is essential but challenging between clinicians, researchers, epidemiologists, and data scientists. Incentives should be considered for timely participation.

Theme 2: Infrastructure

The CCTS Biomedical Informatics Core was considered to have provided adequate support, especially when there were questions regarding the data and libraries needed to develop technical methods. Additionally, PACS images were successfully loaded into the CRDW, although with much difficulty due to limited transfer capacity, lack of manual support, and the need to investigate and develop processes to transfer and code images. A secure computing environment (SCE) was created with the University of Illinois’s Advanced Cyberinfrastructure for Education and Research group, which was critical to ensuring patient privacy and security. It was also the first time in our institution’s history that identified patient data was allowed to be placed in such an environment.

The development of a suitable computing infrastructure took considerably longer than expected, partly due to insufficient amounts of RAM, storage capacity, CPU/GPU power, and appropriate data processing software and associated costs. By having multiple projects running concurrently with multiple users, storage space on the SCE quickly ran out, necessitating applications for more space. As a result, the environment became slow, and many applications stalled. In addition, increasing our capabilities required using more of our limited funds. For this reason, the SCE was looked upon unfavorably by our clinicians, researchers, and engineers. It was also considered overregulated and defensively built. One research team member noted that this environment was designed “not at all for the benefit of the scientists… trying to produce results more quickly.” Non-clinician researchers were limited to performing their analyses using only tools approved within the SCE, which hindered them from employing many newer, state-of-the-art tools (e.g., specific machine learning tools and the graphic user interface). Obtaining new software, such as a new data visualization program (e.g., Tableau), required an added layer of information technology management and an additional request. Allowing greater autonomy to the research team, such as a degree of limited sudo access [Reference Napier26], for their software installation within the environment would speed up many processes, including installing text editors and other easily downloadable programs necessary for machine learning.

An additional barrier in the SCE was the absence of GPU support. For technical researchers working with machine learning and NLP models, this prevented them from being able to leverage and experiment with well-known deep learning techniques that may have had a strong performance. Enabling GPU support, in general, would also expedite the processing time for larger machine-learning models, including those that the researchers were able to implement and run using CPUs.

Finally, our imaging team could not use the SCE. Throughout the project, transferring Digital Imaging and Communications in Medicine (DICOM) high-resolution images to the SCE was problematic on many levels. These issues included technical issues and departmental policies regarding the sharing and transferring of data. At the imaging department level, our project was not provided personnel effort to extract DICOM images to transfer to the SCE.

We learned through this process how important it is to work within the structure of our institutional review board (IRB) and CCTS. By nimbly adapting to this workflow, we can hope to be more productive in the future. Much of our research is moving toward cloud computing to mitigate some of these difficulties. Cloud computing will bring unique challenges (e.g., cost containment [Reference McCafferty27] and security [Reference Kuo28]). When doing so, we must continue to commit to protecting health information. Nevertheless, at our institution, we have demonstrated how leveraging cloud computing technology can accelerate public health understanding of emerging health threats [Reference Jalali, Olabode and Bell29] and create a compressive synthetic syndromic surveillance system [Reference Jalali, Huang and Kochendorfer30]. Utilizing cloud-based systems can provide our COVID-19 team with best practices, implementation guidance, source code, and tools to help cloud deployment and augment our research capabilities. Ideally, federated learning environments will allow multiple health systems to build models without having to share private data [Reference Rieke, Hancox and Li31].

Lessons Learned

  • Development of novel secured computing environments take time and may lead to unexpected technical challenges.

  • When using SCE, elements such as storage space, GPU support, and access control should be managed well.

  • SCE should be optimized for desired tasks.

  • Cloud computing should be considered to mitigate challenges, but potential cost overruns need to be monitored and managed.

Theme 3: Data Acquisition/Validation

The research efforts coincided with implementing a new EHR system when UI Health converted from Cerner to Epic in September 2020. Approximately five years of clinical data from Cerner were migrated into Epic. In addition, during our EHR transition over the following year, inconsistencies were compounded in EHR data entry between clinicians while our researchers and data scientists became beholden to how data is entered. For example, data revolving around housing instability and race/ethnicity were not standardized or consistent. Having clinician partners can help elucidate errors and fill in gaps.

Data governance, accountability, and privacy rights remain paramount when overseeing health information. Governmental regulations such as HIPAA, 21st Century Cures Act, and The Common Rule may apply to select research activities. The CCTS has been IRB approved for a CRDW since 2012, and the existing CRDW contains patient data going back to January 2010. All clinical data was ingested into the CRDW and converted to a research common data model (CDM). Prior to September 12, 2020, EHR data was from Cerner. Following that date, EHR data ingested into the CRDW was sourced from Epic. The ACT/i2b2 [32] CDM was chosen as a reasonable common ground for merging data from both EHRs into a research data model format. As needed, the data model has been extended to support various research data requests. While these mappings and the conversion were difficult and time-consuming, the CCTS BI Core has been able to accomplish this seamlessly from the perspectives of most researchers. Some survey respondents noted that the data was received in batches and was not provided within a consistent timeframe. In hindsight, a dedicated pipeline could have been created to update the requested EHR data weekly.

Data transfer was particularly challenging with our imaging informatics efforts. Goals had to be modified over time due to challenges in transferring images. In addition, much paperwork became necessary to obtain information. While the PACS team in the imaging department thought that the PACSGear Media Writer could be used to create and export images to the SCE, it would have resulted in a significant amount of time to individually extract a large amount of DICOM data, given the large cohort. Additionally, IT personnel’s proposed use of a 10-gigabit tunnel for this type of use proved to be too complex for the system to work because of technical issues. Further potential assistance from CCTS to extract the DICOM images was not possible given policy constraints for transfer from the hospital system information systems that could not be changed. Ultimately, concerns were raised at the university campus level. Information Services noted that extraction of PACS images would be complex because of HIPAA concerns and limitations, where direct access to the PACS system is given to researchers.

Eventually, given the inability to extract DICOM images from JPEG, all Chest X-Rays (CXR) were extracted from the PACS system utilizing a scripted method (SikuliX, 2.0.2) and saved in an encrypted laptop as high-quality 24-bit JPEG files (1669 × 1538 to 3032 × 2520 pixels). CXR analysis used for one particular deep learning algorithm predicting comorbidities and clinical outcome details was established in published work [Reference Pyrros, Rodríguez-Fernández and Borstelmann11,Reference Pyrros, Rodriguez Fernandez and Borstelmann12]. A solution to this issue may be to use a web environment with optional de-identification of imaging data to facilitate data distribution within a hospital environment [Reference Aryanto, Broekema, Langenhuysen, Oudkerk and van Ooijen33]. Additionally, there is an increasing use of enterprise-level PACS systems that use cloud storage for more efficient anonymization and upload [Reference McRoy, Patel and Gaddam34]. Additionally, GE Healthcare plans to include the Edison Digital Health Platform, which will optimize the use of AI applications across health ecosystems [Reference Whooley35]. As a result of these issues and challenges, the CCTS has since added clinical images to its CRDW to help UIC researchers appropriately and securely incorporate these data elements into their future research when needed.

NLP has been used widely in healthcare [Reference Koleck, Dreisbach, Bourne and Bakken36,Reference Si, Wang, Xu and Roberts37] and understanding COVID-19 [Reference Wagner, Shweta and Murugadoss38]. The team installed a popular annotation framework, Inception [Reference Klie, Bugert, Boullosa, Eckart de Castilho and Gurevych39], to facilitate annotation on the SCE. While this tool worked well, the notes provided were challenging to utilize consistently, especially given the limited computing resources available in the SCE. The project also suffered from slow progress towards data annotation due mainly to limited annotator availability, proving to be a significant unforeseen bottleneck. Ultimately, the NLP team opted to address this by using techniques designed for low-resource tasks, relying mainly on semi-supervised and active learning methods to identify relevant information in clinical notes.

Lessons Learned

  • Data standardization, validation, cleaning, and analysis can be a lengthy arduous process.

  • Clinician partners can help find errors and fill in gaps.

  • Data governance, accountability, and privacy rights remain important when working with a clinical research data warehouse. Data may not be received in a timely or consistent manner.

  • Data transfer can be challenging, especially with imaging data.

Theme 4: Model Building

We successfully created models and compared our results with other published models. These models were published in journals [Reference Pyrros, Flanders and Rodríguez-Fernández9Reference Pyrros, Rodriguez Fernandez and Borstelmann12] and presented at national conferences. However, translation of such models to the clinical setting proved difficult, partly due to the malalignment of incentives throughout our hospital organization. A robust translational process would allow us to accelerate research results in the clinical environment and the community [Reference Zerhouni40]. Roadmaps and frameworks could help to outline investments and activities required to overcome this barrier [Reference Dougherty and Conway41].

We aimed to identify factors associated with poor outcomes in our patients with a new disease for which we had little clinical experience. With a pending EHR transition, models could be considered a clinical decision support tool to improve care and stored outside the EHR. Some hoped to create a point system, similar to Ranson’s criteria [Reference Ranson, Rifkind, Roses, Fink, Eng and Spencer42], placed online with the ability to create shareable plastic reference cards with simple decision rules based on a developed forest tree plot model. There was modest communication between our research team and the hospital system, leading to limited internal dissemination. Getting buy-in became difficult. There is a well-entrenched and reasonable bias toward published models. However, the timeline toward publication is generally slower and not conducive to implementing a new tool in a clinical setting. Because we had an internal grant, went through the IRB, and used the CCTS Bioinformatics Core, there was probably the assumption that it was all research and not operational. Increased support and involvement of clinical informatics physicians could have helped to provide domain expertise, bridge gaps between researchers and clinicians, promote these processes among leadership, and be engaged in change management processes.

Some committee members noted that our models focused primarily on admitted patients, but there was a more general question regarding use in the emergency department and whether the models can help determine whether it was safe to send a COVID-19 patient home. These were valid questions, and to be successful, we would have needed mortality data from neighboring hospitals, the city, and the state. Furthermore, we could have leveraged national data and studies better, including comparing our results to those at other institutions. A proposed COVID-19 registry offered from ACS (American College of Surgery) for use in REDCap [43] was deemed not valuable for data collection efforts. National datasets, like the National COVID Cohort Collaborative (N3C) and the National Technical Information Service managed Limited Access Death Master File (LADMF) [44], were considered. Unfortunately, we did not have the resources to obtain or get up to speed quickly enough on these outside resources to benefit our initial model creation. Eventually, team members participated in some N3C projects, and the acquisition of the LADMF was undertaken to help with future research efforts.

Lessons Learned

  • Translation of research models to a clinical setting is difficult to achieve.

  • Communication with hospital administrators and clinicians is important to obtain buy-in.

  • National data sources and registries should be leveraged.


As an institution, we continue to work on COVID-19 research endeavors. In the process, we learned about our strengths and deficiencies as an organization. We could group these lessons around collaboration, infrastructure, data acquisition/validation, and model-building themes (Table 1). Newer goals did develop out of this project, and we learned more about how our new EHR presents clinical data to our health system. We also continue to work with CCTS to improve our data translation and warehousing capabilities. There has been progress working on analogous efforts in other disease areas, including diabetes, pancreatic cancer, and HIV.

Table 1. Lessons learned summary


We acknowledge the editorial assistance of the University of Illinois Chicago Center for Clinical and Translational Science, supported by the National Center for Advancing Translational Sciences (NCATS), National Institutes of Health, through Grant Award Number UL1TR002003.

We thank Edward Barbour and Robin Mermelstein, both of the Center for Clinical and Translational Science at the University of Illinois Chicago. They provided valuable insight into CCTS infrastructure and guidance during the writing of this manuscript.

Of note, as of 7/1/2022, Sean S. Huang has been employed at Vanderbilt University Medical Center – Department of Medicine and Biomedical Informatics.

The project described was supported by the National Center for Advancing Translational Sciences (NCATS), National Institutes of Health, through Grant Award Number UL1TR002003. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.


The authors have no conflicts of interest to report.


Mamidi, TKK, Tran-Nguyen, TK, Melvin, RL, Worthey, EA. Development of an individualized risk prediction model for COVID-19 using electronic health record data. Front Big Data. 2021;4:675882. DOI: 10.3389/fdata.2021.675882.CrossRefGoogle ScholarPubMed
Bennett, TD, Moffitt, RA, Hajagos, JG, et al. Clinical characterization and prediction of clinical severity of SARS-CoV-2 infection among US adults using data from the US National COVID Cohort Collaborative. JAMA Netw Open. 2021;4(7):e2116901. DOI: 10.1001/jamanetworkopen.2021.16901.CrossRefGoogle ScholarPubMed
Vaid, A, Somani, S, Russak, AJ, et al. Machine learning to predict mortality and critical events in a cohort of patients with COVID-19 in New York City: model development and validation. J Med Internet Res. 2020;22(11):e24018. DOI: 10.2196/24018.CrossRefGoogle Scholar
Syrowatka, A, Kuznetsova, M, Alsubai, A, et al. Leveraging artificial intelligence for pandemic preparedness and response: a scoping review to identify key use cases. NPJ Digit Med. 2021;4(1):96. DOI: 10.1038/s41746-021-00459-8.CrossRefGoogle ScholarPubMed
Roberts, M, Driggs, D, Thorpe, M, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat Mach Intell. 2021;3(3):199217. DOI: 10.1038/s42256-021-00307-0.CrossRefGoogle Scholar
Wynants, L, Van Calster, B, Collins, GS, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal [published correction appears in BMJ. 2020 Jun 3;369:m2204]. BMJ. [2020;369:m1328]. DOI: 10.1136/bmj.m1328.CrossRefGoogle Scholar
Zhang, A, Xing, L, Zou, J, Wu, JC. Shifting machine learning for healthcare from development to deployment and from models to data. Nat Biomed Eng. 2022;6(12):13301345. DOI: 10.1038/s41551-022-00898-y.CrossRefGoogle ScholarPubMed
Baker, RE, Mahmud, AS, Miller, IF, et al. Infectious disease in an era of global change. Nat Rev Microbiol. 2022;20(4):193205. DOI: 10.1038/s41579-021-00639-z.CrossRefGoogle Scholar
Pyrros, A, Flanders, AE, Rodríguez-Fernández, JM, et al. Predicting prolonged hospitalization and supplemental oxygenation in patients with COVID-19 infection from ambulatory chest radiographs using deep learning. Acad Radiol. 2021;28(8):11511158. DOI: 10.1016/j.acra.2021.05.002.CrossRefGoogle ScholarPubMed
Galanter, W, Rodríguez-Fernández, JM, Chow, K, et al. Predicting clinical outcomes among hospitalized COVID-19 patients using both local and published models. BMC Med Inform Decis Mak. 2021;21(1):224. DOI: 10.1186/s12911-021-01576-w.CrossRefGoogle ScholarPubMed
Pyrros, A, Rodríguez-Fernández, JM, Borstelmann, SM, et al. Detecting racial/ethnic health disparities using deep learning from frontal chest radiography [published correction appears in J Am Coll Radiol. 2022 Mar;19(3):403]. J Am Coll Radiol. 2022;19(3):403–191. DOI: 10.1016/j.jacr.2021.09.010.Google Scholar
Pyrros, A, Rodriguez Fernandez, J, Borstelmann, SM, et al. Validation of a deep learning, value-based care model to predict mortality and comorbidities from chest radiographs in COVID-19. PLOS Digit Health. 2022;1(8):e0000057.CrossRefGoogle ScholarPubMed
Pishgar, M, Harford, S, Theis, J, et al. A process mining- deep learning approach to predict survival in a cohort of hospitalized COVID-19 patients. BMC Med Inform Decis Mak. 2022;22(1):194. DOI: 10.1186/s12911-022-01934-2.CrossRefGoogle Scholar
Sharafeldin, N, Bates, B, Song, Q, et al. Outcomes of COVID-19 in patients with cancer: report from the National COVID Cohort Collaborative (N3C). J Clin Oncol. 2021;39(20):22322246. DOI: 10.1200/JCO.21.01074.CrossRefGoogle ScholarPubMed
McGeorge, D. The 25 Minute Meeting: Half the Time, Double the Impact. Hoboken, NY: John Wiley & Sons; 2019.Google Scholar
Petz, J, Boring Meetings Suck: Get More Out of Your Meetings, or Get Out of More Meetings, Hoboken, NY: John Wiley & Sons; 2011.Google Scholar
Xyrichis, A, Lowton, K. What fosters or prevents interprofessional teamworking in primary and community care? A literature review. Int J Nurs Stud. 2008;45(1):140153. DOI: 10.1016/j.ijnurstu.2007.01.015.CrossRefGoogle ScholarPubMed
American Society of Clinical Oncology. American Society of Clinical Oncology policy statement: oversight of clinical research. J Clin Oncol. 2003;21(12):23772386. DOI: 10.1200/JCO.2003.04.026.CrossRefGoogle Scholar
Ardehali, H. How to write a successful grant application and research paper. Circ Res. 2014;114(8):12311234. DOI: 10.1161/CIRCRESAHA.114.303695.CrossRefGoogle ScholarPubMed
Jacob, BA, Lefgren, L. The impact of research grant funding on scientific productivity. J Public Econ. 2011;95(9-10):11681177. DOI: 10.1016/j.jpubeco.2011.05.005.CrossRefGoogle ScholarPubMed
Wei, Q, Franklin, A, Cohen, T, Xu, H. Clinical text annotation - what factors are associated with the cost of time? AMIA Annu Symp Proc. 2018;2018:15521560.Google ScholarPubMed
Barnard, JA. Protected time: a vital ingredient for research career development. J Pediatr Gastroenterol Nutr. 2015;60(3):292293. DOI: 10.1097/MPG.0000000000000693.CrossRefGoogle Scholar
Young, RA, Dehaven, MJ, Passmore, C, Baumer, JG. Research participation, protected time, and research output by family physicians in family medicine residencies. Fam Med. 2006;38(5):341348.Google ScholarPubMed
Krenzer, A, Makowski, K, Hekalo, A, et al. Fast machine learning annotation in the medical domain: a semi-automated video annotation tool for gastroenterologists. Biomed Eng Online. 2022;21(1):33. DOI: 10.1186/s12938-022-01001-x.CrossRefGoogle ScholarPubMed
Clinical Text De-identification using NLM-Scrubber. NIH National Library of Medicine lister Hill National Center for Biomedical Communications, 2022. (, cited August 19, 2022.Google Scholar
Napier, R. Secure automation: Achieving least privilege with SSH, Sudo and Setuid. In: 18th Large Installation System Administration Conference, Atlanta, GA, USA, 2004.Google Scholar
McCafferty, D. How unexpected costs create a ‘Cloud hangover’. CIO Insight, 2021. (, cited August 19, 2022.Google Scholar
Kuo, AM. Opportunities and challenges of cloud computing to improve health care services. J Med Internet Res. 2011;13(3):e67. DOI: 10.2196/jmir.1867.CrossRefGoogle ScholarPubMed
Jalali, A, Olabode, OA, Bell, CM. Leveraging cloud computing to address public health disparities: an analysis of the SPHPS. Online J Public Health Inform. 2012;4(3). DOI: 10.5210/ojphi.v4i3.4325.Google ScholarPubMed
Jalali, A, Huang, S, Kochendorfer, K. UI Health: Spatiotemporal Big Data Analysis of Opioid Epidemic in Illinois. U.S. Department of Health and Human Services Office of the National Coordinator for Health Information Technology. In: Synthetic Health Data Challenge: Winning Solutions Webinar. Virtual; 2021.Google Scholar
Rieke, N, Hancox, J, Li, W, et al. The future of digital health with federated learning. NPJ Digit Med. 2020;3(1):119. DOI: 10.1038/s41746-020-00323-1.CrossRefGoogle ScholarPubMed
Informatics for integrating biology & the bedside, 2022. (, cited August 19, 2022.Google Scholar
Aryanto, KY, Broekema, A, Langenhuysen, RG, Oudkerk, M, van Ooijen, PM. A web-based institutional DICOM distribution system with the integration of the Clinical Trial Processor (CTP). J Med Syst. 2015;39(5):45. DOI: 10.1007/s10916-014-0186-y.CrossRefGoogle ScholarPubMed
McRoy, C, Patel, L, Gaddam, DS, et al. Radiology education in the time of COVID-19: a novel distance learning workstation experience for residents. Acad Radiol. 2020;27(10):14671474. DOI: 10.1016/j.acra.2020.08.001.CrossRefGoogle ScholarPubMed
Whooley, S. GE Healthcare wants to expand the use of AI in healthcare: here’s how. Mass Device, 2019. (, cited August 19, 2022.Google Scholar
Koleck, TA, Dreisbach, C, Bourne, PE, Bakken, S. Natural language processing of symptoms documented in free-text narratives of electronic health records: a systematic review. J Am Med Inform Assoc. 2019;26(4):364379. DOI: 10.1093/jamia/ocy173.CrossRefGoogle ScholarPubMed
Si, Y, Wang, J, Xu, H, Roberts, K. Enhancing clinical concept extraction with contextual embeddings. J Am Med Inform Assoc. 2019;26(11):12971304. DOI: 10.1093/jamia/ocz096.CrossRefGoogle ScholarPubMed
Wagner, T, Shweta, F, Murugadoss, K, et al. Augmented curation of clinical notes from a massive EHR system reveals symptoms of impending COVID-19 diagnosis. Elife. 2020;9:e58227. DOI: 10.7554/eLife.58227.CrossRefGoogle ScholarPubMed
Klie, J-C, Bugert, M, Boullosa, B, Eckart de Castilho, R, Gurevych, I. The INCEpTION platform: machine-assisted and knowledge-oriented interactive annotation. In: Proceedings of System Demonstrations of the 27th International Conference on Computational Linguistics (COLING 2018), Santa Fe, New Mexico, USA, 2018.Google Scholar
Zerhouni, EA. Translational research: moving discovery to practice. Clin Pharmacol Ther. 2007;81(1):126128. DOI: 10.1038/sj.clpt.6100029.CrossRefGoogle ScholarPubMed
Dougherty, D, Conway, PH. The “3T’s” road map to transform US health care: the “how” of high-quality care. JAMA. 2008;299(19):23192321. DOI: 10.1001/jama.299.19.2319.CrossRefGoogle Scholar
Ranson, JH, Rifkind, KM, Roses, DF, Fink, SD, Eng, K, Spencer, FC. Prognostic signs and the role of operative management in acute pancreatitis. Surg Gynecol Obstet. 1974;139(1):6981.Google ScholarPubMed
ACS Covid-19 Registry. American College of Surgeons, 2022. (, cited August 19, 2022.Google Scholar
NTIS LADMF. United States of America Department of Commerce, 2022. (, cited August 19, 2022.Google Scholar
Figure 0

Figure 1. Survey questions.

Figure 1

Table 1. Lessons learned summary