Dashboarding to Monitor Machine-Learning-Based Clinical Decision Support Interventions

Abstract

Background Existing monitoring of machine-learning-based clinical decision support (ML-CDS) is focused predominantly on the ML outputs and accuracy thereof. Improving patient care requires not only accurate algorithms but also systems of care that enable the output of these algorithms to drive specific actions by care teams, necessitating expanding their monitoring.

Objectives In this case report, we describe the creation of a dashboard that allows the intervention development team and operational stakeholders to govern and identify potential issues that may require corrective action by bridging the monitoring gap between model outputs and patient outcomes.

Methods We used an iterative development process to build a dashboard to monitor the performance of our intervention in the broader context of the care system.

Results Our investigation of best practices elsewhere, iterative design, and expert consultation led us to anchor our dashboard on alluvial charts and control charts. Both the development process and the dashboard itself illuminated areas to improve the broader intervention.

Conclusion We propose that monitoring ML-CDS algorithms with regular dashboards that allow both a context-level view of the system and a drilled down view of specific components is a critical part of implementing these algorithms to ensure that these tools function appropriately within the broader care system.

 

Scroll to top

Keywords

dashboards - machine learning - clinical decision support - clinical informatics - statistical process control

Introduction

Machine learning (ML) based algorithms are increasingly being used in clinical care.[1] [2] [3] [4] These algorithms can be embedded into health care providers' workflows through clinical decision support (CDS) in electronic health record (EHR) software to match patients to the resources they need.[1] [2] [4] [5] [6] These interventions offer opportunities to advance the quintuple aim of health care: reducing costs while improving quality, equity, and patient and provider experience.[7] [8] [9] To make the most of these opportunities, implementation of such tools must be coordinated with users and stakeholders[6] [10] [11] [12] [13] [14] [15] [16] and continuously monitored to ensure that the models maintain fidelity in the face of data drift, adaptive system dynamics, and deviations from a priori assumptions about patient and provider behavior.[4] [13] [16] [17] [18] [19] [20] [21] [22] There is a growing understanding that such solutions must be carefully monitored and adjusted to achieve the full potential of ML-based CDS (ML-CDS) interventions[21] and that such monitoring needs to take into account not just the ML algorithm but patient care system features as well.[22]

 

Scroll to top

Objective

Existing monitoring of ML-CDS is focused predominantly on the ML outputs and accuracy thereof.[17] [20] [21] Improving patient care requires not only accurate algorithms but also systems of care that enable the output of these algorithms to drive specific actions by care teams, necessitating enhanced monitoring.[6] [10] [22] In this case report, we describe the development of a dashboard that bridges the monitoring gap between model outputs and patient outcomes.

 

Scroll to top

Methods

Setting

As part of a larger study, a large academic medical system in the Midwestern United States implemented an ML-CDS tool in 2021 with the aim of preventing falls after emergency department (ED) visits. An algorithm scores all eligible older adults and outputs a risk score between 0 and 1 for return visit to the ED for fall within 6 months. During the discharge workflow, providers receive an interruptive alert when eligible patients are scored above a threshold. Providers are prompted to refer the patient to an evidence-based, multidisciplinary review at an outpatient geriatrics clinic, which they can order or reject with a single click.[11] [23] System capacity constraints necessitated optimizing the intervention to give an average of five referrals per week. The operational leadership team updates the threshold used to flag patients based on ED volumes and referral rates to avoid overwhelming the falls clinic. Details of this ML-CDS, the underlying data, and effectiveness evaluation have been described elsewhere.[11] [23] [24]

Since implementation, the ML-CDS model has been monitored at weekly operational meetings to track its use and adoption. To facilitate this review and differentiate between areas that require closer attention and statistical noise, we developed a dashboard that allows operational leaders to visualize the whole process at both a context level and a more detailed level in a “zoom-out/zoom-in” manner.[15] [16] [25]

 

Scroll to top

Data and Development

We began our development of a dashboard by identifying discrete metrics to monitor each successive action between risk score calculation and patient completion or noncompletion of the recommended follow-up care. These metrics included the number of patients eligible, patients flagged, patients referred, scheduled clinic appointments, and completed clinic appointments.

Data came from the EHR with additional manual review that occurred within REDCap. For example, if a scheduling note indicated “patient not interested at this time,” then the order would be temporarily classified using regular expressions as not scheduled for reason of patient not interested and that classification would be manually confirmed by the study team. Sometimes, orders were marked as canceled as duplicate orders, but a completed referral was found for the same patient; in those cases, the referral visit information and status were added to the referral order to indicate that the patient was seen at the outpatient falls clinic. The data structure is described in detail elsewhere.[24]

Within the dashboard, many metrics (e.g., referral rates) were calculated at a visit level, although some outcome measures (e.g., was the outpatient visit completed?) were aggregated at the patient level. For example, a patient could be referred but ultimately be ineligible for a falls prevention visit because they already completed one after a prior ED referral. To create a patient-level view, we used an alluvial plot[26] [27] to give a “zoomed-out” overview that allows us to depict the multistep process through the care system and scheduling process.

Real-world data are noisy, and it may not be immediately clear whether changes in performance are due to issues with the underlying model (like data drift), changes in EHR configuration, or simply random fluctuations.[22] [28] [29] To that end, we incorporated control charts to help distinguish between meaningful changes in outcomes and normal variation.[29] Control charts have been applied in health care quality improvement as early as the 1990s by Berwick[30] and continue to be a valuable tool in such initiatives.[22] [28] [31] When applied to ML-CDS, control charts offer implementers the ability to “zoom in” from the overall view to a more focused examination of specific steps of the intervention. This allows relevant parties to assess whether recent changes—such as a presentation at faculty meeting about the ML-CDS—have had a significant impact on the system's performance. Additionally, control charts can aid in detecting data drift in the ML algorithm and help identify areas for improvement.

 

Scroll to top

Iterative Development of Dashboard

It is important to continuously refine data definitions with stakeholders to ensure data accuracy.[13] To achieve this, we used an iterative software development approach, integrating regular stakeholder meetings and feedback to inform subsequent programming iterations.[32] This approach facilitates continuous improvement and adaptability by incorporating stakeholder input into each stage of development.[13] [15] [32] [33] While the dashboard was being developed, drafts were reviewed regularly by the project team, which included a physician, research specialists, and systems engineers, all of whom were responsible for implementing the ML-CDS intervention. During these reviews, participants were encouraged to provide open-ended feedback, including feedback on the dashboard's usability and the operational insights it offered (e.g., “why are so many patients already in treatment being identified?”). This feedback guided updates to the dashboard for the next meeting and ensured user acceptability of the final dashboard. The dashboard was created with RMarkdown[34] [35] [36] [37] [38] and was informed by the usability principles laid out by Wilkinson[39] and Dowding and Merrill.[25]

 

Scroll to top

 

Scroll to top

Results

Dashboard Visualizations

Additional figures on the dashboard included a distribution of falls risk scores, and how many repeat patients were receiving multiple referrals. Finally, some tables on the dashboard reflected the number of referrals that required study team review for categorization. A full example dashboard is available in the [Supplementary Material] (available in the online version only); the code is available at www.hipxchange.org/CDSDashboard where we encourage readers to review an example dashboard, manipulate randomized test data, and explore how the code might be adapted to their use cases.

 

Scroll to top

Operational Outcomes

The dashboard drives discussion in regular leadership meetings about ways to evaluate and potentially improve the broader intervention.[14] By breaking down the patient flow into measurable process steps, we gained a better understanding of the workflows involved in scheduling to understand scheduler work as performed, not just work as imagined.[40] For example, when canceling a referral because the patient was ineligible for the intervention due to being in other physical therapy, some schedulers indicated the reason in a discrete flowsheet, which was the intended workflow as designed; most schedulers indicated the reason in a free-text comment, which we anticipated would happen; and some schedulers indicated the reason in a separate scheduling note, which we did not expect and had to adapt to.

Secondly, the dashboard, especially [Fig. 1], guided tuning the ML-CDS risk threshold for identifying patients to refer. We were able to track whether we were referring the intended average number of patients per week (five referrals) to the clinic. If this number was too low, we adjusted the threshold risk used to flag patients for providers. After adjustment, we were able to easily gauge the effects on flagging, referral, and scheduling rates ([Fig. 1A], B, and D). Control charts were able to identify empirically whether the number of referrals was meeting our operational goal more concretely than looking at total referrals or asking the clinic for their impression.

Fig. 1 A series of control charts. We monitored to ensure that a manageable number of patients were being flagged (A) so as not to contribute to alert fatigue by flagging too many patients; that providers were referring a consistent number of flagged patients (B); our operational goal to make five referrals per week (C), and that those patients were making appointments at the specialty clinic (D). Monitoring these items together allowed us to ensure that risk thresholds were set appropriately. Following conventions for control charts,[45] we identify deviations from expectations as repeated observations at or beyond the control limits for each graph, particularly for time frames after known changes to the machine-learning-based clinical decision support (ML-CDS) tool or emergency department (ED) workflows or changes in ED patient populations.

Additionally, the alluvial chart ([Fig. 2]) allowed us to identify common reasons for referral noncompletion and act as appropriate. While some of these reasons were outside our control, such as having received the intervention at another health system, some of the reasons had actionable responses. For example, the number of patients who reported not being interested in the intervention was higher than we had anticipated, so we undertook an overhaul in the messaging surrounding the referral as part of patients' discharge summary, which preliminary qualitative evidence suggests helped, but those changes are too recent to evaluate at this time. We also found that uncompleted referrals were being written for patients in end-of-life hospice care or patients who are already established with the mobility and falls clinic, which we were then able to exclude from the CDS criteria.

Fig. 2 An alluvial plot to visualize patient flows from one stage of the intervention to another. The width of each curved arc between points is proportional to the number of patients represented by that path. Colors roughly correspond to desirability of outcomes, but primarily provide contrast and improved readability. The “Manual Review Needed” node includes patients with referrals that could not be automatically classified and require a chart review from the project team to identify the status and reason for noncompletion of the order. Typically, these were patients for whom scheduler documentation was documented in a nonstandard field and a deeper chart review of scheduling notes was needed. Tracking this figure gives an overview of the reasons patients may not complete the order.

 

Scroll to top

 

Scroll to top

Discussion

We built a dashboard to monitor the performance of our intervention in the broader context of the care system. Dashboards are ubiquitous, but applying quality improvement dashboards focused on outcomes beyond ML algorithm accuracy is the subject of recent discussion.[22] Our investigation of best dashboard practices elsewhere, iterative design, and expert consultation led us to anchor our dashboard on alluvial charts and control charts. Translating our clinical intervention into a dashboard helped us identify gaps in the long chain of events that make up this intervention. As a result, we could allocate resources strategically to improve care.

By quantifying and monitoring the intervention at a patient level as well as individual component parts—such as referral rates ([Fig. 1B])—we gained valuable insights into the functioning of the system. This approach allowed us to identify when the intervention was successful in achieving its intended outcomes (completed patient referral) or if adjustments were necessary.[41] [42] Equally importantly, we were able to quickly and easily ensure that the system was configured correctly by using control charts to identify and follow up on large deviations from expected results.[28] [43]

 

Scroll to top

Limitations

This dashboard was developed for a single-use case in one system as a quality improvement component within a wider research effort.[44] While the strategy of combining alluvial and process control elements may be useful for monitoring other ML-CDS interventions, iterative design and customization will be needed to create site-specific tools that incorporate these general principles.

The ML-CDS intervention was geared at identifying fall events in the next 6 months. Thus, we did not incorporate model accuracy estimates in this dashboard due to long latency between scoring and outcome. Other analyses are being done to evaluate model performance and impact to patient care and outcomes.[44] Model error analysis would be a useful component in similar dashboards identifying more immediate-term patient outcomes, as long as they are contextualized in the care system.

 

Scroll to top

Conclusion

ML-CDS interventions are a powerful tool to improve patient outcomes without adding additional burden to providers. We propose that monitoring ML-CDS algorithms with tailored dashboards that allow both a context-level view of the system and a drilled-down view of specific components is a critical part of implementing these algorithms to ensure that these tools function within defined limits.

 

Scroll to top

Clinical Relevance

Implementation is often just as, if not more, important to the success of an ML-CDS as the area under receiver operating characteristic curve of the underlying model.[6] [10] We believe that the tools and ideas presented here will be useful to practicing clinical informaticists in their work.

 

Scroll to top

Multiple Choice Questions

1. When building a tool, what is a useful graph to show the end-to-end patient flow from start to intervention with losses to follow-up?

   • Pie charts.

   • Alluvial plots.

   • Statistical process control charts.

   • Create a “Table 1” to display the data in tabular format.

   Correct Answer: The correct answer is option b.

2. What is one threat to the ongoing success of a machine-learning-based clinical decision support tool?

   • Incorrect patient flagging.

   • Data drift.

   • Software and system updates.

   • All of the above.

   Correct Answer: The correct answer is option d.

 

Scroll to top

 

Scroll to top

Conflict of Interest

None declared.

Protection of Human and Animal Subjects

This larger intervention has been deemed minimal risk by our institutional review board and is registered at clinicaltrials.gov as NCT05810064; https://www.clinicaltrials.gov/study/NCT05810064.

• References

• 1 Bates DW, Saria S, Ohno-Machado L, Shah A, Escobar G. Big data in health care: using analytics to identify and manage high-risk and high-cost patients. Health Aff (Millwood) 2014; 33 (07) 1123-1131
• 2 Chen JH, Asch SM. Machine learning and prediction in medicine: beyond the peak of inflated expectations. N Engl J Med 2017; 376 (26) 2507-2509
• 3 Obermeyer Z, Emanuel EJ. Predicting the future: big data, machine learning, and clinical medicine. N Engl J Med 2016; 375 (13) 1216-1219
• 4 Brewer LC, Fortuna KL, Jones C. et al. Back to the future: achieving health equity through health informatics and digital health. JMIR Mhealth Uhealth 2020; 8 (01) e14512
• 5 Berlyand Y, Raja AS, Dorner SC. et al. How artificial intelligence could transform emergency department operations. Am J Emerg Med 2018; 36 (08) 1515-1517
• 6 Salwei ME, Carayon P, Hoonakker PLT. et al. Workflow integration analysis of a human factors-based clinical decision support in the emergency department. Appl Ergon 2021; 97: 103498
• 7 Berwick DM, Nolan TW, Whittington J. The triple aim: care, health, and cost. Health Aff (Millwood) 2008; 27 (03) 759-769
• 8 Sikka R, Morath JM, Leape L. The quadruple aim: care, health, cost and meaning in work. BMJ Qual Saf 2015; 24 (10) 608-610
• 9 Nundy S, Cooper LA, Mate KS. The quintuple aim for health care improvement: a new imperative to advance health equity. JAMA 2022; 327 (06) 521-522
• 10 Carayon P, Wooldridge A, Hoonakker P, Hundt AS, Kelly MM. SEIPS 3.0: Human-centered design of the patient journey for patient safety. Appl Ergon 2020; 84: 103033
• 11 Jacobsohn GC, Leaf M, Liao F. et al. Collaborative design and implementation of a clinical decision support system for automated fall-risk identification and referrals in emergency departments. Healthc (Amst) 2022; 10 (01) 100598
• 12 Ancker JS, Edwards A, Nosal S, Hauser D, Mauer E, Kaushal R. with the HITEC Investigators. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Med Inform Decis Mak 2017; 17 (01) 36
• 13 Foster M, Albanese C, Chen Q. et al. Heart failure dashboard design and validation to improve care of veterans. Appl Clin Inform 2020; 11 (01) 153-159
• 14 Safranek CW, Feitzinger L, Joyner AKC. et al. Visualizing opioid-use variation in a pediatric perioperative dashboard. Appl Clin Inform 2022; 13 (02) 370-379
• 15 Nelson O, Sturgis B, Gilbert K. et al. A visual analytics dashboard to summarize serial anesthesia records in pediatric radiation treatment. Appl Clin Inform 2019; 10 (04) 563-569
• 16 Radhakrishnan K, Monsen KA, Bae SH, Zhang W. Clinical Relevance for Quality Improvement. Visual analytics for pattern discovery in home care. Appl Clin Inform 2016; 7 (03) 711-730
• 17 Singer SJ, Kellogg KC, Galper AB, Viola D. Enhancing the value to users of machine learning-based clinical decision support tools: A framework for iterative, collaborative development and implementation. Health Care Manage Rev 2022; 47 (02) E21-E31
• 18 Duckworth C, Chmiel FP, Burns DK. et al. Using explainable machine learning to characterise data drift and detect emergent health risks for emergency department admissions during COVID-19. Sci Rep 2021; 11 (01) 23017
• 19 Mišić VV, Rajaram K, Gabel E. A simulation-based evaluation of machine learning models for clinical decision support: application and analysis using hospital readmission. NPJ Digit Med 2021; 4 (01) 98
• 20 Cieslak D, Chawla N. A framework for monitoring classifiers' performance: When and why failure occurs?. Knowl Inf Syst 2009; 18: 83-108
• 21 Bedoya AD, Economou-Zavlanos NJ, Goldstein BA. et al. A framework for the oversight and local deployment of safe and high-quality prediction models. J Am Med Inform Assoc 2022; 29 (09) 1631-1636
• 22 Feng J, Phillips RV, Malenica I. et al. Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare. NPJ Digit Med 2022; 5 (01) 66
• 23 Patterson BW, Engstrom CJ, Sah V. et al. Training and interpreting machine learning algorithms to evaluate fall risk after emergency department visits. Med Care 2019; 57 (07) 560-566
• 24 Hekman D, Maru AP, Patterson BW. Dashboarding for Machine Learning-based Clinical Decision Support Implementation and Monitoring Toolkit. HIPxChange.org. 2023 . Accessed March 16, 2023 at: https://www.hipxchange.org/CDSDashboard
• 25 Dowding D, Merrill JA. The development of heuristics for evaluation of dashboard visualizations. Appl Clin Inform 2018; 9 (03) 511-518
• 26 Jagannath S, Rifkin RM, Gasparetto CJ. et al; CONNECT MM Registry Investigators. Treatment journeys of patients with newly diagnosed multiple myeloma (NDMM): results from the Connect MM Registry. Clin Lymphoma Myeloma Leuk 2020; 20 (05) 272-276
• 27 Friendly M. Visions and re-visions of Charles Joseph Minard. J Educ Behav Stat 2002; 27 (01) 31-51
• 28 Slyngstad L. The contribution of variable control charts to quality improvement in healthcare: a literature review. J Healthc Leadersh 2021; 13: 221-230
• 29 Pimentel L, Barrueto Jr F. Statistical process control: separating signal from noise in emergency department operations. J Emerg Med 2015; 48 (05) 628-638
• 30 Berwick DM. Controlling variation in health care: a consultation from Walter Shewhart. Med Care 1991; 29 (12) 1212-1225
• 31 Thor J, Lundberg J, Ask J. et al. Application of statistical process control in healthcare improvement: systematic review. Qual Saf Health Care 2007; 16 (05) 387-399
• 32 Sotirovski D. Heuristics for iterative software development. IEEE Softw 2001; 18 (03) 66-73
• 33 Caldiera VRBG, Rombach HD. The goal question metric approach. Encycl Softw Eng 1994; 528-532
• 34 R Core Team. R: A Language and Environment for Statistical Computing. 2020. Accessed December 13, 2023 at: https://www.R-project.org/
• 35 Wickham H. ggplot2: Elegant Graphics for Data Analysis. 2016. Accessed December 13, 2023 at: https://ggplot2.tidyverse.org
• 36 Allaire JJ, Xie Y, McPherson J. et al. rmarkdown: Dynamic Documents for R. 2021. Accessed December 13, 2023 at: https://github.com/rstudio/rmarkdown
• 37 Xie Y, Dervieux C, Riederer E. R Markdown Cookbook. 2020. Accessed December 13, 2023 at: https://bookdown.org/yihui/rmarkdown-cookbook
• 38 Wickham H, Averick M, Bryan J. et al. Welcome to the tidyverse. J Open Source Softw 2019; 4 (43) 1686
• 39 Wilkinson L. The Grammar of Graphics. 2nd ed. New York, NY:: Springer;; 2005
• 40 Fairbanks RJ, Bisantz AM. Understanding better how clinicians work. Ann Emerg Med 2011; 58 (02) 123-125
• 41 Donabedian A. The quality of care. How can it be assessed?. JAMA 1988; 260 (12) 1743-1748
• 42 Carayon P, Schoofs Hundt A, Karsh BT. et al. Work system design for patient safety: the SEIPS model. Qual Saf Health Care 2006; 15 (Suppl. 01) i50-i58
• 43 Noyez L. Control charts, Cusum techniques and funnel plots. A review of methods for monitoring performance in healthcare. Interact Cardiovasc Thorac Surg 2009; 9 (03) 494-499
• 44 Hekman DJ, Cochran AL, Maru AP. et al. Effectiveness of an emergency department-based machine learning clinical decision support tool to prevent outpatient falls among older adults: protocol for a quasi-experimental study. JMIR Res Protoc 2023; 12: e48128
• 45 Juran JM. Quality-control handbook. New York, NY: McGraw-Hill; 1951. . Available at: https://cir.nii.ac.jp/crid/1130000795132214656

Publication History

Received: 22 August 2023

Accepted: 28 November 2023

Accepted Manuscript online:
29 November 2023

Article published online:
28 February 2024