Taking AI and Machine Learning to the Next Level in Healthcare

By Tony Ambrozie, SVP and Chief Digital & Information Officer, Baptist Health South Florida

There have been spectacular advancements in data insights and machine learning capabilities in the last ten years—all driven by cloud economics, availability of large quantities of data, and massive investments. While we’ve seen certain industries taking advantage of these advancements, others, including healthcare, have far more opportunities than they do successes in the machine learning space.

So, where are we today in healthcare?

We are exponentially increasing the amount of data being generated:

1. Firstly, we are putting sensors into and onto seemingly everything: houses, buildings, streets, cars, planes. Not to mention people.
   
   
2. Secondly, sensors are getting increasingly sophisticated and intelligent, which means they can tell us a lot more than they used to.

While some data is just noise, a lot of it is very valuable on its own and even more so in aggregation to other data from different sources. 

Not only do we have the data we need, we now have significant needs to utilize machine learning to drive enormous value in the healthcare space: 

1. Healthcare operations must become significantly more efficient, reducing costs without reducing the quality of care. Efficient staff and physical resource scheduling immediately comes to mind in the clinical space (with the added bonus of positive outcomes to patients), while efficient supply chain management rises to the top in the administrative operations space. This is something other industries, such as financial services, hospitality, travel, and entertainment parks, have done successfully for some years.
   
   
2. The need for diagnostics support using narrow AI has been increasing exponentially in the last 2-3 decades. Expansion in sophistication and prevalence of tests, deeper understanding of complex medical conditions as well as compounding interactions of many factors, including medications, means there is more and faster way to analyze now than ever before. For narrow applications, AI has proven it can help. While notable commercial applications providing relatively high accuracy exist and new ones appear constantly, the road to an AI-rich environment in this space is still long. Commercial vendors and startups working with physicians and the FDA and supported by vast amounts of VC capital will ultimately fulfill the need.
   
   
3. Finally, systems assisting physicians, including by minimizing the administrative EHR burden, would be very valuable. Moreover, systems actively assisting during physician activities (consult or even surgery) and either providing recommendations, summaries, updates, notifications, or alerts in real-time based on the situational context would be priceless.  And, increasingly, this will mean the use of what we would call real-time intelligent assistants.

It is the intelligent assistants where I think we will see a tremendous amount of value. This is where humans and computer “smart” systems would actively and dynamically be collaborating in real-time on relatively sophisticated tasks.

It is the intelligent assistants where I think we will see a tremendous amount of value. This is where humans and computer “smart” systems would actively and dynamically be collaborating in real-time on relatively sophisticated tasks. The systems would not merely respond to requests or input from the human but proactively, independently, and dynamically process the actual task context, actions, and current flow state, efficiently collaborating with humans.

This capability will be powered mainly by comprehensively injecting real-time Machine Learning insights in all existing applications, and creating specialized capabilities that don’t exist today.

Active human-system collaboration is already happening in some basic, generic systems. We are seeing the start of it in more complicated systems and we will progressively see it built into other commercial systems of increased sophistication. As developers and vendors of digital capabilities, we must actively pursue the injection of ML insights into all those cases where it makes sense to enable the system to “collaborate” with humans.

At the end of the day, through, for such intelligent assistants to see adoption, the systems would have to prove effectiveness and accuracy. If, for example, a surgery assistant gets in the way during an operation instead of helping, you wouldn’t use that person next time. Similarly, a system’s incorrect activity could lead to very adverse outcomes, so it would not be used. Therefore, the feature must work very well and be adequate to the time/place/action before making it comprehensive. Less but very good is better than more but buggy.

Despite all these exciting opportunities, so far, a significant size of advancements in applied ML have been happening mainly in distinct but limited areas, such as marketing personalization, and mostly by technology companies, startups, financial and streaming giants. Unfortunately, everybody else, including in healthcare, has done a lot less than they could and should. Too much talk and too little walk in too many quarters, as they say.

How do we change that?

First of all, from a technical perspective, for AI/ML to be impactful, the predictions must get injected, preferably in real-time, into the operational systems. There are two parts to getting this benefit:

1. Exposing the recommendations coming out of ML so they can be easily consumed, preferably through APIs.
   
   
2. Changing the operational systems to consume this data. The insights must be consumed correctly and effectively at the point they add the value – either in transactional systems (which therefore must be enabled to consume the insights) or sometimes directly by humans. That part is more expensive, especially if those systems are not modern enough.

Secondly, we need to move our cultures and decision-making processes to be more data-driven versus opinion-driven. There are culture and change management aspects to be managed, as well as trust on insights (models explainability can go a long distance towards establishing that trust).  Sure, data can be interpreted differently or misinterpreted, but that does not mean we should manage out of opinions and gut.

Thirdly and finally, we must be deliberate and strategic in where we focus our data efforts:  

• Aligned with whatever strategic direction and goals we have (those are still very much important in an agile world)
  
  
• Where is the biggest challenge?
  
  
• Where is the biggest opportunity for positive impact?
  
  
• To support regulatory or compliance requirements 

Rigorous prioritization is critical, as trying to do everything at once just because everybody wants the outcomes of ML immediately only leads to diffused efforts that lead nowhere.

In conclusion, renewed and robust efforts to introduce AI and ML are something we must all do. And since we know how to do it by now, it is now a matter of will and execution.