What are bedside cyborgs? Part 1 of 3
According to Elon Musk, modern technology has transformed us into a society of cyborgs who live in symbiosis with a global web of electronic devices. While you might prefer less colourful language, there's no question that the internet, personal computers and smart devices have changed the world over the course of a single generation.
In this short blog series, we'll look at how healthcare is lagging behind many other industries in embracing the Information Age. We'll also talk about why that needs to change, and soon. Finally, we'll discuss the barriers to exploiting emerging technologies at the point of care, and how they might be overcome
All in all, this is a blog about why we need 21st century healthcare professionals to aspire to be less the stereotypes of old and more like the bedside cyborgs of the future.
Which technologies hold the key to 21st century healthcare?
There are a lot of exciting things happening in the world of health tech at the moment, from robotic surgery to RNA therapeutics. In this series, though, we'll be focusing on the humble search engine. This is, after all, the innovation that has already revolutionised many aspects of modern society, placing a world of knowledge at our fingertips and fundamentally changing the way we live. We’ll cover some game-changing recent advances in search technology, and take a look at the science behind them. Most importantly, we’ll talk about what this technology could do for the world of healthcare, and take a look at some of the key barriers to exploiting it in the medical domain.
It’ll be a journey that takes us into the world of intelligent machines so convincing that one Google expert publicly declared them sentient. We’ll take a peek at what really lies behind “The Cloud”, and why it’s forecast to be a trillion-dollar market by 2030 . We’ll even touch on quantum computing and Chinese satellites. But we’ll do it all from the ground-up: no background in tech assumed, everything tied back to the world of frontline clinical care.
How do modern search engines perform with medical data?
Historically, healthcare data has been a sticky wicket for search technologies. To get a feel for the current state of the art, let’s kick off with a little practical exercise.
First, open up each of the following links as new tabs in your web browser: www.google.com, www.bing.com, www.webcrawler.com, www.ask.com. Copy-paste the following query into each of those engines and hit search:
How do I tell the difference between Waldenstrom's macroglobulinemia and multiple myeloma?
Then take a look at the results.
As long as you’re not reading this too far in the future, you should notice an immediate difference between the top-tier search engines (Google Search and Microsoft’s Bing) versus the other two. Specifically, Google and Bing don’t just serve up relevant web pages. They directly answer our question via a “featured snippet” that appeared before the full search results. That’s huge, and we’ll talk about why in a later post.
But that's not all. Take a quick look at each of the top 10 hits from the four search engines. When I ran this exercise earlier today, every one of Google and Bing’s top 10 hits contained the answer to our question. Conversely, only one from WebCrawler did, and none from Ask.com. Given that we're not exactly comparing chalk and cheese here - all four companies are big names in the space and have been developing internet search capabilities since the 1990s - that's a pretty significant difference.
The purpose of that exercise is to show that things are moving very fast in big tech right now. If we'd have run that experiment a few years back, the different wouldn't have been nearly so marked. We'll talk about what's changed later, but for now let's worry less about how Google and Microsoft have pulled ahead of the pack, and more about why it's important.
What's the impact of better search at the point of care?
As busy healthcare professionals, you might ask, do we really need to concern ourselves with this stuff? After all, a few extra seconds scrolling through a couple of pages of search results hardly seems like a big deal.
To find out, here’s a quick-and-dirty experiment I ran earlier:
Way back in 2012, Google studied a group of healthcare professionals and found they each ran an average of six medical searches per day. There’s actually strong evidence that this has increased a lot over the last decade, but I couldn't find a more recent absolute figure so I stuck with the original.
Hence, I came up with six random questions that medical professionals might ask during their day-to-day work. I also jotted down the answers they might want to find. Then I ran each question on Google and Ask.com, and compared how long it took me to find the target answers. Here are the results:
What’s the most effective treatment for ischaemic stroke?
Thrombolysis / clot retrieval
How do I manage granulomatosis with polyangiitis?
How do I decide between PCI and CABG?
What proportion of small cell lung cancers respond to monoclonal antibodies?
Response rates for different MABs for SCLC
How do I tell the difference between erythema nodosum and necrobiosis lipoidica?
Comparative description of lesions
What vessel supplies the lower mesentery?
Inferior mesenteric artery
* cases where Google did not provide a direct answer to the question via a featured snippet
This isn’t exactly tier-1 evidence, but it’ll do for our purposes. Based on the results above and the findings of Google’s 2012 study, we’ll conclude that using an industry-leading search engine could save healthcare professionals 202 seconds per day compared with a less advanced alternative.
Now let’s look at that in the context of a major healthcare system. The NHS employs about 500k doctors and nurses and about 170k allied health professionals. Extrapolating our conclusions across the whole workforce, advanced search capabilities could save the NHS (202 ÷ 60) x 670k ≃ 2.25 million minutes of clinical time per day. That’s 250k ÷ 60 = 37.5 thousand hours per day. In a system where - according to the UK government’s own Health and Social Care Committee - “persistent understaffing poses a serious risk to staff and patient safety”, that’s already a pretty big deal.
But let’s also quantify that in monetary terms. If those additional 37,500 hours were covered by agency staff at an average of £30/hour, that would total £1.125M per day, or £410M per year. Which is about the cost, by the way, of a brand new NHS hospital. And if we were to go a step further and translate those same assumptions to the US, where there are approximately 20 times more healthcare professionals, the figure would be $9.44BN per year.
If search engines are already so good, what's the problem?
It might seem like advanced search is a fait accompli. After all, we can use it any time we like. But the truth is that information accessible through internet search engines is only a fraction of what healthcare professionals need to do our jobs.
There’s good news in that: it means there are huge benefits still to be realised by implementing advanced search (and related technology) in the healthcare sector. The time and efficiency gains will dwarf the numbers we’ve just discussed. The bad news, though, is that many of those benefits depend on making personal health information (PHI) accessible to the type of technologies that Google and Bing are using. That’s proving really hard to do.
We'll talk more about both the challenges and benefits of exposing PHI to bleeding edge information management technology in the final post of this series. In the meantime, though, there’s lower-hanging fruit to be had. This takes the form of local clinical guidance, treatment protocols, organisational policies, etc.
These local documents are first-line knowledge sources for healthcare professionals. They generally supersede internet sources because they are tailored to reflect the availability of local expertise and resources, the needs of locally prevalent minority groups, etc. But they generally reside on the intranet, where search facilities can be all but non-existent. So that’s going to be our first case study for applying bleeding edge search technology to health information.
Before we get to that, though, we’ll need a couple of technical primers to set the scene. In the second half this post, we’ll start to dig into how internet search engines work their magic. In the following, we’ll look at some of the emerging AI techniques that are already powering next generation search and are set to fuel a new era of intelligent technology. After that, we’ll loop back around and start looking at how they can be implemented at the point of care.
How does Google work?
Needless to say, we’re not going to get down into the weeds here. Modern search engines are the product of decades of work from some of the brightest minds of our time. Instead, as we discuss different technologies throughout this series, we’re aiming for an intuitive sense of how they work rather than a comprehensive technical description. So let’s start with a quick mental exercise to explain our first key concept: ‘indexing’.
Imagine we’re in a hospital, doing a ward round. Even if you’re not from a clinical background, you probably have an idea of what this involves: traipsing around all the beds in the ward and consulting with each patient in turn. We keep a ward list as we do our rounds where we make a note of anything that has changed. For example, is this the same patient who occupied that bed yesterday? If so, any major updates? If not, who’s in there now and what’s their story?
Where we find a new patient, or in the case of significant updates for an existing patient, we’ll jot down a few salient pieces of information. A typical note on my (cardiology) ward list might read: Mr X, 74yo, STEMI with PPCI LAD 2/7 ago, a/w follow-up to LCx. Index event. EF 55%. H/O HTN, otherwise well. Don’t worry if that sounds like double dutch - the point is that there’s a lot of key information squeezed into a couple of lines.
Now here’s the important bit. Mr X is, of course, a human being associated with a lifetime of information. We could write several tomes about him and still only cover a fraction of what there is to say. But with some judicious summarising, it’s possible to capture most of the salient clinical features in just a line or two. In fact, the tiny snippet in italics above contains pretty much everything we need to refer Mr X for the right tests, instigate the right interventions, discuss him with colleagues, set a discharge date… and so on.
Back to the world of search engines, and the beds in our analogy are internet addresses, the patients are web pages, but it’s the same basic idea. Our ward round is equivalent to a ‘crawl’ of a search engine’s target domain. Our note-taking is analogous to ‘indexing’ new and updated pages. A good index will contain everything needed to get the job done. A bad index, on the other, can cause all sorts of problems.
How did web pages get indexed when the internet went mainstream?
Until less than a decade ago, lexical search ruled the internet roost, including for industry leaders like Google. A lexical search algorithm looks for literal matches of words, subwords or characters. It often indexes web pages by a “dictionary” showing how many times each word appears in the page. For example this page’s index might read:
.And so on.
There’s a clever trick you can use to make this approach more effective, where you divide the number of word occurrences in one web page by the number of occurrences across a group of web pages. This cancels out the massive preponderance of words like “the”, and surfaces any words that are unique to a particular page. It’s called “term frequency/inverse document frequency”, or TF-IDF.
Lexical approaches like TF-IDF are fast, effective and still widely used, but they only go skin deep. For example, faced with the search term bacterial pneumonia, your basic lexical search would rank pneumatic drill higher than chest infection. The underlying meaning of the terms just doesn’t come into it. In a domain like healthcare, where we’re positively drowning in lexically dissimilar synonyms like pneumonia and chest infection, that kind of approach can only get you so far.
That said, you can overcome some of the limitations of lexical search with a few clever tricks. For example, you could use a thesaurus to look up synonyms for your search term, and include those synonyms in the search. You can even use domain ontologies like SNOMED CT, which aim to capture both synonyms and relationships between entities. But in the end, the complexity of natural language just defies any manual attempt to marshall it into a computer-friendly structure. These approaches simply don’t scale beyond a certain point.
Google’s original competitive edge was a different type of clever trick. They counted the number of “backlinks” to any given web page and used this to enhance their lexical searches. Backlinks are links to a given web page contained within other web pages. For example, if I stick a link to https://www.nhs.uk/ here - that’s a backlink for the NHS home page. If there are loads of backlinks for a particular page spread across the web, Google realised, it’s probably an important page and should be ranked highly, and vice versa. It’s not exactly rocket science, but this new algorithm - affectionately named “BackRub” - was enough to set Google on their stellar trajectory.
How do industry-leading search engines do indexing now?
Internet search may have started relatively low-tech, but the field has evolved hugely since the late 90s. As we saw earlier, the most advanced search engines are capable of more than just serving up a list of relevant web pages. In many cases, they can pinpoint a direct answer to our question. This is really quite remarkable when you stop and think about it. It means that they have gone beyond simply matching up letter and words, and are actually making sense of the contents of web pages.
The winds of change really began to blow in 2015, when Google started to use AI in its "RankBrain" algorithm. RankBrain was a game-changer because it included “semantic” search capabilities. Semantic search involves indexing and / or ranking documents based on the underlying meaning of text, rather than simply matching letters or parts of words. Compared with lexical search, it’s a seismic conceptual shift. At its core, lexical search is basically about counting letters. Deriving meaning from words and sentences, on the other hand, requires something else entirely: machine intelligence.
What exactly is machine intelligence?
We’ll wrap up here for today, but In the next post we’ll review some major recent breakthroughs in artificial intelligence that are allowing machines to grasp the meaning behind language. After that, we’ll bring everything back into a medical context to talk about some of the blockers for implementing this technology in healthcare, and some of the ways they can be overcome.
To get a feel for just how powerful modern language processing can be in the meantime - or if you think the world isn’t scary enough already - check out this article in the Guardian written by an industry-leading AI algorithm.
See you next time!