It’s safe to say there are too many manual processes in medicine. While in training, I hand wrote lab values, diagnoses, and other chart notes on paper. I always knew this was an area in which technology could help improve my workflow and hoped it would also improve patient care. Since then, advancements in electronical medical records have been remarkable, but the information they provide is not much better than the old paper charts they replaced. If technology is to improve care in the future, then the electronic information provided to doctors needs to be enhanced by the power of analytics and machine learning.
Using these types of advanced analytics, we can provide better information to doctors at the point of patient care. Having easy access to the blood pressure and other vital signs when I see my patient is routine and expected. Imagine how much more useful it would be if I was also shown my patient’s risk for stoke, coronary artery disease, and kidney failure based on the last 50 blood pressure readings, lab test results, race, gender, family history, socioeconomic status, and latest clinical trial data.
We need to advance more information to clinicians so they can make better decisions about patient diagnoses and treatment options, while understanding the possible outcomes and cost for each one. The value of machine learning in healthcare is its ability to process huge datasets beyond the scope of human capability, and then reliably convert analysis of that data into clinical insights that aid physicians in planning and providing care, ultimately leading to better outcomes, lower costs of care, and increased patient satisfaction.
Applied Machine Learning in Healthcare
Machine learning in medicine has recently made headlines. Google has developed a machine learning algorithm to help identify cancerous tumors on mammograms. Stanford is using a deep learning algorithm to identify skin cancer. A recent JAMA article reported the results of a deep machine-learning algorithm that was able to diagnose diabetic retinopathy in retinal images. It’s clear that machine learning puts another arrow in the quiver of clinical decision making.
Still, machine learning lends itself to some processes better than others. Algorithms can provide immediate benefit to disciplines with processes that are reproducible or standardized. Also, those with large image datasets, such as radiology, cardiology, and pathology, are strong candidates. Machine learning can be trained to look at images, identify abnormalities, and point to areas that need attention, thus improving the accuracy of all these processes. Long term, machine learning will benefit the family practitioner or internist at the bedside. Machine learning can offer an objective opinion to improve efficiency, reliability, and accuracy.
At Health Catalyst, we use a proprietary platform to analyze data, and loop it back in real time to physicians to aid in clinical decision making. At the same time a physician sees a patient and enters symptoms, data, and test results into the EMR, there’s machine learning behind the scenes looking at everything about that patient, and prompting the doctor with useful information for making a diagnosis, ordering a test, or suggesting a preventive screening. Long term, the capabilities will reach into all aspects of medicine as we get more useable, better integrated data. We’ll be able to incorporate bigger sets of data that can be analyzed and compared in real time to provide all kinds of information to the provider and patient.
The Ethics of Using Algorithms in Healthcare
It’s been said before that the best machine learning tool in healthcare is the doctor’s brain. Could there be a tendency for physicians to view machine learning as an unwanted second opinion? At one point, autoworkers feared that robotics would eliminate their jobs. Similarly, there may be physicians who fear that machine learning is the beginning of a process that could render them obsolete. But it’s the art of medicine that can never be replaced. Patients will always need the human touch, and the caring and compassionate relationship with the people who deliver care. Neither machine learning, nor any other future technologies in medicine, will eliminate this, but will become tools that clinicians use to improve ongoing care.
The focus should be on how to use machine learning to augment patient care. For example, if I’m testing a patient for cancer, then I want the highest-quality biopsy results I can possibly get. A machine learning algorithm that can review the pathology slides and assist the pathologist with a diagnosis, is valuable. If I can get the results in a fraction of the time with an identical degree of accuracy, then, ultimately, this is going to improve patient care and satisfaction (I write this as my own mother has been anxiously awaiting her own test results for over a week).
Healthcare needs to move from thinking of machine learning as a futuristic concept to seeing it as a real-world tool that can be deployed today. If machine learning is to have a role in healthcare, then we must take an incremental approach. We must find specific use cases in which machine learning’s capabilities provide value from a specific technological application (e.g., Google and Stanford). This will be a step-by-step pathway to incorporating more analytics, machine learning, and predictive algorithms into everyday clinical practice.
Initially, our goals need to match our capabilities. Training a machine learning algorithm to identify skin cancer from a large set of skin cancer images is something that most people understand. If we were to learn that radiologists are being replaced by algorithms, then people would be understandably hesitant. This must be bridged over time. Radiologists won’t ever become obsolete, but radiologists of the future will supervise and review readings that have been initially read by a machine. They will employ machine learning like a collaborative partner that identifies specific areas of focus, illuminates noise, and helps focus on high probability areas of concern.
How do we reach the threshold needed to trust machine learning? Medicine has a method for investigating and proving that treatments are safe and effective. It’s a long process of trial and error—and basing decisions on evidence. We need these same processes in place as we look at machine learning to ensure its safety and efficacy. We need to understand the ethics involved in handing over part of what we do to a machine.
A “What If” Scenario on the Potential of Machine Learning
A few months ago, I gave a presentation about the future of analytics and its potential impact on clinical care. In my slides, I showed a hypothetical EMR running predictive algorithms while a doctor was examining his patient. A pop-up box displayed the real-time diagnosis, pathology results, and treatment options, as well as each option’s potential effectiveness and cost for this patient.
While the patient in this case may have been hypothetical, it was modeled after my father who passed away several years ago, from prostate cancer. I chose this scenario to demonstrate outcomes that could have been possible had machine learning been available at the time.
In my dad’s actual case, his doctor initially gave him two years to live. This was based on a combination of the doctor’s experience with similar patients and the treatment options available at that time. What he did not know was that I was going to take an active role overseeing my dad’s care. I researched clinical trials and new treatment options. I studied side effects and trial results. Were treatments keeping people alive longer? If so, was it for a few weeks, a few months, or longer?
My dad had a great oncologist, but he was caring for thousands of patients with many different kinds of cancer. He could never have put in the time and effort needed to learn all the new drugs and treatment options coming out for all these cancers. Many times, I presented treatment options and clinical trials that my dad’s doctor wasn’t aware of.
With my years of training and expertise, I could cull the literature and recommend the best options for my dad. In other words, I was the human algorithm, the doctor’s brain, who had the means and, most importantly, the motivation and time to work in concert with my dad’s physician to develop the optimal plan, which ultimately extended Dad’s life nine years.
With an analytics platform and machine learning running in the background, the human algorithm—the extra layer of a back-up physician—wouldn’t be necessary. The analytics engine would have infinitely more data than any one person could ever process. It would have a library of patients like my dad, with his diagnosis and tissue type. It would have treatment options available with predictions of how long they would be effective, mortality rates, side effects, and cost. Regardless of all the effort by a human caregiver, an analytics platform could put in infinitely more work behind the scenes and deliver decisive information to the physician in real time.
Data Drives Machine Learning
As more data is available, we have better information to provide patients. Predictive algorithms and machine learning can give us a better predictive model of mortality that doctors can use to educate patients.
But machine learning needs a certain amount of data to generate an effective algorithm. Much of machine learning will initially come from organizations with big datasets. Health Catalyst is developing Collective Analytics for Excellence (CAFÉ™), an application built on a national de-identified repository of healthcare data from enterprise data warehouses (EDWs) and third-party data sources. It is enabling comparative effectiveness, research, and producing unique, powerful machine learning algorithms. CAFÉ provides a collaboration among our healthcare system partners, big and small.
Another possibility for smaller entities will be their ability to merge their data with larger systems. At some point, we may see regional data hubs with datasets customized for geographical, environmental, and socioeconomic factors, that give healthcare systems of all sizes access to more data.
As larger datasets begin to run machine learning, we can improve care in more specific ways for each region. And considering rare diseases with low data volumes, it should be possible to merge regional data into national sets to scale the volume needed for machine learning.
Let’s Move Machine Learning from Theoretical to Clinical Reality
We already see applications of machine learning in healthcare that are advancing medicine into a new realm. It’s exciting to think about where it can go. Someday, it will be commonplace to have embedded machine learning expertise that analyzes not only what’s going on with patients in real time, but also what’s going on with similar patients in multiple healthcare systems, what applicable clinical trials are underway, and the efficacy and cost of new treatment options. It may sound futuristic, but the analytics engine that can present all this information at the point of care is available now.