#AI for healthcare? Great but let's get these things right
Whether you are IBM Watson or some other AI system, you seek to find answers to questions. That humans can't answer quickly or analytically.
- What's the likelihood of this autoimmune disease in that patient?
- What does the latest evidence say about this treatment protocol for that cancer?
- From which zip codes will the next generation of asthma patients come from?
When we ask such cognitive (thinking) questions, there's a big underlying assumption. That the data we feed into the AI system is reliable. That it's complete. And it's relevant. More important, labelled or catalogued correctly. So that the AI system understands what exactly that data means.
We know that Facebook AI can see photos. And Amazon AI can predict what you may want to buy next. Behind such AI applications are underlying systems supplying reliable data. Standards fuel connections between a variety of applications.
For example, if Android can't integrate with Facebook, computer vision would be pointless. You'll be still at the level of figuring out how to upload your cat video.
And that's exactly where healthcare is stuck today.
Before we journey to the future with AI, we need to pause and examine the past.
There's a reason standards exist. That reason is to make us talk.
Standards are the reason you see YouTube videos inside your Facebook app. It's how iTunes on your phone can play songs on your BlueTooth speaker. Also how your old VCR played VHS tapes.
AI for healthcare needs better foundation
Healthcare has many standards. And yet our systems (both software and hardware) hardly talk to each other.
We continue to drag our feet with really old systems in healthcare. It's like uploading home videos from VHS tapes to Facebook.
We don't seem to see beyond present-day competition to re-imagine the industry. So we hold onto data of our clients and lock them up. Creating islands of information.
Do you HL7?
Consider a dominant data standard called Health Language 7 (or HL7). Most global software and hardware systems aren't yet HL7-complaint. But should be.
[Aside: Read about FHIR that's taking HL7 standards further - What is FHIR and why should you care]
Even if systems are HL7-enabled, they aren't ready to connect with others. That's why every time a hospital goes live with an electronic health record (EHR), millions of dollars are spent in integration.
HL7 looks something like this.
If you take a moment to read through, it can tell you that it's about a patient in Birmingham, Alabama. With a certain body height and weight. The patient's on aspirin and has unspecified chest pain.
This information is labelled. It can now be understood by systems that follow HL7 standards.
And because someone else can understand our data, we begin to communicate. Sync up by exchanging information back and forth.
If we begin to communicate then we have interoperability of information.
If there's interoperability then we can do amazing things. Like watching YouTube videos on our Facebook feed. Even though they are two different companies.
If we have interoperability then we lay the foundation for good data. That makes AI efforts meaningful. To repeat, data that's reliable, accurate, complete, and relevant.
As of today, there's a joke that runs in industry circles.
Question: What's the fastest way to get a medical chart from a hospital to the one across the street?
Answer: Make it into a paper rocket.
The problem of garbage-in-garbage-out
If you are in healthcare, you would know that we produce reams of medical information. Most of it is not pertinent.
Templated data from EHRs. Unnecessary lab tests. Excessive imaging studies (CT-scans, X-rays). Many prescriptions that often have harmful interactions with each other.
Now add to this the vast unknown. Such as the microbiome that we know very little of. We've only recently discovered that we are 90% microbes and 10% human.
If you ask a seasoned clinician, she'll tell you the problem. That there's very little wisdom from the kind of data that exists today. Doctors often face a big task to wade through noise. To find what's relevant to the medical case at hand.
Now consider channeling all that data into an AI.
Facebook AI learns from billions of photos that we upload. All relevant data. But a healthcare AI would struggle with the today's data. It'll create worrisome patterns.
Fast forward to 2025.
You'll have an AI system magnifying the nonsensical data it learnt from back in 2017. Imagine yourself as a patient treated by AI. Or a doctor assisted by it. Ahem.
Do you see the challenge of relying on AI decision making based on such data?
So what do we do?
Of course, artificial intelligence will help the industry. I hear that hospitals plan to recruit for artificial intelligence positions.
But before we spend money on expensive AI scientists, we have some work to do.
0) Agree on standards and implement them. We have many working standards, particularly in the US. But its implementation is in pockets. Driving standards will pave way for cleaner data. Boring but critical.
1) Create pertinent data. Whoever we are in the industry - doctor, nurse, vendor, biotech company - we must want to create pertinent data. For example, EHRs must simplify their design. So that doctors can capture important data effortlessly.
2) Label your data. For historical data to be relevant, we must label and catalog. In a way that computer systems can understand. Consider crowdsourcing to scale this task quickly and on an ongoing basis.
3) Be ready to share. Make your systems interoperable. If you have hardware, explore ways to extract data so that it's HL7-ready. For software, create HL7 engines so that others can read your data. Open up your APIs.
Everything that's going on today seems to signal the pervasive influence artificial intelligence will have on all industries. The pace of technology innovation is changing. At a scale we haven't seen before.
However, healthcare needs to get AI right. If not for ourselves, then for the several millions of future patients.