Machine learning, or deep learning, are common buzzwords today. Nym bucked the common fashion trend and built a medical text analysis product based on a big data system that has no machine learning at all. Nym’s system automates a major component of a hospital’s bill collection interface with/from insurance companies. This young company is already a significant factor in the market.
Amihai Neiderman, founder and CEO of the company (31) earned two college degrees, one in mathematics and the other in computer science while still in high school. He did his 10-year army service in the cyber unit of 8200. “It was a magical time for me. Both satisfying and challenging. I was making an impact. And forming great friendships.”
When Amihai left “8200”, he was an entrepreneur looking for a killer idea. "At that time, my wife Narin, who is a doctor, was engaged in medical research. She would disappear for hours to Ichilov Hospital to search the files for the patients who were suitable for her research. The files were already computerized, but the search algorithm was ineffective. She had to use all sorts of tricks, such as searching for the name of a doctor she knows to be performing these types of surgeries.”
"It bothered me because in my army unit, everything that could be automated, was automated, and everything that could be streamlined, was streamlined. So, I wrote them a 'poor' language analysis algorithm.
“I wrote it over one weekend and holiday, and it worked. My wife was the youngest intern in the department, and this helped her produce quality research and progress rapidly. I enjoyed the fact that I could do something which provided a quick and significant improvement…and of course enjoyed the added benefit that my wife would be able to spend more hours at home.”
Her studies were so successful that Nerin began traveling to conferences and lecturing. "And as I travelled with her, I began to understand the world of healthcare a little bit. In particular I understood why it was so difficult to bring technology to bear on new products. In the military, it’s all about iteration. Fast prototyping, then adjustments to match the reality on the ground. In medical care, that can’t be done. People can die from careless mistakes. As a necessity, new products are developed extremely carefully, and sales and implementation cycles are incredibly long. It occurred to me though, that doesn’t have to be the case on the administrative side of the healthcare world."
Encouraged by the initial success, Neiderman decided to set up Nym, together with his partner Adam Rimon. "He's a friend of mine from the Unit, and after his release from the army he studied computational linguistics at Tel Aviv University. This is a field that significantly preceded the natural language algorithms of commercial companies but remained within the halls of academia. A part of this research focuses on basic linguistic questions, such as how to understand subtext and sarcasm, or hypothetical questions which don’t necessarily end in a question mark.
"Happily, the clinical language doctors use does not include jokes or complex cultural references - and therefore, in this sense, it is easier to understand. Medical language must be precise and understood by any doctor who reads the case."
Nym's developers are mainly in Israel. "Many of them served in 8200," says Neiderman. "We have several developers who have advanced degrees in computational linguists. They have developed and are enhancing our core language comprehension engine. We also have medical doctors – all with computer science backgrounds or who are learning how to code, who take part in development. In addition, our sales, product and customer success teams are based mainly on the U.S. East Coast.”
The Nym team worked for two years improving their core technology before venturing out to the market. "I met at a conference with a huge company that provides shorthand services to doctors, through a transcriber sitting next to them while they dictate their charts. I was sitting next to the CEO who when I mentioned my military service immediately became interested. 'What gun did you have?' He asked, and I told him I had a really good computer. He insisted on inviting me to his house, shooting beer cans in his backyard. Quite an introduction to Georgia. My shooting was better than I thought it would be. I still have the can.”
Our first deal came from this entertainment when we signed with ScribeAmerica. "They offered to pair their service with ours in order to look more innovative. And they paid for the privilege.” The following year many additional deals were signed including Geisinger, a network of 14 hospitals and a medical school, reputationally a leader in Healthcare IT. Nym's investors include Google Ventures, Bessemer, Lightspeed and Dynamic Loop Capital. A total of $23 million was raised.
"We have a nucleus of customers who are very satisfied with the service. There are literally dozens waiting to work with us. We are building our implementation team rapidly which will allow us to add new customers with more speed and efficiency. Each of these deals is potentially several million dollars. Alongside our partnership with ScribeAmerica, we are approaching customers directly. Our technology currently covers emergency rooms, cardiology and radiology departments best."
To understand the market in which Neiderman operates, one needs to understand the healthcare system in the United States. “Hospitals are private, for-profit organizations, or at least are managed as such. Insurance companies are also private corporations, including Medicare (Federal insurance program). Each side wants to make money, and there is distrust between the two.
“Insurance companies want to know exactly what treatment was administered, why it was administered, and what parts of it the insurance company paid for. But they don’t have enough staff to read each individual discharge letter, while at the same time, doctors have no interest in sending over the discharge letter which outlines their entire thought process. This makes them vulnerable to appeals, and even legal liability.”
The interface agreed upon between both sides is based on what is known as medical codes. Each treatment, diagnosis and use of medical equipment by the hospital receives its own unique code. The list of codes is sent to the insurance company, and each code is individually paid. Coding a medical record takes a few minutes, so coding medical records can take up an hour or more of a doctor’s time. Accounting for the cost of an hour of each doctor’s time, this can add up to a massive loss for the hospital or for the doctors themselves in a private clinic.
The doctor is not required to do the coding by himself, however, and an entire medical coding industry has grown. “There are 250,000 medical coders for the US healthcare system. Many of these coders are based in the Philippines, India, and there is even a group of ultra-Orthodox medical coders in Israel”. The medical coder reads the discharge letter and translates it into codes, after having being certified in a medical coding course which takes several weeks.
“As the medical coding system grows evermore complex, it becomes more unreasonable for a person to do the medical coding,” says Neiderman. “Today the US healthcare system has approximately 200,000 codes, a number which has grown significantly over the past 3-4 years. As the codes themselves grow more complex, applying them requires more and more medical knowledge, which medical coders might not possess.”
If the coder makes an error, the claim is denied, and it doesn’t always make sense for the hospital to appeal to reopen the case. “Hospitals end up losing billions of dollars a year to claim denial as a result of coding errors. If there are too many appeals, the insurance company may put a hold on payments for several months, until the hospital ensures its coding system is operating properly. The hospital may even wind up with low cashflow. If a coding error is sent to a government entity, the insurance company may determine this is fraud.”
Did no one try telling doctors in advance to just write in code?
“Doctors are under tremendous amounts of stress, and their days are extremely chaotic. There are different medical dialects, different formats. Young doctors will not have time to learn how to write in code while also working, which is demanding from their very first day on the job. When attempts were made to ask doctors to write in a more formulated manner, the doctors resisted, or ended up spending too much time on the task.”
Nym is not the first company trying to create an algorithm to decipher text written by doctors and automatically convert it into prepared database. Companies such as IBM and Microsoft, alongside numerous other smaller companies, are all working the problem. “It’s a challenge defying solution for 20 years already,” said Neiderman. “But they all have one thing in common: They developed models based on Deep Learning and Machine Learning.
“There are two problems here. The first is that it’s very difficult to reach the level of accuracy required for medical texts, particularly because each sentence is rich with information. When the doctor mentions a drug, is it a drug the patient takes at home or was it administered during the patient’s visit? You also need to determine the level of confidence in which things are said. If a diagnosis is mentioned, is that the doctor’s diagnosis, or something they wish to count out? Perhaps it’s something the patient mentioned he believes he has, and the doctor noted it even though they believe it to be something else?
“The second hurdle is that Deep Learning algorithms can easily turn into black boxes. When you submit a claim to the insurance company, and your only explanation is ‘Because that’s what the computer said’ – the insurance will deny the claim.”
Can’t someone build Machine Learning algorithms which can explain their rationale?
“When dealing with these masses of data and with Deep Machine Learning, probably not.”
Which means that the world will need to take a step back from the idea of Machine Learning in many fields.
“That’s definitely a possibility. For instance, if we ever want to implement our algorithm to support medical decision making, I don’t believe that there is a place for statistical models. You can’t tell a patient ‘We’re going to amputate your leg because that’s what we did for ten people with similar cases to yours.’
“When Amazon created a Deep Learning product for medical texts, they saw that the results changed based on the patient’s name. That shouldn’t happen. Other systems which practiced on historical medical charts mistakenly learned that ‘She’ is always the nurse, never the doctor.”
Rimon and Neiderman implemented classic computational linguistic thinking into the coding problem. “We realized that we would not be able to understand and code medical texts without first understanding the medical narrative – What brought the patient to the hospital? What symptoms did he tell the doctor he had? What did the doctor think of that? Who said each sentence, was it from the doctor to the patient, the patient to the doctor, the patient’s mother to the doctor?
“Some of the companies which tried to crack this nut developed tools “For coders, by coders”. At Nym, the people who write the rules are doctors. They don’t read a medical chart the same way a medical coder does, they read it as a doctor would, after having passed a medical coding course. External audits show Nym’s engine consistently reaching 97-98% accuracy. Thanks to that, alongside our ability to explain just how our algorithm works – rule-based, and not Machine Learning-based – we made quite a bit of noise in the industry.
“The downside is that we don’t know how to do this for all charts, we currently only have 60% coverage. It was hard to be in the position where we had to tell customers ‘No’, but we wanted to stay focused – and it paid off.
“Meanwhile, some of the charts we return, saying that we were unable to code them using the Nym engine, which makes the medical coders happy. They feel that they are receiving the more complex cases, which only a human can code. For what we code, however, customers can rest assured. And if an insurance company will appeal a claim or run an audit, they have full automated documentation and auditing of each and every decision.”
How did COVID affect your activity?
“At first there was some uncertainty, when we weren’t sure of how the market would react, how much hospitals would be interested in implementing new technologies during the pandemic. Then in March, India and the Philippines entered a lockdown which lasted several weeks. A large part of those medical coders we are replacing are from India and the Philippines, and as a result of the lockdown, many hospitals lost the ability to code files and receive reimbursements overnight.
“They began approaching us, asking to use our solution. We discovered that one of the reasons to use Nym is our availability – since it is not human-dependent, Nym can code 24/7 even during a global pandemic like COVID. We weren’t ready for such a large number of customers, so we currently have a waiting list made up of several hospital and physician groups, just waiting to be onboarded.”
Subscribe to our exclusive mailing list and get the freshest stories from the Lemonade team