Machine Learning, Ethics and Change Management in Healthcare

In the past few years, topics in predictive analytics, machine learning (ML) and artificial intelligence (AI) have increased their presence in business school, industrial engineering and applied mathematics program curricula. The market and students have demanded greater expertise in these topics, and universities have realized that they need to train graduates who can be successful in an AI-enabled future. Training for such a future involves increased technical rigor and emphasis on analytical, modeling and programming skills. There is recognition that the training should also engage AI’s challenges: ethics, quality of training data and transparency of algorithms [1, 2]. In addition, many have argued that universities should emphasize development of “soft skills” – collaboration, networking with multiple stakeholders and effective communication – to empower their graduates to counter trends in AI-based job automation [3].

This shift in content priorities is creating both challenges and opportunities for instructors in various analytics fields, including operations management, operations research, industrial engineering, data science, predictive analytics and ML. Staying relevant requires enhancing traditional operations course content with topics in the applicability of ML models, their interface with operational models, and their implementation in organizations. At the same time, incorporating operational, ethics and human factor considerations in building predictive models helps analytics, data science and ML instructors provide a more complete view of the implications and use of advanced algorithms.

Multidisciplinary cases like “Machine Learning, Ethics and Change Management: A Data-Driven Approach to Improving Hospital Observation Unit Operations,” winner of the 2020 INFORMS Case Competition, provide a canvas for illustrating such important topics. The case is based on an actual project completed in a midsize hospital, and follows the steps taken by the protagonist Dr. Erin Kelly, medical director of an observation unit, in leading a data-driven process change for placing the “right” patients in her unit. Hospital observation units (OUs) are intended to host patients for relatively short periods of time, during which healthcare providers can observe a patient and assess the need for an inpatient hospital admission [4]. Most hospitals have internal rules for placing patients, and the case refers to an “exclusion list” used at the hospital for diverting patients with particular conditions away from the OU and sending them directly to the inpatient wards. Patient case misclassifications, however, can be a substantial problem for OU operations, leading to increased average length of stay and decreased hospital capacity [5, 6].

Theoretically, the problem of predicting whether a patient with particular characteristics will be appropriate for placement in the OU can be addressed by using data analysis and machine learning algorithms. In practice, the complexity of building and implementing ML and predictive analytics algorithms in a healthcare setting is tremendous. The short case – only about 2,000 words – provides a simple setting but rich context, enabling instructors to teach not only about building ethical and interpretable predictive models, but also operationalizing them given the existing organizational structure. Concepts that can be covered with the case include service process flow, resource utilization, Little’s Law, data preparation and dealing with dirty data, data summaries and visualization, analytics problem framing and model mapping, decision trees, evaluation of predictive model performance, bias in predictive models, audience-driven multimodal communication, collaborating on an interdisciplinary team, and change management in a healthcare organization. Instructors can pick a particular concept of relevance to the subject material they teach in a class, revisit the case over multiple classes to illustrate different concepts, or collaborate with colleagues from other disciplines to cover different aspects of the case across a lock-step core curriculum to provide students with an integrated experience. The case supplies realistic data, a detailed teaching note and a variety of resources – R code, Excel model files, output, sample exercises and sample teaching plans.

We highlight three major themes that run through the case and can provide an opportunity for operations management and analytics instructors to enhance the way in which traditional operations management or analytics concepts are taught.

1. The interface of predictive and prescriptive analytics in the data analytics lifecycle. The data analytics lifecycle is a framework helpful in predictive analytics and data science courses [7]. It includes six stages: 1) discovery, 2) data preparation, 3) model planning, 4) model building, 5) communication of results and 6) operationalization. Presenting the technical aspects of model development as part of this process instills in students the important idea that developing predictive models should be based on needs identified through a disciplined discovery process and that one needs to anticipate the challenges with model implementation while the models are still being planned and built.
   
   Stand-alone predictive analytics courses rarely point out the connection between predictive and prescriptive analytics, but prescriptive analytics and operational models in particular can play a critical role in some of the stages of the data analytics lifecycle, specifically for assessing the potential for success of the project before any time or resources are spent. The case provides examples of operational models (an application of Little’s Law and sensitivity analysis) that could be either briefly referenced or discussed in detail by the instructor to determine the range of improvement that could be achieved from operationalizing a predictive model in terms of the number of additional patients that could be treated in the OU if the model is operationalized.

2. Making ethical data-driven decisions using machine learning models. The mechanics of data analysis and coding predictive models are difficult but can be taught in a structured, tractable way, even through online tutorials. Much more challenging to teach are topics that have to do with the intangibles in effective model building, such as handling dirty or missing data, summarizing patient data at a level that is useful for obtaining insights, selecting variables that can be informative, determining a meaningful target variable for the predictive model, and avoiding look-ahead bias.
   
   In recent years, the notion of interpretable machine-learning models has been gaining traction, and it is particularly relevant in the case, where the stakeholders in the project – doctors, nurses, social workers, hospital administrators – need to understand the reasons for the recommendations of the predictive model. The healthcare operations setting and the specifics of the data provided with the case also provide students with an opportunity to realize the moral complexity of operationalizing an algorithm for medical decisions, especially when sociodemographic variables turn out to be predictive.

3. Effecting organizational change to operationalize the findings of the analysis. More than 80% of analytics projects in practice do not get implemented [8]. The success of analytics projects critically depends on framing the project appropriately and clearly communicating value to the relevant stakeholders. Yet, there is a dearth of cases in our community’s literature that allow instructors to link analytics concepts to established concepts in change management. Students need to be aware, for example, that influential stakeholders should be brought on board from the beginning, and that communication with the different stakeholders should be effective to improve the chances that the project will succeed.
   
   For example, communications need to be multimodal. Reports to project sponsors need to be comprehensive to enable project sponsors to evangelize the process innovations they supported. Presentations to other stakeholders may be briefer, “what-you-need-to-know” communications. In-class activities can ask students to map the steps followed by the protagonist in the case to a typical change management framework like Kotter’s [9] to allow them to critically assess the change management process, as well as to understand the differences between a typical change management process in business and the process followed in a healthcare organization.

2020 was a disruptive year for the INFORMS community and university education. The longer-term trends of shifting content priorities have combined with a rethinking of what university education should be about, as well as a heightened sense of social responsibility. The COVID-19 pandemic and events over the past few months have caused many of us, jointly with our students, to feel urgency to use our skills to help solve pressing societal problems.

Healthcare is a sector where our community’s understanding of effective data-driven decision-making can truly make a difference. The classroom use of rich, real examples that illustrate the connections between disciplines and enable students to explore, ask questions, and grapple with the ambiguity and complexity of real-world, data-driven decision-making is critical for remaining relevant and preparing our students for dealing with the challenges of our time.

Editor’s Note: For more information about the case, see the authors’ 2020 INFORMS Annual Meeting Presentation [10].

References

1. Knowledge@Wharton, 2020, “The real threat to business schools from artificial intelligence,” March 12, https://knowledge.wharton.upenn.edu/article/the-real-threat-to-business-schools-from-artificial-intelligence/.
2. Davenport, T. H. and Katyal, V., 2018, “Every leader’s guide to the ethics of AI,” MIT Sloan Management Review.
3. Murray, S., 2016, “Harvard Business School is teaching MBAs about artificial intelligence, deep learning –­ here’s why,” BusinessBecause, https://www.businessbecause.com/news/mba-degree/4100/harvard-business-school-is-teaching-mbas-about-ai.
4. Dwyer‐Matzky, K., Pachamanova, D. and Tilson, V., 2021, “Accounting for capacity: A real‐time optimization approach to managing observation unit utilization,” Naval Research Logistics, special issue on Service Operations Management, forthcoming.
5. Crenshaw, L. A., Lindsell, C. J., Storrow, A. B. and Lyons, M. S., 2006, “An evaluation of emergency physician selection of observation unit patients,” The American Journal of Emergency Medicine, Vol. 24, No. 3, pp. 271-279.
6. Conley, J., Bohan, J. S. and Baugh, C. W., 2017, “The establishment and management of an observation unit,” Emergency Medicine Clinics, Vol. 35, No. 3, pp. 519-533.
7. EMC Education Services, 2015, “Data Science and Big Data Analytics: Discovering, Analyzing, Visualizing and Presenting Data,” Wiley.
8. Morgan, L., 2018, “Why operationalizing analytics is so difficult,” Information Week, March 2, https://www.informationweek.com/big-data/big-data-analytics/why-operationalizing-analytics-is-so-difficult/a/d-id/1331168.
9. Kotter, J. P., 1995, “Leading change: Why transformation efforts fail,” Harvard Business Review, March-April.
10. https://bit.ly/3rTelAq