The Promises and Limits of Precision Health
There’s no doubt that future of health will include better informed clinical decisions driven by data. Precision health mixes data and machine learning to make highly personalized care possible. Imagine computers figuring out if you have a certain disease from your medical records, or a making diet recommendations by combing through your DNA. It’s an area that One Medical is exploring as we build the future of care.
Through our explorations, several questions came up:
- Clinical Effectiveness: How do you determine what data are useful? Not all data lead to meaningful clinical decisions.
- Data Completeness: Do the right data sets exist? Precision health is only as good as the data that drive it.
- Human Need: Where are the humans in all this? Even if we could, it’s questionable if we should replace our human health providers.
Currently, scores of companies utilizing precision health promise to help consumers avoid injuries, improve diet, and detect diseases. The real impact of some of these services on people’s health is unclear.
Dr. Tamer Fakhouri, a provider I work closely with, has a great example:
Let’s say you have a genetic condition that causes your LDL cholesterol to be 2–3x what a normal person’s is. As part of routine care, you already got tested for cholesterol. Your doctor puts you on a statin and recommends that you eat heart-healthy foods. Does it matter what gene is responsible for your cholesterol? The treatment is the same.
Contrary to the saying that it can’t hurt to test, too much testing does hurt. Atul Gawande wrote in his seminal article, “Overkill”, that “Millions of people are receiving drugs that aren’t helping them, operations that aren’t going to make them better, and scans and tests that do nothing beneficial for them, and often cause harm.”
One example of this is prostate cancer. According to Malcolm Thaler, a provider at One Medical, “screening for prostate cancer with PSA testing leads to many false positives (or true positives that never become clinically significant) with all the risks of unnecessary testing and treatment, along with all the accompanying stress and anxiety.”
As precision health industry matures, we must take care that new products and innovations don’t give people worse health at higher costs. The takeaway here is that in order to make precision health useful, data need to be tied to clinical outcomes. Oncology, the study of cancer, does a great job of this. Because this field has rich datasets matched with clear outcomes (treatment x helped the patient live y days longer), the data are actionable and useful.
Lastly, genomics is just one part of the kit. Treatment, lifestyle, and social data are great sources to mine from and if they could automatically be brought into the health record system, it would be a huge win.
Data form the backbone of precision health and there are challenges on this front.
- Data accessibility: Health organizations usually don’t like to share data. Difficulties around privacy and ownership are still being worked through.
- Data interoperability: There are few standards and rules for how datasets talk with each other. Even figuring out which patient is which is fraught with difficulty.
- Data cleanliness: Dirty data needs to be reformatted and cleaned before it’s useful for analysis. Dealing with inaccurate and conflicting information is a huge headache.
- Data structure: Most data is unstructured (free text), not structured (organized). This makes it hard to arrive at meaningful conclusions.
Let’s dive into the last point, data structure, because data need to be structured before precision medicine can do its magic.
Currently, solutions roughly fall in 3 buckets:
- Force providers to type data in a structured way.
- Use manual labor to translate unstructured data into structured data.
- Use machine learning to automatically dump data into the right buckets.
The first solution creates clinical burden for a role that is already stressful and busy. The second solution is not scalable across a large population (although done with a lot of success in oncology). The last solution is still in progress. Say an algorithm can predict with 90% confidence that the doctor wrote the patient has diabetes. It’s not OK to be wrong 10% of the time.
While there are no easy solutions, there are many organizations that are tackling this. Companies like Oscar Health, Apple, and One Medical are pushing the healthcare industry towards more interoperable, accessible, and actionable data.
Lastly, what about the humans in all this? Let’s say an algorithm could predict with 100% accuracy if someone has lung cancer. Do we still need human health providers?
At One Medical, we know that diagnosing conditions is just the first step. Synthesizing that information and convincing people to follow through on treatments requires warm, caring people. In a national Harris Poll survey, 80% of respondents wanted to have a single trusted person or resource to help with all of their healthcare needs.
We need precision health to enable more humanness in healthcare, not less. The biggest opportunities in precision health are not in mining down to the last DNA sequence and replacing providers, but cross-pollinating data and eliminating cognitive and administrative burden for our healthcare workers.
“Look to bring people along with technology” — Vineeta Agarwala, Director of Project Management at Flatiron Health
The magic world of algorithms and hyper-personalized care won’t happen tomorrow. How we can prepare for this future and thoughtfully design a human healthcare experience around it?
One Medical is hiring talented people who want to revolutionize healthcare. Visit onemedical.com/jobs. Thanks to my co-workers Tamer, Kimber, Kyle, David, Wes, and Patrick for all the thoughtful discussions and support on this article!