Emergency Room Efficiency with Chess Engines

The Chessable Research Awards for the Spring 2023 cycle had two winners, undergraduate student Sarah Kudron and graduate student Adam DeHollander. Applications for the Fall 2023 cycle open May 1, 2023.

In this blog post, Adam DeHollander reports on his research about how chess engines might make hospital emergency rooms more efficient. He begins with his own experience in an emergency room.

– Dr. Alexey Root

During the summer of 2021, I was a patient in a hospital emergency room (ER) where I witnessed a disturbing scene. The waiting room was overcrowded, with dozens of patients sitting on the floor due to a lack of available chairs. What was even more concerning was the fact that many of these patients appeared quite ill and many were elderly.

ER crowding has been a growing problem even before the COVID-19 pandemic. For example, in 2016 over 90% of ERs in the US regularly experience crowded conditions. ER crowding is much more serious than long wait times; crowding causes adverse health outcomes and even increased deaths.

Researchers have been working to improve ER efficiency for decades using a variety of methods, such as machine learning and simulation. However, there are multiple limitations of the current studies. First, when researchers choose to improve a certain variable, it usually causes other variables to get worse. For example, in chess imagine both sides have castled on the same side of the board. If you start moving the pawns in front of your king towards your opponent’s king, then there are some pros and cons to this decision. It may help you create an attack which means you are improving your position. However, it also makes weaknesses which makes it is easier for your opponent to attack you. If you look at this plan from an attacking perspective it is a good move, whereas from a defensive perspective it is a bad move. The same concept occurs in the ER. If researchers improve the waiting time for one group of patients, it usually causes another group of patients to spend more time waiting.

Another major problem with the current studies is they often have a difficult time accounting for uncertainty. For example, let’s say you are coming up with a plan in chess. In this plan you assume your opponent is going to castle and make some simple development moves. If your opponent makes these moves, then your plan will work well. However, instead your opponent makes a brilliant piece sacrifice that you didn’t anticipate, and then you lose the game. Even though you had a seemingly great plan, if something happens that you did not expect, it can cause major problems. In the ER there are numerous uncertain events that can occur such as new patients arriving, a patient requiring more tests, and tasks taking longer than expected.

The current methods used in the literature are not able to optimize multiple variables simultaneously or account for uncertainty, however, chess engines, such as Stockfish or AlphaZero, are able to perform these tasks. When a chess engine evaluates potential moves, it considers multiple variables such as number of pieces, king safety, control of the center, and so on. It also thinks many moves ahead to consider all future events that could occur. Therefore, I wondered if the same technology that is used to play chess could also be used to make the ER more efficient.

To apply chess engines to make the ER more efficient, one first needs to convert ER language into chess language. The nurses and physicians become the pieces. The ER is the board. Every decision that is made in the ER becomes a player’s moves. Long wait times equate to losing the game.

For example, let’s analyze a position in the “ER game”. Say there are ten patients waiting to see a physician which means you have ten possible moves to choose from in this position. Move one would be letting the first patient in line get to see the physician next. But it may be better to play move two, because the second patient is a lot sicker than patient one, so it is important to treat them quickly. Decisions like this can be made by looking at the current information, however, we may need to look ahead multiple moves to make a good decision. For example, while move two currently looks good, if we look ahead we may see that patient two would require at least forty-five minutes with the physician, whereas each of the other patients could be completed within five minutes. If we look ahead and see a lot of other patients are going to need to see this physician soon, it may be a bad idea to perform a long forty-five minute procedure. Instead, it may be good for the physician to see multiple quick, easy patients first.

One of the most difficult aspects of relating the ER to chess is evaluating the position. In chess the game always ends and there are three possible results—either White wins, Black wins, or it is a draw. In the ER the game never ends and there is not a simple win/loss/draw result. Instead, I developed a complex set of equations that are used to evaluate the ER position. A major advantage of the evaluation function I developed is that it considers the current position and it incorporates future information. This idea is also used when evaluating a chess position.

For example, the simplest method of evaluating a chess position is counting how many points you have (pawns are one, knights are three, and so on). In the current position, let’s say the points are equal and your opponent then uses their queen to take your queen. Now how do you evaluate the position? If you simply count up the points currently on the board, then your opponent would be ahead by nine points. But assume you have a move you can play where you take their queen. In that case, your opponent isn’t actually ahead by nine points, because if you consider future states when evaluating the position, then you will see the game is still equal. In the ER evaluation this is done by taking into account how long each patient has currently waited and also predicting how long they will need to wait in their future tasks. The evaluation function also looks at how many patients are in the waiting room while also predicting future information such as new patient arrivals and estimating how long it will take until a bed becomes available.

While the ideas of this research have been developed, it takes a long time to implement these ideas. Chess engines like Stockfish have taken years to develop. It is not possible to directly solve the ER decisions using an existing chess engine due to some key differences between chess and the ER. Currently I have developed an ER simulation which represents the ER and allows one to develop the technology to play the “ER game” like a chess engine. I will start by testing the ER game engine in this virtual simulation to ensure it makes high-quality moves. After illustrating the merit of this research through simulation, the goal is for the game engine to be implemented in ERs in real-time. Anytime a decision would need to be made in the ER, the game engine would analyze the position and quickly output the move it would play. Additionally, it would explain why it choose that move and it would show the evaluation of each alternative move. Importantly though, the game engine would not actually make the move. The hospital staff would look at the move the game engine suggested and the reasons for choosing that move, but then the hospital staff would make the final decision. This approach allows humans to continue making decisions in the ER. The difference is this research will provide the hospital staff with accurate evaluations of each action from an efficiency perspective. The staff can then combine their intuition with these suggestions to make more efficient decisions.

In 1996, Deep Blue became the first computer to beat a world champion. Since then chess engines have not only rapidly grew stronger, but they have also became a highly utilized asset among chess players. Chess engines are used to analyze games, develop openings, and play training games. Chess engines allow all chess players to have a world-class chess player at their disposal to help them improve their game. My research will hopefully inspire the same revolution to occur in ER decision-making. An ER game engine won’t just allow us to win the game, it will also help us save lives.

Applications for the Fall 2023 cycle of the Chessable Research Awards open May 1, 2023. The Chessable Research Awards are for undergraduate and graduate students conducting university-level chess research. Chess-themed topics may be submitted for consideration and ongoing or new research is eligible. Each student must have a faculty research sponsor. 

Each winning undergraduate student gets $500, and their faculty research sponsor also gets $500. Each winning graduate student gets $1,000, and their faculty research sponsor gets $500. There are three cycles of Chessable Research Awards given each year. For more information, or to apply, please see the Chessable Research Awards page.

You May Also Like