When I think about traditional market research, I picture 10 people sitting around a table answering questions, eating M&M’s, and being watched through a one-way mirror. I also picture two expert physicians being interviewed about how they care for patients, write (or at least follow) treatment guidelines, and teach fellows in their hospital to practice data-driven medicine.
But that is the traditional approach, and while there will always be value in those methods, we have a responsibility to innovate. We can no longer extrapolate from the opinions of a few to determine the truth. It just doesn’t make sense to use small sample sizes to derive the size of a market, understand how medicine is practiced “out there,” or investigate the natural history of a disease.
With the ever-increasing availability of healthcare data, it seems that we should use this information – always in a HIPAA compliant manner — to augment current market research techniques.
But it is harder than most people think to make use of big data. Everyone says that they do “big data analytics” now, but data mining is not analytics, and it is simply not enough. Analytics is the means to transform “big data” into something meaningful and actionable. We must collect/aggregate the best quality of data and we must apply the best transformation techniques including state-of-the-art machine learning techniques. But that is just the beginning. From there we must extract actionable intelligence, and finally, we must ensure that we can measure the impact of our outputs.
When I think about how to best use analytics in this framework, I think we must
How do we do that with analytics? Let me tell you (briefly) about three projects we completed to predict, identify, and intervene – all to improve the lives of patients.
Amyotrophic Lateral Sclerosis (ALS), as you likely know, is a neurodegenerative disease that only has one outcome. So far there is no cure, but we are starting to have more treatments available. It turns out that those treatments are probably more effective if you give them earlier in the course of disease. But ALS has always been thought to be symptomatic only 12-18 months prior to diagnosis. Our client asked us to see if we could find a way to accelerate diagnosis. So far, we have demonstrated that there is a clear signal – recognizable only through machine learning – as early as five years prior to diagnosis. That’s right: up to 25% of patients have symptoms long before the 12-18 month window. With these analytics, we are going to PREDICT who is moving in the direction of ALS, get them diagnosed sooner, treated earlier, and if all goes well, improve those patients’ course of disease.
Multiple Sclerosis (MS) is a heterogeneous disease which can take years to progress, or can have a much more aggressive course. It can attack our ability to feel, touch, or to use our muscles. There are many treatments, and the market is a crowded, competitive space. We were approached by a client to IDENTIFY patients who were either naïve to treatment, or had lapsed treatment. We did that, but in the process, we also discovered that the client’s long-held belief that the US had ~400,000 diagnosed patients was incorrect. We were able to IDENTIFY ~200,000 additional diagnosed patients, as well as ~100,000 undiagnosed patients. In this case we identified patient subpopulations, as well as a vastly bigger market than had been previously known. We can also use our analytics to identify the right question, the right data set, the right patients, and the right physicians.
Non-alcoholic steatohepatits (NASH) presents a challenge in that a liver biopsy is still the gold standard for determining level of fibrosis. The “holy grail” of hepatology at the moment is to find biomarkers that accurately define the level of fibrosis. Though there are scales like APRI, Fib4 and MELD, they are not yet in a position to replace the liver biopsy. That makes the use of big data tricky since many large claims databases still don’t have associated EMR data, and therefore biopsy results. We were asked by a client to see if we could build a “claims-equivalent” to identify stage of disease, or approximate level of fibrosis. We did just that, and we also demonstrated that the specialists the client thought would be caring for their patients were in fact not. In this case, we were able to guide the client how to INTERVENE at the right point in the patient’s course of disease, as well as at the right provider who cares for the patient.
There is one last thing to mention. We have found that we need physicians and data scientists working together to ensure that we analyze healthcare data in a meaningful way, and extract measurable value. It is critical, in this age of burgeoning healthcare data, that we create our teams from people who know healthcare and medicine, and people who know data analytics. With those experts working together, we will be more successful in this process I have described today: Predict. Identify. Intervene.