Imagine you go see a patient in the urgent care who presents with shortness of breath, wheezing, but is otherwise well. You do a full exam and assessment, an EKG, a CXR, and you find nothing wrong.
The O2 level was slightly low at first but the patient had a past history of smoking and after some breathing treatments the SOB improves and you can’t hear the wheezing. You’re finishing up the note to discharge the patient and you get a pop-up on your EMR:
Have you considered Pulmonary Embolism in this patient? If it’s a possibility, measure the patient’s ambulatory SO2. Based on the current information on this patient, she has a 73% chance of having a PE.
You tell your nurse to ambulate the patient and, sure enough, they desat down to 91%. You go back to your computer and there is an assessment tool for PE and the pop-up in your EMR even tells you what steps to consider next.
How did your EMR know this? It’s the same EMR you’ve used for a decade.
What you don’t know is that your medical group recently hired a company who has installed this add-on tool to help catch commonly missed diagnoses. The pilot program include PE, aortic dissection, and Lyme disease.
You didn’t have to do anything different. You didn’t even know this pop-up window existed. But the add-on software has been screening everything you wrote and dictated. It was looking through the patient’s history and collecting other information from nursing notes.
Because artificial intelligence, machine learning, big data, and deep learning are becoming hot topics, the exact definitions are fiercely debated. I’ll explain the terms as I go along but understand that I have no intention on being scientifically accurate with this post.
Over the years, as technology has advanced, the term AI has morphed and so for something to be considered artificial intelligence, it has to display some autonomy in analyzing information and displaying outputs to the user.
That little pop-up was an AI tool which was added on to the EMR.
So is a Well’s Criteria calculator popup also an artificial intelligence tool? Maybe in the early 90’s. But, now, the bar is set higher and a static calculator gets a less sexy name, a ‘helper tool’.
There is far more to that little Pulmonary Embolism popup than its ability to interpret the text you entered in the EMR. Let’s talk about machine learning.
Let’s go one step further from artificial intelligence. Let’s dive down into machine learning. This is the concept that a program or software can learn from the information it collects.
It’s more than just recognizing patterns and recalling associated diseases. If that’s all the PE pop-up did, it would be nothing more than a helper tool; just a little calculator, of sorts.
Machine learning requires the program to sift through thousands of research articles, patient medical records in the EMR, historical PE cases, and PE outcomes.
So, the more data can be mined by this software, the more patterns will be recognized. In order for this to happen complicated algorithms need to be executed which happen in the background.
So that’s amazing already. This 3rd party software is able to sift through a ton of shit and then, based on the symptoms you entered into the EMR, it can recommend what to do next and what the chance of PE is. Great, is that it? Oh, fuck no.
There is one more thing that was programmed into that disease recognition pop-up, deep learning. Think of deep learning as the next level of data analysis. Think of it as the 6th sense a doctor develops after practicing medicine for a decade.
More importantly, deep learning isn’t static, it grows. That 3rd party software not only looks at all the information you provided but it follows every patient who was eventually diagnosed with a pulmonary embolism. It goes back in time and sifts through the symptoms and presentations and it comes up with novel ways of determining relationships.
It might, for example, recognize that 60% of patients who obtained an eventual PE diagnosis went to their PCP for a cough the days prior. It can even look to see which patients had no improvement from medical intervention and which ones were at highest risk for complications.
Because of this perpetual learning, 3 months from that last visit you had with the PE patient, if you see another potential PE patient, the popup will give you a far better accuracy than just 73%. It can even tell you if it’s worthwhile getting a CTA right away or if the symptoms and presentation are too mild to warrant immediate workup.
The same software, when deployed in the pulmonologist’s office, can come up with the right statistics to help that doctor decide if heparin infusion is the appropriate management for that particular PE patient.
The DNA of Artificial Intelligence
No doubt that the human body is fucking amazing. We can learn new shit and invent stuff and feel emotions and get stronger. But at the very core of our being, all there is DNA.
Software, no matter how sophisticated, is just code typed into a computer. Someone has to come up with that code, someone has to program in those algorithms.
For that exact reason, doctors won’t be replaced with artificial intelligence. Instead, AI can be used to make some advances in medicine, of which we haven’t had a whole lot of in the past few decades.
Clinicians are needed to bridge the knowledge gap. For centuries medicine has been an apprenticeship. We can teach another doctor how to do something, how to recognize something, but we don’t know how to relate the same knowledge to a software engineer.
That, right there, is a huge opportunity and gap in technological advancement of medicine. If I can learn how to communicate with a programmer by translating my knowledge into statistical values or visually build my decision tree, I can help improve the programming of that machine learning algorithm.
Jobs for Medical Professionals
Data science will be an interesting field in healthcare over the coming years. Software engineers and health scientists are taking on some of the work to help ‘train’ these programs. But there is only so much someone can do who doesn’t see patients.
Before deep learning can be programmed or any kind of machine learning designed, information has to be translated from a clinician’s mind into written code.
Then you need clinicians who can supervise these algorithms and point it in the right direction. You need clinicians who can test the system and force it to stumble.
It’s eventually going to come down to the edge cases. The edge cases, like the pneumothorax in the undiagnosed Marfan Syndrome patient, are the cases which programmers and clinicians will have to spend the most amount of time one.
Eventually, every clinician will have to learn how to communicate with artificial intelligence programs. It won’t be some futuristic shit that’ll take over medicine. Instead, it’ll feel natural and happen gradually, much like the way we learned how to ask the right questions in a web search to get the desired result – ‘keyword search’.
Every startup is focused on big data. They want to sift through tons and tons of data to come up with something spectacular, which they hopefully will.
But nobody has done something obvious and simple. Nobody has built a $0.99 app which can help you decide if you need to go to the doctor for your URI symptoms.
Imagine how many patients would use such an app in a single day of a cold and flu season. We’re talking millions of dollars of revenue. Even after they get the results, they can go to their doctor and be seen. Or they can save their copay.
And the software doesn’t have to be unsupervised. With that kind of revenue, each and every interaction could be checked by a clinician to make sure nothing falls through the cracks and that it’s all on the up-and-up.
The opportunity for entrepreneurship is huge in this space. You can focus on acute diseases, on chronic diseases, on post-op cases, on pre-op cases, on substance abusers, on those with depression, on insomnia patients, on IBD patients, or, on cancer patients.
There are research programs you can join to become a pioneer in AI in medicine.
There are medical journals dedicated to this new branch of medicine.
There are conferences which try to address the political conflicts of such technological encroachment.
There are specialty specific institutes which are developing their own medical intelligence arms.
And there are a ton of medical startups which are basing their business model off of artificial intelligence.