Machine Learning in Medicine
Let me tell you about a real patient, but with a different name. Let’s call her Cheryl. Cheryl came to the emergency room on a sunny Tuesday. Her electronic health record says she's 52 years old. She came to this emergency room in Maryland with a foot sore. Physicians investigated the foot sore and she ended up staying in the hospital for 22 days. Here's what happened.
Day 1 She came to the emergency room for a foot sore. The physicians inspected her and they saw no real reason for medical concern. They decided to monitor her in case her foot sore was infected. So they put her in the general ward.
Day 3 Cheryl starts developing symptoms of what looks like mild pneumonia. They give her the usual treatment of antibiotics and all is well. But then her condition starts to worsen.
Day 6 Cheryl develops what's called Tachycardia. That means in medical speak her heart rhythm has accelerated dramatically. She then has trouble breathing.
Day 7 Cheryl experiences septic shock. That means her body is in crisis.
Incidentally, mortality in septic shock is one in two. Now it's only at this point that physicians get really concerned and they transfer her to the intensive care unit (ICU). The ICU is where the most critically ill patients get cared for. While here, physicians give her every possible treatment to stabilize her. But her condition only worsens. First her kidneys start to fail. Then her lungs fail. And on day 22 she dies.
Cheryl did receive the right set of treatments. The problem is she received them only too late. What Cheryl experienced was an infection that turned into sepsis. Let me tell you a little bit about what sepsis is, sepsis occurs when an infection releases chemicals in your blood to tackle the infection – so the body releases chemicals to fight the infection. Now this chemical can trigger a negative inflammatory response. What it can then do is cause a cascade of changes leading your organs to fail, leading to death.
Sepsis is the 11th leading cause of death. More than breast cancer and prostate cancer combined. Sepsis is preventable, if treated early. So than why did Cheryl die? Physicians find it very hard to recognize sepsis. In fact, a Harvard study shows with 93 leading academic experts that when they were given seven cases of patients with and without sepsis, they couldn't agree.
This problem of timely treating sepsis was tackled by leading machine learning experts to see if sepsis can be identified and treated earlier. So let me explain – what is machine learning?
A good start at a Machine Learning (ML) definition is that it is a core sub-area of Artificial Intelligence (AI). ML applications learn from experience (data) like humans without direct programming. When exposed to new data, these applications learn, grow, change, and develop by themselves. In other words, with Machine Learning, computers find insightful information without being told where to look. Instead, they do this by leveraging algorithms that learn from data in an iterative process.
Contrary to popular belief, machine learning is not a magical box of magic, nor is it the reason for $30bn in VC funding. At its core, machine learning is just a thing-labeler, taking your description of something and telling you what label it should get. Which sounds much less interesting than what you read in the news.
So patients like Cheryl rallied Machine Learning experts at Hopkins to design what they called Targeted Real time Early Warning Score, or TREWScore.
It’s goal is to identify sepsis as quickly as possible. And since it is based on Machine Learning, it is using data from past patients. This avoids the need for tools to have to experiment on new patients. Also, the past ten years have ushered in the introduction of electronic health records capturing every single measurement, every single lab test that is ever done when you walk into the clinic or hospital. TREWScore analyzes this data. From thousands of patients it tries to identify subtle signs and symptoms that appear in patients with sepsis.
What TREWScore also needs to do is to figure out how to think about every signal in the context of every other signal. Let me give you an example. Let's look at the example of creatinine, a waste molecule that your kidneys filter out. So when your body has sepsis it affects your kidneys. It deteriorates your kidneys ability to filter out creatinine, so creatinine level rises. But there are many other things that can affect your kidneys ability to filter it out. For example, if you have chronic kidney disease you're very likely to have high creatinine levels. So now what TREWScore has to do is to figure out if your creatinine high because of sepsis or because of chronic kidney disease or the numerous other factors that lead to higher levels. But that's not enough, you need to do this for every single signal that exists in the electronic health record and map out every signal in the context of every other signal to identify signs and symptoms that occur more often in patients with sepsis than those without.
Let's return to Cheryl.
Research by Sauerwein et al have shown that for every hour treatment is delayed in sepsis, mortality goes up by 4 percent. The timing for sepsis treatment is critical. Suchi Saria from Johns Hopkins went and took Cheryl’s data and ran TREWScore on it. TREWScore would have diagnosed the sepsis 12 hours before the physicians did. Many would say that is the difference between life and death.
Last year 6000 patients’ data was shown to TREWScore. On average, TREWScore would have detected most patients’ sepsis on average 24 hours prior to the sharp onset. In two thirds of these patients this sepsis was detected prior to any organ dysfunction whatsoever. And to put this result in context, that is a 60 percent increase in performance over what was state of the art technique by physicians. So what TREWScore is really doing is giving physicians a much longer window to come in and intervene in order to prevent organ dysfunction and mortality.
TREWScore is still being validated and attempts are being made to do real time integration so every physician has access. What this will do is give physicians a second pair of reliable eyes. It's hard to scale up physicians. It is much easier to scale up computers and tools that represent the best expertise from the best physicians everywhere.
So to summarize, sepsis is one preventable killer in many pressing medical problems like described above in sepsis. The answers for knowing whom to treat, when to treat and what to treat it might already be in the data – and machine learning is making this happen.