Imagine a world WITHOUT AI in healthcare

I’ve followed the debate around AI in healthcare with some enthusiasm over the past few years, myself going from sceptic to data scientist to now AI health-tech co-founder. What I found frustrating about this area of medicine is that it’s split rather oddly into two areas of focus; either
a) what the utopia of the cutting edge of AI can do for healthcare in 5 or 10 years time, or
b) what the dystopian dangers are of AI and its clinical risks.
The third and much neglected area of this debate however is the state of the world today, a world where for the most part there isn’t any AI in healthcare. I think the true state of medicine today is neglected in public debate simply because we don’t talk about it enough.

So let’s take a deep dive into the world today, a world of healthcare WITHOUT AI.

 “It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t so. “ – Mark Twain

‘Healthcare’ as an industry is impossibly large, and we could never hope to cover the entire ‘world’ in a single post. However, the fundamental unit of any healthcare system, anywhere, is the patient-clinician consultation. Every other part of the system (referrals, tests, investigations, access, billing, insurance, etc) is a downstream derivative of that. So we will focus on the consultation as it is today.

Sir William Osler, father of modern medicine famously said;

“Listen to the patient, they are telling what is wrong.”

What he was saying, 150 years ago, was that listening to the patient’s story is the single most effective way of deriving the correct diagnosis. And even after 150 years of innovation in medicine, that is still the case. Let us breakdown the consultation into its three component parts;

  • the patient interview where the patient relays their symptoms and medical conditions which we call “taking the history”
  • the physical examination of the relevant part of the patient’s anatomy
  • the investigations (blood tests/scans).

Now ask yourself the question, “if you could perform only one of these tasks for every patient, which task would yield the correct diagnosis most often?” And by how much so relative to the others?

In todays world, most lay people go for tests, especially scans. We think of imaging like CT and MRI as a straight forward way to show “what’s really going on”, and we can cut straight to the chase. But in fact, it’s the absolute opposite.

In study after study after study after study, the correct diagnoses is most often yielded from the patient history alone ie what the patient says, 4x-20x more valuable than any other single part of the consultation.

A big reason why AI to date has failed to make much of a dent in healthcare is simply because most AI is focussed on tests – in radiology. Of the 500+ FDA approved AI algorithms, 76% are in radiology, most of the rest in imaging in cardiology. Even dramatic improvements in this area will yield very little to the world of healthcare as a whole.

So the fundamental element of success in healthcare is simply listening to the patient. Let’s look at how easy we make that right now.

1 in 2 adults globally are prescribed at least one medication every day. Only around half of those of patients actually take them and the biggest factor in non adherence is lack of communication between the clinician and the patient. Many patients don’t even understand their own “stories” today, so how can they possibly tell them?

It gets worse – half of the patient’s story is the written record. The mean patient electronic medical record is 16,800 words, about the same length as Macbeth, and around half those words are duplicated from other records.
Moreover, even if we could read a tome of Macbeth for every patient, we may not even understand the story. Non-native language speakers accessing care have worse outcomes than native language speakers. For example, your risk of obstetric trauma on admission as a non-native language speaker is twice that of a native one.

And even if we could somehow get past all these existing barriers – information overload, patient education and the language barrier, we’ve added a whole bunch of distractions on top. ~90% of medical offices now are fully digital. Studies show digitisation of medicine reduced dedicated patient time by 8%.

In the last two decades, we gave physicians computers and told them to multi task. Except humans actually can’t multi task.

For example, conversing on a mobile phone while driving profoundly impairs cognition and reaction times, as bad, or possibly worse than, being drunk. Human beings have a set limit of information we can hold in our minds at any given time, and even minor distractions (Eg talking on a phone) add up to a sufficient distraction to cause major harm.

Now think again about the physician typing and clicking endless forms on their digital record system while they try to talk to you. How can they really listen if they have 10% less time to do so and are only using half their brain to do so even then?

And even if the physician can listen to your story – it’s getting harder and harder to actually offer the right treatment.
Guidelines are hard to access, and physically impossible to follow. One study calculated the time it would take for a primary care physician to perform every guideline-directed task for an average clinic – the answer was 27 hours, a day.
It’s also impossible to keep up with new research – 1,000,000 new biomedical research papers are published every year, or 2,739 a day. Most of those won’t be relevant to every physician, but just searching through that volume of noise is near-impossible.

So, what does all of this add up to? 44 minutes.
Every 44 minutes in the UK alone the NHS is successfully sued for medical negligence. A case of serious harm or death due to a significant medical error. A literal plane crash, week on week on week. And that’s just the tip of the iceberg globally.

This is the state of healthcare today. I could give you hundreds of examples of stupid errors I’ve seen over the years professionally- any clinician can. I can give you several ones from my own family as well.

Here’s some from the press:
A 3-year old boy had vital cancer scans delayed, simply because no one sent his scan instructions to his family in Romanian.*

A 30-year old actress recently died after being misdiagnosed by a physicians associate with anxiety, when she had a massive pulmonary embolism (lung clot on the lung).**

This is the world of healthcare WITHOUT AI.

Phew.

But it doesn’t have to be this way.

In 1935 the US army was flight testing the B-17, the first prototype for the modern airliner. The first of its kind longrange airliner, it was the most complicated plane ever built at the time. The very first test flight crashed killing some of the crew. It was even said at the time these planes were simply too complex to fly.

But then the culture changed, co-pilots and checklists became the norm, and today flying on a plane is around 2,000x safer than getting in a car. Errors are now vanishingly rare in aviation.

We have a moral imperative to do the same for healthcare, with the tools we have right now. AI may be the greatest tool humankind ever invented, but it’s a tool nonetheless. How we use it, recklessly or safely, is up to us. What world we create with AI is up to us.

But the alternative, the the world of healthcare we have today, is not a world any of us should want to persist.

Dr Dom Pimenta MD

*Google Translate came out 8 years ago with Romanian

**I tested GPT4 on 20 sample questions from the multiple choice PA exam, it scored 100%, without the choices, and corrected the exam itself.