Dr. Shuhan He, Program Director of the Healthcare Data Analytics program, publishes research with Dr. Pedram Safari, instructor of A.I./machine learning, calling for A.I. to help reduce diagnostic uncertainty 


When the facts are on the table, practicing medicine can be like a game of chess. There’s a right move and a wrong move to make, based on the rules and positioning of one’s pieces.   

But in the world of emergency medicine, where patients are rushed into an examination room with life-threatening conditions and little information to go on, doctors are playing poker.   

Medical poker.  

At least that’s the way Shuhan He, MD, an emergency room doctor at Massachusetts General Hospital sees it. Dr. He, the program director of the MGH Institute’s Healthcare Data Analytics program, is using artificial intelligence (AI) to eradicate as much of the uncertainty as possible.   

“In traditional medicine, you spend time to deeply gather information like vital signs, history of present illness, medication use, or even social history to fully see the board so you can make the right move,” says He. “In emergency medicine, we play poker. We don't have all the cards. We don't have all the information and we have to make a decision - that can be life or death - very quickly, and we need a way to make the right move that gives the most probability of success.’  

In the paper, “Entropy removal of medical diagnostics,“ recently published in Scientific Reports, lead author He argues for the need to reduce entropy – or diagnostic uncertainty – by using AI to help clinicians determine the best course of action. And with medical error blamed for an estimated 98,000 deaths in the United States each year, He says AI can actually help save lives.   

“Someone has chest pain. That could mean a heart attack, or it could mean multiple other things that could all kill you,” said He. “So there needs to be a better way to quantify that uncertainty. There are no tools to tell you ‘Here are the next best guesses.’ That's what we're set out to change.”  

The research is the first look at taking machine learning and data analytics math being used and applying it to medical calculators, which are evidence-based tools to help doctors make decisions. Most clinical decision support (CDS) calculators use information doctors have already obtained, such as, “Did the patient have amnesia for more than 30 minutes? If yes, perform a CAT scan of the head.” However, He and his research partners are creating CDS calculators that take in information a doctor doesn’t have.  

“No diagnostic metrics that specifically measure the reduction of diagnostic uncertainty,” the authors write, “which often leads to decision paralysis and the ‘shotgun’ diagnostic approach, over-testing, delayed diagnosis, and patient harm, have been extensively researched.” 

pedram stands in front of a grove of trees
Dr. Pedram Safari, mathematician and term lecturer in the Healthcare Data Analytics program, helped author the Scientific Reports paper with Dr. Shuhan He that calls for using artificial intelligence to reduce diagnostic uncertainty.

He and his fellow authors, including Dr. Pedram Safari, a mathematician and term lecturer in the Healthcare Data Analytics program, examined the online databases of 533 studies that looked at 623 decision-making tools surrounding 267 diagnoses.   

Most of the time, the machine learning approach showed that the outcome was the same, which is a useful metric for most cases where doctors already measure sensitivity (proportion of people with a disease who have a positive test) and specificity (proportion of people without a disease who have a negative test).   

“However,” added He, “AI and entropy can do additional things like tell us the optimal sequence of tests – and what test should come first – to save time that sensitivity and specificity cannot.”  

Like many facets in life, AI continues to make a growing statement into medicine. 

Entropy is already in use in engineering to reduce information redundancy, increase transmission efficiency, and detecting and correcting error, most notably in telecommunications,” said Safari. “It also helps with pattern recognition in Machine Learning. So, it makes perfects sense to use. Here is good enough reason to make use of it in medicine just the same to reduce diagnostic errors.” 

“It's a shift in computing and also thinking,” said He. “When you say chest pain, it can mean a lot of things. It can mean pain in your chest. It can mean a feeling of an elephant on your chest. It can mean a rash on your chest. There's always the scope of probabilities. So computing is moving towards a probabilistic answer. That's why we're developing these calculator tools to help capture the uncertainty and probabilities which is what we deal with every day in emergency medicine. And that gives us better quantitative tools to solve the extremely hard problems as fast and quickly as possible.”  

He hopes the paper will help the medical community adopt a new way of thinking.   

“We should learn about uncertainty and how to manage it,” said He. “There's not much formal learning of how to play the quantitative game of medical poker and uncertainty management.  

“How do you manage when you don't know all the next best steps, when you don't have the complete information? What's the best move? And that's what I want people to learn in medicine. Every other sector such as finance, mathematics, and telecommunications has formal methods to manage uncertainty. But medicine doesn't have a formal method. In contrast, we generally use a keyword approach. For example, if a patient says, ‘thunderclap headache’ that, therefore, means brain bleed.’  

However, this approach doesn't tell you about all the options that are on the table besides the one linked disease to this keyword. so, this is an opportunity to improve our approach.”

All of which reaffirms the importance of healthcare data analytics, the newest program in the MGH Institute’s School of Healthcare Leadership.   

“Interpretation of data is so important,” said He, the program’s director. “We do it every day with lab values, vital signs, medication dosing, and this is hopefully the first step in many of getting more people in medicine to understand a wider breadth of how numbers can inform medical diagnosis teaching to a more probabilistic thinking.”  

Do you have a story the Office of Strategic Communications should know about? If so, let us know.