Making Healthcare More Human with Artificial Intelligence

 

As part of the Intel AI Day events late last year, I had the good fortune to moderate a precision medicine panel discussion featuring four healthcare luminaries. This is a topic that is close to my heart, given that I have spent much of the past 10 years focused on building analytics capabilities and technologies that bring transformational value to healthcare.

That’s very much the case with the organizations represented in the Intel AI Day Precision Medicine Panel featuring —Mayo Clinic, Penn Medicine, Kaiser Permanente, and Cigna. All of these organizations are using artificial intelligence (AI) to make healthcare more human via personalization and optimization, leading to improved outcomes for both patients and healthcare systems.

In this post, I will highlight a few of the key themes that bubbled to the surface in this panel discussion. These themes build on topics I explored in a September 2016 post that focused on augmenting human capabilities with AI. That’s really what we are talking about with personalized medicine—augmenting the capabilities of clinicians with data-driven insights. We’re not talking about replacing people with machines. We’re talking about making people better at the things they do best.

Personalizing Healthcare with Help from AI

Let’s start with personalization. There are many ways to use AI to personalize the healthcare experience. One way is pretty straightforward: give clinicians real-time access to all of the clinically relevant information they need to make fully informed care decisions. While this may sound easy to do, in practice this often isn’t the case. In a study carried out by my data science team at Apixio, we found that for patients who have key clinical conditions that a doctor always needs to know about, such as a history of heart attack, blood clots, or hepatitis, the structured problem list is missing this information 63% of the time.

Why so? In some cases, the relevant information isn’t document in the coded data because it has not been used to justify medical billing. In other cases, the condition may have been noted in a doctor’s dictation but never converted to a medical code, resulting in a false negative. Interestingly, structured clinical data is impacted by the converse problem as well: false positives. For example, a shocking 30 to 50 percent of patients who have a diagnosis for heart failure in their problem list do not actually have heart failure.

AI can help here. Machine learning techniques can cluster patients with similar characteristics into natural groupings that indicate the presence of certain conditions. At the same time, machine learning can classify patients based on those who are actual heart failure patients and those who have attributes indicating that they are at an elevated risk of heart failure, essentially calling out situations in which the a patient’s coded data is not consistent with their clinical history. When doctors have this kind of information in real time, they are poised to make better care decisions.

Let’s take another personalization example. For patients who have been in the hospital, an AI system can act as a one-on-one coach to help them with aftercare. We’re not talking about a single coach who dishes out advice for the entire team of healthcare patients in the same manner. We’re talking about a personalized coach who learns from its experiences and tailors its interactions and advice for each patient based on the tactics that are likely to work best for the patient. In data science, this type of coach is known as a “contextually intelligent agent” or an application that understands the context of the subject.

For one patient, the AI system might use a humorous avatar to outline steps to follow to stay healthy in the days after hospitalization. For another, the system might use firm instructions on the exact things to do. And for still another, the system might call for the involvement of family members to help the patient stick to post-hospitalization recommendations. This level of personalization is made possible by having the contextually intelligent agent follow the patient through the continuum of care.

AI Allows for Optimization

Let’s move now to one of the other key themes from the Intel AI Day panel discussion: optimization. Optimization takes many forms in personalized medicine. One form is to use AI capabilities to help care providers use systems and devices more effectively.

Take the example of ultrasound systems, an incredibly important tool in healthcare. To help healthcare providers gain the greatest value from ultrasound systems, AI capabilities can be used to guide the system user to capture the best image and then help the user read the image to gain accurate insights from it. These machine-driven capabilities open the door to the wider and more effective use of ultrasound systems, allowing the machines to be used by EMTs, ER nurses, and other caregivers who don’t have the same level of training as full-time ultrasound technicians.

Another form of optimization leverages a concept called “transfer learning.” This process takes an AI system that has been trained to do one broad task, such as general image recognition, and teaches it to do a more granular task, such as recognition of specific skin conditions.

Researchers at Stanford University did just this when they took an image-recognition algorithm developed by Google and trained it to identify signs of potential skin cancer. Google had trained the algorithm to identify 1.28 million images from 1,000 object categories, teaching it, for example, to differentiate cats from dogs. Building on that foundation, the Stanford researchers retrained the algorithm on a much smaller and more targeted set of images, 130,000 pictures of skin disease. This kind of transfer learning system has a modest data appetite and it doesn’t usually require any specialized computing hardware, so it is broadly accessible to data scientists and application developers.

The result of this effort was pretty amazing. In a paper published in the Jan. 25 issue of the journal Nature, the researchers reported that their model matched the performance of 21 board-certified dermatologists in diagnosing the most common and deadliest skin cancers. This is a great example of the power of deep learning. As they explain in a January 2017 news release, the Stanford researchers didn’t write a program that told a computer what to look for in the skin disease images. Instead, they let the algorithm figure that out for itself—and it did a darn good job at that task.

Helping Healthcare Professionals Succeed

Here again, it’s important to note that we are not talking about replacing medical professionals with AI systems. Instead, we are talking about augmenting their capabilities. In this case, we could potentially enable a wide range of medical professionals to use systems that flag patients with potentially cancerous skin conditions, so those patients can be referred to dermatologists for further investigation.

In short, we’re talking about making healthcare more human with artificial intelligence.

 

Editorial Note:

Bob has recently co-authored a chapter in the book, Demystifying Big Data and Machine Learning for Healthcare, released on January 27, 2017The chapter on machine learning is a guided tour of machine learning in healthcare.  It answers: What are the methods? How are they used?