ChatGPT? What AI in Healthcare should be concentrating on

The hype surrounding ChatGPT led us to ask: What should AI in Healthcare be concentrating on?

Is AI the future of Healthcare? 

Over the last few weeks there has been lots of noise about ChatGPT and AI effectively taking over the world. I can vouch to be a paid-up fan of ChatGPT and think these tools have the potential to revolutionise all sorts of industries. However, we need a conversation about healthcare and the role of AI.

A famous Irish saying goes along the lines of the tourist visiting Ireland and asking a local for directions to Dublin. He is met by the answer “well I wouldn’t start from here”.

Healthcare is in exactly that position with regards AI.

The quite frankly scandalous investment in capital and IT infrastructure, coupled with at best shoddy implementations of EHRs across the NHS in particular, make implementing AI at scale a tall order.

But more importantly than that, we need to think harder about what other industries are doing with AI. Process automation is the top answer here. Rooting out mundane tasks and allowing machines to handle them is what most other industries are focussing on. In health, we are talking about blue sky dreams of machines diagnosing complex diseases and preventing future pandemics. That is for the birds right now.

What healthcare needs to focus on

Having spent over a decade working with over 100 clients in health, and the last five years focussing on how technology can help solve healthcare issues I can confidently say that the focus of AI initiatives in healthcare to date are missing the point.

And that point is: Data Quality.

There is little point analysing data that is incorrect.

There is little point having a machine predict something based on data that is incorrect.

In short, there is little point doing anything with incorrect data except one thing: fix it.

Why is healthcare data incorrect?

There are numerous reasons to point to but the big factors are:

  1. Data migration into EHRs: loading in historically inaccurate data with little effort made to clean it beforehand has led to an explosion of data quality errors in patient records.

  2. Implementation of EHRs: I’ve lost count of the numbers of clinicians who tell me that 20 years ago they used to come to clinic with a pen, write in the notes of each patient they saw and move on. Today, they have to turn up 15 minutes before a clinic to make sure the computer is working and has the relevant patients loaded, then they spend most of the consultation trying to find the right code on the EHR to record what is happening and the whole thing tales 25% longer. So guess what? They don’t bother. Which leads to a situation where the EHR systems get progressively worse in terms of data quality over time.

  3. Workforce: shortages of key staff, particularly administrative staff who can now be paid more working in retail than health mean that either highly expensive clinicians need to record administrative data or it doesn’t get recorded at all.

  4. Turnover: the rates of staff sickness and turnover make it almost impossible to train someone on how to minimise data errors because as soon as you’ve trained them, they’ve moved on to a better paying position elsewhere.

  5. Lack of standardisation: Put two doctors in a room and ask them to define the exact terminology for anything and I’d wager my house they can’t do it. We need to understand these people are highly intelligent and need flexibility in the way they diagnose. Restricting them to one way of ordering a specific test will either fail or the wrong test gets ordered.

  6. Validation: this is currently a hugely manual and expensive task which means that very few things can actually be checked. In a highly regulated sector like health, this often leads to governmental targets being manually checked at the expense of clinical risk, which in turn feeds the distrust of using the electronic systems amongst clinicians. And so, the die of poor data quality has been cast.

What AI should focus on

Let’s start from where we should have started rather than where we are. Healthcare technology needs to focus on high volume interactions that improve the quality of data. If the quality of the data goes up then perhaps we can talk about finding cures for cancer etc.

The highest volume interactions in healthcare are consultations -be that in Primary Care Gp visits or acute care outpatient appointments. At present these consultations are recorded in two separate places: coded data in EHRs (which are almost certainly incorrect) and clinical documentation such as outpatient letters (which are almost certainly correct).

With that in mind, the biggest challenge to achieving accurate data is making use of this unstructured clinical text that is currently not being utilised.

To put this a different way, think of the problem from a doctors perspective: how can we make it easier than when he or she could just scribble notes on paper in the past? Lets work it through:

Firstly, it’s common for a doctor to record or dictate a consultation through speech these days. These dictations use software to convert them into very accurate records of what occurred. The problem is that these dictations are extremely difficult to be translated into an EHR coded dataset, which means they largely go ignored from an IT perspective, but are the actually the thing that humans use most often.

Imagine being able to talk to a patient, have a machine record the conversation, classify the outcomes, next steps and medications and present the clinician with a bullet point summary for approval at the end of every consultation. Once approved, the machine records the data in the EHR for you and orders all necessary next steps. That’s AI and thats what we need to to focus on.

In addition to the future, we also need to think of the past. The huge amounts of quite simply rubbish data currently sitting is IT systems needs to be corrected. A simple summary of such data should be produced using AI algorithms combining data from EHR and unstructured data from clinical documents. This data should either be sent to patients directly for approval or raised with them at their next clinical visit and corrected on EHR systems.

In short AI should focus on fixing data quality first, and grow from there.

Challenges / Priorities for Implementing AI in healthcare

The first and most important challenge is acknowledging the problem: data quality in healthcare data is poor and needs rectified before anything else is even contemplated. After that, a few key issues remain:

  1. Investment: it is simply not acceptable that machines take 20 minutes to load. Investment in basic infrastructure and IT support will need to be a priority.

  2. Education: not a single clinician studying medicine today is taught anything about how AI can help them in their jobs. In fact, there isn’t even a data analytics module. Future clinicians need to understand how technology can help them and demand it exists in their local organisations.

  3. Incentives: the NHS has the potential to have the most advanced data in the world despite it currently being wrong. Prioritising making it right, and giving access to correct data should override some of the ludicrous information governance (IG) rules currently holding AI back.

  4. Adoption: when something has been proven to work it should be “pushed” out to all organisations without delay.

Reply

or to participate.