“Premature optimization is the root of all evil.”

Donald Knuth, Stanford computer scientist

A fellow AI enthusiast and I had a robust discussion this week about the “optimization problem” in artificial intelligence as it relates to healthcare. Optimization as a concept in AI parlance is defined as problem solving and planning in an iterative trial-and-error process to achieve a defined goal. In optimization for games, for example, the goal can easily be the maximum score possible. Another well known problem or challenge in optimization is the traveling salesman problem.

Reinforcement learning is a method used in optimization algorithms and this optimization is used in sales by optimizing prices and inventory. In the recent era, DeepMind of Google that was deployed to defeat the human Go champion is an example of “deep” reinforcement learning with a component of generative adversarial networks (GANs). In essence, it is reinforcement learning combined with deep learning. In addition, what was impressive about AlphaGo and DeepMind was that it could forego short term gains in order to have a longer term win (which often is the case in clinical medicine).

In healthcare, one example of this optimization approach is use in revenue cycle management (RCM). This tedious RCM process involves the use of software for each of the steps of RCM that ranges from initial scheduling to final payment. Artificial intelligence in the form of robotic process automation can be used to evaluate the myriad of variables in each step of RCM and return the most “optimal” result. For instance, AI can help identify reasons why a claim was denied and devise a strategy to avoid that set of circumstances that led to a claim denial. An optimization strategy rather than a single deployment would have been very useful during the early days of the pandemic when supply chain and resource allocation were very challenging.

A significant issue in optimization in clinical decisions is the “fuzzy” nature of maximum score. For instance, is six months of relatively pain-free life better than two years of pain-filled therapy for cancer? Much of this “maximum score” is dependent on the individual as well as timing so there is no universal maximum score as exists in games. Optimization in clinical medicine, therefore, will need to be “precise” for the individual and condition at that specific time.