Facing the Fax: The New Era of Machine Learning in Healthcare
When you ask Cohere Health leadership what their vision is for the future of healthtech, they are quick to quip.
“Hopefully there won’t be any fax machines,” said Chief Technology Officer Niall O’Connor.
“And no more handwriting,” added Director of Machine Learning Anne Nies.
But don’t take their banter for a lack of seriousness. In fact, O’Connor identified the fax machine as an enemy of modern medicine. A massive amount of healthcare communications are still conducted via fax, and the clunky 1964 invention is perpetuating the generation of mountains of unstructured data.
This outdated technology contributed to prior authorization being named by the Medical Group Management Association as one of the most burdensome and slowest processes in healthcare last year. Patients must wait for an administrative check with insurance, often via fax, to take the next step in mitigating what might be an urgent problem. According to a 2020 American Medical Association report, 90 percent of physicians state that prior authorization has had a negative impact on patients, and 30 percent report that the practice has “led to a serious adverse event for a patient in their care.”
“Imagine a chronic pain patient going to a doctor and being sent home for seven to 14 days to wait for an administrative check with their insurance. That is cruel and unusual,” said O’Connor.
What is Prior Authorization?
Cohere Health is tackling the problem of prior authorization with machine learning and artificial intelligence. By skipping the fax machine, digitizing information and offering providers real-time feedback on prior authorization requests, patients receive access to appropriate care four days sooner on average. Ninety percent of prior authorization requests through Cohere are getting instant approval. Through the platform, prior authorization becomes an invisible process that reduces deviations from evidence-based care, rather than a barrier to access.
Perhaps unsurprisingly, this is not the only challenge in the medical information management industry. Much of healthcare is governed by medical policy and clinical guidelines, which is why Cohere Health uses a combination of traditional business rules and machine learning models to automate and scale decision-making processes. The company’s engineers often find they’re able to make a significant impact on this work quickly, and new team members are often pushing code to production in their first week.
The company is also using digitization to tap into crowd intelligence and create a network of information to support providers. “Individually, doctors might be wrong sometimes, but we can use data to help them draw on the knowledge of the broader population, and there is opportunity for much better health outcomes,” said Nies.
Want to be part of the team pushing healthcare forward? Built In asked these two leaders to describe their work and how it fits in with their vision for the future.
How does your work solve the problem of prior authorization and what impacts have you personally seen?
O’Connor: We looked at a maddening set of clinical guidelines that determine medical necessity and clinical appropriateness. There are guidelines set by the government and the Centers for Medicare & Medicaid Services, local rules, and different commercial payers. They are all similar but with slight deviations. It’s impossible for any clinician to know them all.
We’ve codified all of it. It reads like Boolean logic, so we turned it into Boolean logic. We simulate the guidelines as they are written by all these different bodies, and we can give providers real-time feedback about whether their plan is clinically appropriate or fits the definition of medically necessary. Sometimes it means that they have to change the dosage or order an X-ray before an MRI. They need to try a conservative therapy before surgery. We can give real-time feedback, which means that we focus on the approval for them.
The idea of having a unique care path for every single patient is not beyond reach, because personalized treatment is the outcome of digital data in patient care.”
We’re trying to create the most clinically appropriate and medically necessary requests from the get-go. This is all done digitally, not via fax machine, because we can’t nudge a fax machine. This allows clinicians to schedule the patient for a treatment that same day.
We have a consumer focus on this and it shows in the feedback we get. We have a provider net promoter score that is in the range of consumer applications. We’ve taken something that people have to use for their job and made them actually satisfied.
Tell us what you think healthcare could look like in a decade. What are you working on right now that would make this possible?
O’Connor: There’s so much administrative work; the AMA has found that physicians and their staff spend an average of 13 hours each week completing prior authorizations alone. Even when you are at your appointment, it feels like they are checking you in for a flight with 1,000 keystrokes at the computer to fill out a form. We’re using a human instrument, the physician, to do non-clinical, non-patient-facing work.
In my vision of the future, I imagine having these patient-physician conversations recorded by a machine with a high degree of accuracy and privacy protection. The physician can reference your chart on a device where they can see structured information from prior encounters. It’s all machine readable, and the insurance company can turn around instant decisions so the patient can get treated immediately. Then the patient goes home and has already been mailed their treatment or medication, so they don’t have to stop at the pharmacy separately. We have devices that help with patient adherence, or monitor health and record information that is important to discuss at your next appointment.
Right now, so many companies want to make a healthcare innovation then run up against fax machines and give up. But if we can solve that data capture problem upfront, it unburdens our clinicians and enables the kind of healthcare future promised in Star Trek to become a reality. It’s attainable because it’s been done in other areas.
The idea of having a unique care path for every single patient is not beyond reach, because personalized treatment is the outcome of digital data in patient care. Machine learning can get us greater levels of specificity, which allow us to create individualized plans.
So many companies want to make a healthcare innovation then run up against fax machines and give up.”
Nies: One of the best healthcare experiences I’ve had was seeing a doctor in the same building as the lab, phlebotomist and pharmacy. My doctor could put in an order and I could walk to get tests and then my doctor could send a prescription for me to pick up at the pharmacy before leaving. Why can’t we have that kind of experience across the board? How do we get the patient to the next step without having to wait for weeks to get what they need? Imagine how radically different the outcomes would be. Imagine how much easier it would be to adhere to medical plans. If you have to go out of your way, you might not do it. Getting better at the upfront data capture and putting machine learning in the process makes it a more guided experience.
One piece that is exciting to me is getting the right care to the right person. I think we’ve all had the experience where you feel like you are just a checkbox on a list and you aren’t actually a human being. How can we use machine learning to do a better job of understanding opportunities to give people personalized care?
Prejudice in Healthtech
What is most exciting about the work you are doing?
Nies: A lot of the problems we are working on haven’t been solved. In other fields there are really robust machine learning models, and you’re just improving them. In healthcare a lot of those models fall apart. There is a huge opportunity to make an impact and try new things out.
In other fields there are really robust machine learning models, and you’re just improving them. In healthcare a lot of those models fall apart.”
O’Connor: Sometimes there isn’t a direct tie to the metrics driving success. The more we can get discrete and specific information out of unstructured sources, the more we can understand variations and outcomes. It’s rare that you are so coupled to the impact.
What should people who are interested in joining this team know?
Nies: We move fast. That’s something special in the machine learning space, especially in large corporations: Usually you’re either improving a model that already exists or you’re waiting a long time between when you built it and when it gets used. Here, you get to see your work not just existing, but being used to make people’s lives better.
O’Connor: Anyone who joins the company will be pushing code to production right away. That’s unusual in a lot of companies. Regardless of your years of experience, we expect you to grow and develop very quickly. Nothing gives me greater joy than to watch an individual growing in their career.