One Physician’s Journey Finding the Power in Healthcare Analytics
Dr. Tara Grabowsky, shares a passionate call to “end misdiagnosis”
I always knew I would be a doctor. I skinned my knee at 7, and took notes on the changing scab day after day. My father had a heart attack when I was 11. I got to know the inside of a hospital. I loved the concept of humans evolving from apes. I was told that as a girl, I “should” do science because I was good at it. Be a role model. In high school I took that aspiration to a whole new height – I wanted to be an astromed – a physician practicing medicine on a space station. It sounds silly now when I reflect back on that. I went to Catholic girls school. I was told that I should live a life of service. Science and service. I went to medical school.
But that’s not the whole story is it? I went to medical school to combine my drive to serve and my love of science. But then I saw true human suffering, and was at an age to comprehend it. I watched a 32 year-old English teacher die of metastatic melanoma. I couldn’t find a balance between the science/complexity of oncology that I loved and the doctor – the caregiver- that I wanted to be. So I pulled back – to primary care medicine. And I LOVED my patients. Phyllis, who was 97 when I was caring for her, came in with a blood pressure of 240/110 every time I saw her. For those of you who don’t know what that means, I will tell you. It means stroke, right then, in my exam room. She was on six blood pressure medications. I REALLY wanted to talk about her blood pressure, but she didn’t. She just wanted to talk about how cute she thought my toes were. Lisa had Down’s. She made it to 52. Being thanked by the minister at her funeral was one of the great honors of my life. I still have a watercolor she painted in my office. There are so many more – Stacy, battling bulimia, Gretchen, finding a new definition of beauty during menopause, Gary being wheeled out of my office to the ER in new onset atrial flutter.
But I always thought there was more. More science, more research. More impact. Something new. Something different. Something bigger than caring for astronauts on a space station.
The hardest thing about being a physician is not necessarily losing a patient. Sometimes that can be the most beautiful part of practice. I was witness to terrible deaths – the chaos of people running around, chest compressions, too many drugs. And I was part of serene deaths steeped in love and family and gratitude. So no – the hardest thing is not necessarily losing a patient. The hardest thing for this physician was the fear of missing a diagnosis. The fear of shortening someone’s life. The fear of prolonging suffering.
This is what I want to solve. I want to end misdiagnosis – telling a patient he has multiple sclerosis when it is really ALS. I want to end missed diagnoses – ignoring the increasing weakness and telling the patient she is simply anxious.
This is not a small goal, but we are on our way. Machine learning and AI is all the rage in 2018. But we have to remember that some people – for example the defense and intelligence communities – have been using and developing these capabilities for 4 decades. So – what is a foray into the future in healthcare is reaching into the past in the defense world. We have repurposed those analytic capabilities and jumpstarted the future of healthcare analytics. Why would we start from scratch when we invented support vector machine learning almost 25 years ago? We already know how to integrate disparate data systems. We already know how to transfer learnings from one data set like claims over to other data sets with notes, images, and genomics. What is new is that we have moved that over to healthcare data and sat physicians at the table with satellite engineers, information theoreticians, and image scientists.
At HVH, we are using these capabilities to find patients with rare diseases before they are diagnosed. We have done this in ALS, Hunter Syndrome, and Pompe disease, just to name a few. We are able to predict who will develop Ankylosing Spondylitis. We are proactively altering the course of disease – we are predicting who will have the aggressive form of the disease and enabling them to start on medication sooner. We are enabling payers to refer patients to proper specialists earlier. We are providing the foundational analytics to support the creation of educational programs to alert physicians a patient who is about to need a change in treatment.
Next, we will turn that around and look at an undiagnosed patient in the ER, and by using their record in the context of ~240M other people, understand what disease the patient has.
But truly, even all of that is old news. There are two things to consider here when contemplating the future. One is the data, the other is the analytics.
The raw material of data is changing. First, there is the volume to consider. Just one test generates a terabyte of data. Moreover, the types of data are evolving. When we started HVH ~4 years ago, we only had claims data readily available. Now, we can work with EMR data, -omics, imaging, pathology, microbiome and lab data. And there is patient-generated data; wearables are generating a whole new realm of information to be captured. But what comes next? We are already moving beyond pure healthcare data to think about the health of a patient. We have to understand the dynamic environment – just like the self-driving car has to take every relevant element into account (laws, obstacles, human safety), so too is the future of healthcare analytics. We are moving toward integrating data on everything that impacts a person’s being. We need to think about social media, where people live, levels of commuter stress, pollutants, social networks, the emotional journey, points of shame, geospatial data, environmental data.
But that’s just the data. You also have to know how to use it. Everyone is talking about machine learning and artificial intelligence in health care. Everyone is buying data in vast quantities. I just don’t think they are getting the most out of it. And this is why: buying the data is just the first step. You can’t just buy the land. You have to install the oil drills, pump the oil, relocate the pumps when they run dry, refine the oil… You need different specialists to build the drills, maintain them, refine the oil… you get the idea.
The key to getting the most out of your data is threefold
- You must have a team that includes both analytic and medical experts. If you just have analysts you get incredible p values, but they tell you that toenail fungus will cure cancer. If you rely on medical knowledge only to answer a question, you lose the power of 240 million patient records. You lose the opportunity to discover something new because you went in with biased clinical expectations
- You must pick your data source – or sources -- based on the business challenge at hand. It does not work to fit the question to the same data source day after day. I simply don’t believe in “garbage in, garbage out.” If the first data source is garbage, then add another, and another. Turn the garbage into treasure. In today’s data landscape it is simply not necessary to accept “garbage out.”
- You must have deep analytic knowledge – I was drawn to this world – and you know it must have been compelling to pull me away from clinical practice – because I had no doubt this was the right formula to forge the path to the future in the field of predictive analytics. Machine learning was a powerful tool in the beginning. But now we are moving into true AI systems. We need to advance the fields of deep learning, the next generation of neural networks, and reinforcement learning. We are moving beyond the concepts of testing, training, and validation data sets. Now the AI is becoming truly autonomous –rewarding and penalizing algorithms so that they continue learning correctly. We must expand these continuous learning processes to be able to handle the increasing types and amounts of data.
So – are there obstacles to all of this? Of course. Data aggregation is not where it needs to be. The quality of data collection, and the regulations surrounding it, are different in every country on the planet. There is much work to be done on the ethics of using these data, the governance surrounding this, and the willingness of people to share their data.
We must keep working, find more data sources to bring into the analyses, continue inventing new methods of ML and AI. We must continue to sit impossibly different specialists next to one another at the table to solve these problems. I hope my next colleagues include roboticists, social anthropologists, public health experts, or a perhaps a field I have never heard of. The point is – we need MANY perspectives to understand something as dynamic as human health.
Do I miss my patients? Yes. I still have the 4-inch stack of their goodbye letters. But we are opening a new era, scripting a new definition of “predictive preventive medicine.” I am not exaggerating when I say that I want to end misdiagnosis.
When I think about it, maybe practicing medicine in space wasn’t such a far out dream after all…