Where Are We: AI for Healthcare-- Part II
Part II: How Do We Evaluate AI Innovators? 

Artificial intelligence (AI) can be applied to numerous healthcare fields benefiting a variety of healthcare stakeholders including physicians, hospitals, health systems, payers, pharmaceutical companies, and medical/wellness device makers and application developers. The AI technology itself has myriad of flavors and boasts different degrees of sophistication. These attributes make this early stage innovation ripe for being hyped and more difficult for healthcare customers and investors to discern true innovators from pretenders. 

A good assessment starts from more accurately categorizing AI-for-Healthcare solution providers based on key attributes and targeted business use cases, though there are multiple means to slice and dice this market. One that I use focuses on whether an AI-solution provider sells its technology as a tool/platform, or it merely sells an application or service with AI-embedded/enhanced functions and features. The tool/platform providers usually have a broad interest in and collection of data sources; their algorithms can be broadly applied to business cases; and they sell their solutions to enterprise customers so that the latter can use them to accomplish tasks/goals. AI-enhanced application/service providers, on the other hand, normally specialize in analyzing one particular data type; as a result, their algorithms usually target a specific business use case; and they can offer solutions as a tool to enterprise customers as well as sell it to end users as a device, a test, or a service. 

Let's illustrate this approach by using two examples. Health Catalyst, for instance, is an AI tool/platform provider.  Health Catalyst not only helps its clients build a health data warehouse, but also teaches clients to use its analytics to understand patient risks, identify clinical care gaps, improve patient engagement efficiency, collect payment faster, and so on. It does so as a technology partner to enterprise customers.  

By contrast, BioBeats is an AI-enhanced wellness application/service provider. It uses AI to enhance the effectiveness of its wellness application and service based on biometrics collected by smartphone sensors and consumer self-reported data. As a corporate wellness solution, it might have access to client-provided employee data. Nonetheless, its AI algorithms work on a narrower set of data sources, analyzes patterns and generate insights only for wellness improvement's sake. Lastly, although its clients are enterprise customers, these clients are also channels through which BioBeats' app is eventually used by consumers. 
In summary, Health Catalyst is a "horizontal play" in the AI-for-Healthcare market, whereas BioBeats is a "vertical play." 

I use this evaluation matrix to compare three horizontal AI-for-healthcare players-- Flatiron, JVion, and Cloudmedx, and three vertical players—Ensodata.io, Buoy Health, and Zebra Medical Vision. The comparison is summarized in two tables below.​
Evaluation Example
A Landscape View of AI-for-Healthcare Solutions
​​Certainly not all AI players can be easily classified in this way. Many horizontal players have the capability to specialize in certain verticals or business use cases, and vertical players may also adapt their algorithms to other data sources and expand their business to become a platform in the future. The sketch on the right is a landscape view of this AI-for-healthcare market. ​​

​Scroll down to read Part I​​
Where Are We: AI for Healthcare-Part I
Data Digest
Part I: How Do We Define AI? 

AI ≠ Virtual Assistant ≠ Machine-Learning ≠ NLP

Is Artificial Intelligence (AI) a game changer for healthcare? No doubt. Are we at the top of AI-for-healthcare hype? Absolutely.
Over the last 18 months, more than 100 startups and incumbent analytics companies have showed up, claiming to possess AI capabilities that can benefit healthcare enterprises. Their websites and PRs are filled with buzzwords such as deep learning, natural language processing, virtual assistant, etc. All these things are NOT equal and they cannot be used in an interchangeable fashion as is done today. 

Let's try to define artificial Intelligence in clearer terms and within the healthcare context, shall we? 

An artificial Intelligence solution must have three core elements: Sources of data, analytics engine, and Application. Comparing and contrasting different AI solutions must examine all three aspects. 

Data sources are the foundation of AI. In theory, the more diverse  the data sources, the more intelligent an AI solution would be.  AI can work on one single data set—for instance, scanning patient claim data to map a patient population's risk profiles. It can be done through machine learning—very elementary one though.  A better AI solution can analyze multiple data sets simultaneously and add much more nuances--and richness-- to a risk profiling model to make it more accurate, reliable, and practical.  

Analytics engine is the secret sauce of any AI solution. No matter it is called machine learning, deep learning, or artificial neural network model, it has to perform some types of analytics. We may never know exactly how robust an AI technology is technically (unless there are peer-reviewed publications, which is rare today), but we can see whether such technology produces something new, meaningful, and actionable.  

If format of analytics output is new but insights are simply old information organized and presented in a more meaningful way, then such AI is simply a descriptive or summary analytics tool. One can argue that data dashboard features or risk scores are machine-learning enabled, thus an AI solution. However, merely describing data in a more intelligent way is a low-level AI.  

If AI helps produce new insights that is previously hidden or unknown, such discovery can potentially lead to new means of patient care. Discovery analytics has been the growth engine of the pharmaceutical industry for the past half century and it is now moving to consumer health applications. For instance, an AI solution may identify causes of insomnia by analyzing a person's sleep data, his sleep environmental parameters, and personal behavior attributes. This is a huge step forward compared with just a sleep quality score. 
An even bigger leap is the ability to predict an outcome and advise on the best actions to take advantage of it (if positive), or mitigate/prevent it (if negative). AI Solutions today claim that they can predict a patient's probability of contracting chronic conditions, re-admitting to ER, or skipping preventive care episodes. However, such predictive capability is rudimentary and unreliable, in part because their analytics engine lacks  richer, more diverse datasets to work with.  

Lastly, applications, not technology or data sources, are what drive AI's market success in healthcare. If a technology breakthrough does not support applications that address a customer's immediate and most significant business pain points, healthcare customers may brush aside such innovations as premature, and innovators are stuck with endless pilots without seeing a signed contract. 

In summary, AI is the product of data, analytics, and applications. An AI solution's strength reflects the merits of its data's diversity, analytics engine's superiority, and applications' commercial viability.  

Machine learning and natural language processing are technologies integral to AI's analytics engine. Virtual assistant is a generic AI application that may or may not possess superior analytics capability. Its success is dependent on whether healthcare-pecific virtual assistant applications can address customer's immediate and significant pain points. 

Medical practices offer telehealth services
21% plan to offer in 2017 while 36% will not
Source: MGMA Stat Poll in Jan, 2017​​
Medical practices use patient provided data
Source: MGMA Stat Poll in Jan, 2017​​
U.S. seniors who don't have access to telehealth or don't know if they have it at all.
Source: HealthMine Survey in Sep, 2017​​
Have Feedback? 
Upcoming Posts
Part III: White Space Opportunities in AI for Healthcare
Part III will examine the white space in AI applications and pinpoint market opportunities based on customer demand and technology feasibility.