The Case For Mathematical Optimization In Health Care: Building A Strong Foundation For Artificial Intelligence

“Enthusiasm for the potential impact of AI [artificial intelligence] on hospital operations is often based on its impact in other industries. However, non-health care companies invest in AI after having digitized and optimized their operations with a variety of older mathematical methods. In contrast, hospitals may invest in AI while still scheduling patient appointments using fax machines and allocating resources based largely on anecdotal experience.

[..] For decades before the advent of modern AI, the operational management of capital, labor, and resources of large manufacturing, retail, airline, and most other large-scale industries were being designed and refined with rigorous methods such as mathematical optimization. Optimization is a family of algorithms that allocate resources while minimizing costs or maximizing benefits in the presence of constraints. Airlines use such algorithms to create efficient staff and flight schedules; Amazon uses them to route packages and minimize the cost of on-time delivery; and Google Maps uses them to find the shortest travel time between two locations.

For such mature organizations, the continued use of optimization after decades of refinement allows only marginal improvement. Thus, the use of AI offers productivity gains for problems to which optimization does not apply. Airlines use it to predict which passengers won’t show up; Amazon and Google Maps use AI to create personalized recommendations for each customer. On the other hand, the delivery of health care, in particular the operation of hospitals, lags decades behind in the use of mathematical optimization. This raises the question: In an industry where 30-year-old fax machines are routinely used but 30-year-old optimization methods are not, will the use of cutting-edge AI reap the same benefits as in more operationally mature industries?

[..] Our hypothesis is that AI and ML [machine learning] have grown more quickly in health care not only because they are easier to conceptualize but also because siloed hospital data systems are better set up to deliver data for AI problems than optimization problems. We illustrate this with two problems from hospital operations based on our own experience: improving predictions of surgical procedures durations with AI/ML and reducing delays in admission to the post anesthesia care unit (PACU) with optimization.

Although the technical details of AI and optimization are obscure for most health care providers, what AI models do is intuitive: They make diagnoses, clinical predictions, or operational predictions. What optimization models do is more difficult to explain and conceptualize. Although their uses are relatively routine and include improving hospital resource planning, improving surgery schedules, and efficiently managing capacity for emergency procedures, these are general statements for which the precise description requires familiarity with ideas such as minimizing an objective function and formalizing constraints. Furthermore, large companies such as Apple, IBM, and Amazon have invested heavily in marketing to make AI familiar and tangible, with technologies such as Siri, Watson, and Alexa. Few such public marketing campaigns exist for optimization.

Even for a technical audience, the conceptualization of AI is more straightforward than that of optimization. Defining and solving problems in AI, for example, defining a supervised learning problem, typically involve narrow sets of relatively structured data with one desired outcome or result. For example, a supervised learning model can be trained on historical surgical patient data to determine procedure duration. Defining an optimization problem, however, requires more robust data collection, since this model translates real-world objectives and constraints into a mathematical model, typically with broader, vaguer goals than those of a problem in AI. For example, mathematically defining the goals of PACU bed congestion reduction requires a great deal of initial research to begin framing the model.

[..] There are several major challenges in health care akin to those in the airline industry that could benefit from a similar “optimization first” methodology. Such challenges impact organizations operationally, clinically, and financially. Operationally, optimization can be used to maximize time used in the operating room or match nursing skills to the appropriate patient cases for the day. Clinically, optimization can address population health problems such as matching organ donors and receivers or designing radiation treatment plans that minimize harm to the patient. Financially, optimization can specify how to allocate funds to various service lines of a hospital.

[..] Groundbreaking advances in AI will have a significantly larger impact when deployed in an operational model that has been optimized. Three closely related current challenges to deriving value from AI models are that they may: rely on data that are unstructured and dispersed across multiple platforms, be designed without a clear understanding of how much value their use will generate, or make predictions upon which health care systems may have little ability or willingness to act or that may suffer from unintentional bias. For each of these reasons, institutions may benefit first from structural changes to systems based on optimization and culture change before attempting to reap the benefits of the applications of AI.”

Full post, Reddy A and Scheinker D. Health Affairs Blog, 2020.11.13