2019 Resolution: To Go Beyond AI Hype in Healthcare

Alt text

I’ve recently come across a few articles describing how healthcare industry is moving beyond Big Data and AI hype and is starting to look at machine learning as a means to achieve a goal rather than a goal in and of itself. In 2015-2016 it was sufficient to come up with a novel machine learning algorithm in order to spark interest of company’s management, customers and investors. In 2017-2018, many of these algorithms were put to work and - as it often happens with novel technology – did not prove to be viable. The reasons range from the lack of underlying infrastructure to support industrialized machine learning to a high failure rate of these models. This was a hard pill to swallow because by that time companies have already invested heavily in new technology, hired teams of expensive experts and - in many cases – sold customers on a promise of breakthrough innovation powered by actionable AI. In 2018, after a series of painful setbacks many application initiatives were sunset as business leaders became more conscious about spending their “innovation” budgets.

2019 has started with a question that is very often overlooked at the time of hype: what is the business problem we are trying to solve and what outcome do we want to achieve by solving it? In agile development framework, this question is often referred to as a “user story” that has to have a clear role (who is asking), an ask (a problem statement) and a value (desired outcome). For example, as an online shopper, I want my packages to be delivered to my apartment’s front door, so that I don’t have to go through every apartment in my complex looking for them (Nota Bene: my recent painful experience). This sounds like a very reasonable problem statement but in the last 3-4 years questions around AI were not about a customer or a business, they were about the technology. If I carry on with my delivery example, a “hype” problem statement of online shopping would be something like this:

“We have just created a new delivery mechanism that can deliver anything to your doorstep. Make your first purchase of groceries and we will deliver them to you for a flat $50 fee in less than 15 business days”.

20 years after the birth of online shopping, this sounds ridiculous, although, I’m not sure that this was far from truth when companies just started shipping online orders. There are many issues with this “hype” problem statement above and the most glaring issue is that it is not about a customer. First, if a family spends $200 on groceries per week paying 25% of that to have it delivered is way more expensive than to drive to a store yourself. Second, by the time this order arrives the family would probably drive to a store anyway because no one wants to wait 3 weeks to eat. Plus, most of the fresh produce and dairy would go bad. Not to mention that at the time when consumers have just started to shop online, food was not something they would trust a delivery guy to pick out for them.

But does this mean online shopping is not useful? We all know that it is, so much so that Amazon’s market capitalization is now higher than Argentina’s GDP. But Amazon didn’t start with acquiring high-end grocery stores and shipping food. It started by shipping books – a product that is often too heavy to carry, is usually cheap enough to cover its cost if an order gets lost, is hard to damage, does not go bad and is something people typically can wait a few weeks for. Sure, shipping books doesn’t sound as exciting as shipping groceries, clothes or makeup but this was the only product back in late 90s that could be reliably shipped to a customer and make her want to order again.

With AI in healthcare, we are now likely where Amazon was with online shopping in early 2000s. We have built a great new technology but it is too unreliable, too expensive or too inconvenient to use in the majority of use cases. But this doesn’t mean that the technology cannot or should not be implemented in healthcare. It just means that we have not articulated a specific problem we should solve for our customers with our current AI capability and haven’t decided what realistically we should do with our solution.

Sandcastle Example

Let’s take an example of a machine learning algorithm that can predict a medical event in patient’s health. Depending on the richness of the data, scientists can come up with a model predicting an inpatient treatment regimen that leads to the shortest time until hospital discharge while somehow guaranteeing low risk of hospital re-admission. This sounds like a golden capability – not only does it help patients get well sooner, it reduces overall financial burden on healthcare to keep patients in the hospitals and increases the amount of “patient throughput”. On the surface, it looks like all healthcare players could benefit from this technology. Hospitals can admit more patients without increasing physical capacity and capitalize on resources typically spent on identifying the reason for admission and prescribing an appropriate treatment regimen. Payers will spend less reimbursing hospitals for beds and will have peace of mind knowing that hospitals are not prescribing unnecessary procedures just to increase a bill. Most importantly, patients will spend less time at the hospital and skip initial waiting period when a team of physicians is working to determine the right set of procedures.

This sounds like a great solution with real business and health benefits, but just like with online grocery delivery in early 2000s it comes with issues few customers would tolerate. Firstly, in order to build a model that can predict the right set of procedures it has to be fed with lots of patient data points. Most hospitals are currently not equipped adequately to collect such information on patients. This is dangerous because in theory a machine learning algorithm could be trained on data that has a high degree of granularity and detail, but once it is deployed in “real world” its prediction will start failing due to imperfection of real world input data.

Secondly, because most hospitals do not have sufficient data to achieve reliable and repeatable prediction they would want to have real people double check what a machine is recommending (to avoid liability) and then administer these procedures. This will most likely increase spend on headcount instead of decreasing it and will almost certainly add to patient’s waiting time.

Thirdly, just like with online grocery, there is a large trust barrier people have towards machine prediction and trusting machines with your own lives is not something most people feel comfortable with. At least not yet.

Finally, it is important to understand and follow each player’s incentives in this scenario. For machine learning to be accurate it has to be fed, as we saw earlier, with a lot of granular patient data. This means that hospitals have to be equipped with costly technology that enables complex data collection. The good news is that some hospitals like NewYork-Presbyterian David H. Koch Center are already starting to do so by digitizing inpatient experience and collecting data on patients in real time. But inpatient health digitization is still considered cutting edge for most hospitals and they may not have access to large investments for such initiatives.

Specific problem + realistic solution

We can see that even though it is possible to build a sophisticated machine learning algorithm predicting optimal treatment regimen and it makes sense from the financial and business perspectives, it may still not prove to be viable, just like it was not viable to deliver groceries to your doorstep in early 2000s. So, what would be an analogy of Amazon book delivery for today’s digital health?

Following customer-centric principals and using an inpatient example, a smaller but more realistic use case could be to make physician’s time-consuming job more efficient. For example, according to one of the Quora answers given by a radiologist, an inpatient CT scan interpretation takes between 2 and 4 hours if analyzed by a specialist. If it was reduced to, say, 30 minutes, this could bring efficiency to the hospital’s operation and result in patients discharging sooner. Current machine vision capabilities and available training data are not sophisticated enough to read CT scans and interpret them automatically. However, there are machine learning algorithms that can highlight “abnormal” areas, so that a specialist can focus her attention on just one subset of a scan. This may not result in 30-minute interpretation time but it could potentially reduce it to 1-2 hours. Compared to a 2-4-hour timeframe this is still a significant reduction that generates efficiency. It is also more realistic from the operationalization perspective and the trust element is being dealt with because a radiologist still inspects the scan.

Machine-aided interpretation of inpatient CT scans does not sound as grandiose as automating treatment pathways but it is a pragmatic application of current AI capabilities that provides real value to the user. In time, as input data becomes richer (thanks to ongoing health digitization), AI capabilities mature and customers trust machine automation technology more, we will observe automated diagnosis and treatment predictions. But until then, healthcare industry needs to learn how to earn customer’s trust by “delivering books” – reliably, inexpensively and quickly.