Healthy Nuggets #15: another chatbot bubble?

Welcome back. It's been a long week at work and home so this might be a short one (although after saying that last week, it got to 1800 words, and it's raining and there's nothing on telly, so let's see...).

Also I'm not going to apologise for giving a few mentions (with nothing in return, full disclosure) this week where people have been doing great things. Onwards.

GIANT Health Event posted a very interesting and thought-provoking piece from Fast Company this past week, entitled: "Are we being led into yet another AI chatbot bubble?"

As mentioned in past Nuggets, but with particular deeper dives in #8 and #10, there was a chatbot bubble in 2015-16, in which some of us were either pioneers or sycophants, depending on your point of view. Some quarters of big tech spotted opportunities and signed up all content partners willing to experiment. The industry still thrives, although often perhaps unfairly compared to the explosion in apps from 2009-2012 mostly driven by social media. Some people (myself included) viewed chatbots as miniature app stores, with their own ecosystem and methods of communication that operating systems couldn't match.

It didn't quite work out that way but all the learnings from that period are still valid, and perhaps why people like Fast Company are taking a pragmatic view as LLMs and GenAI more generally, make language-based products fashionable again.

The article includes various views by Michelle Zhou , also quoted in past Nuggets, depicting LLM-based bots as "information retrieval systems" rather than genuine interaction.

"...chatbots are either too low or just ill-defined. Today, we mainly use chatbots as internet search helpers or productivity tools, but what enterprises (schools, hospitals, businesses, etc.) really need is something that can actually stand in for a human being"

Back in the day when Skyscanner first implemented bots on the Messenger platform, we had a very thin layer of human interaction as people already sent us direct messaging if they needed support. We maybe blurred the line between chatbot and human, but inevitably, on the first day, someone caught wind of this and asked what the respondent was wearing.

On the other hand, the early notion that Alexa might respond differently if phrases are appended with "please", showed that at least some people are willing to treat bots as if they're the real thing.

The promise of AI assistants may be mostly about the extensibility of human knowledge. Businesses with a limited number of employees need to service a far larger number of customers. Advanced AI assistants offer a chance for all users to get high-touch one-on-one attention from businesses or institutions. But the soft skills Zhou mentions may be part of a future wave of chatbots—one that we’re still years away from experiencing.

We've been working, as mentioned in Nuggets #14 last week, on the grey area here between structured data (known sources or curated data made available by users) and unstructured data (the colour around that data which makes it personal, and is thus pretty unpredictable) and it's by no means a solved problem as Zhou says. If I regularly get a headache on Saturdays, and then mention in passing that I was watching boxing until 1am, then that's not exactly regular structured data useful to an AI service. But knowing that I didn't get to sleep until 2am is the tangible result of it borne out by data. Do they belong together?

No doubt we'll be returning to this topic many times on Sunday evenings.

Continuing that theme, an article by Maya Zlatanova (hat-tip to Martin Sandhu for sharing these things in his weekly digest, as always) shows early promise in extending LLMs to delivery more specific information. In this case it's leveraging documented findings from research, effectively indexed by AI, to give a GPT-like experience for people at the feasibility and planning stage for clinical trials.

ChatGPT is great but had little to offer as in the clinical research industry we need to be able not just to get an answer to we need a verified answer that we can track back to data sources we can trust.

This is when my team at TrialHub saw a great opportunity to customize the technology behind ChatGPT and combine it with other technologies we've built throughout time and our pre-validated data.

This is ringing bells, as like many early adopters, I tried to use GPT for research but found (a) its sources are often out of date - although this past week it appeared to catch itself up to mid-2023; and (b) if pushed, it might just make them up. It even created newspaper articles with fake URLs for me and took 5 conversation turns to admit they might not be real.

In the world of clinical evidence, this simply isn't on, so extending an LLM and specifically training it with known good data, makes a ton of sense. More below.

And this past week my friend Stefan Sabev and his co-founder Stephen M. Walker II went public with their tooling platform which helps to do the work outlined above. It's improbable that ChatGPT or any other general LLM will cater for many specific requirements like health journals, but that doesn't mean an LLM can be used to make them more searchable.

Klu.ai - to quote the Sifted article from their launch day, is there to "build the picks and shovels for GenAI" and gives many content owners a head start in adoption. Standard integrations for stores such as Jira, Drive, Dropbox and individual flat files, then picking your own particular AI partner or LLM to use as the foundation, can get a simple retrieval function going in a few minutes. I've been trying the private release for months and can vouch for that.

Stephen also showed the domain expertise offered by Klu with a deeper dive into AI safety, particularly in the wake of a meeting with public officials in advance of the UK Government's AI Safety Summit next month.

As always you should read for yourself but I'll give you a brief takeaway to get you hooked - thinking about how the media continually feed us doom-scrolling about AI coming to kill us, when it generally still makes a lot of stuff up (as above) or can't think outside of specifics (as above). Perhaps we're at the 1994 moment where someone realises you can put nuclear bomb instructions on the Internet and someone else might find it. Anyway, as Stephen writes:

It's hard not to review the facts of 2023 and not think that we were all played. If AI is a great risk, and GPT-4 is taking over the world... then where are makers of the technology behind HQ misinformation via deepfakes, identity theft, or autonomous driving snafus?

From a safety perspective, ChatGPT garnered more headlines for its ability to hallucinate court cases and write (or not write) student homework assignments.

A clear-eyed skeptic might think we're watching the early innings of smart tech CEOs fast track regulatory capture.

Thankfully the Internet remains largely unregulated, outside of genuine crime, and hopefully AI will be allowed to find its way without big tech dictating the path.

There has also been a lot of talk in Nuggets about distributing healthcare by using technology to open up self-monitoring and reduce the burden on clinicians. Where AI can't replace doctors and nurses, perhaps humans, guided by tech, can do it themselves.

Luscii - an OMRON Healthcare service was in the news this past week with its platform being formally recognised by NICE - National Institute for Health and Care Excellence as a suitable platform to do just that, in the domain of acute respiratory conditions.

Such platforms were shown in the NICE research paper to save ￡872 per person compared with in-patient care, and ￡115 per person cheaper than other forms of at-home care, which is a significant benefit for a critical and growing range of healthcare. I've mentioned Lenus Health several times.

And another quick mention for moment: time, re-imagined which, as its founder Filip Filipov often says, is 'relentless' in continuing to improve week on week. The latest release focuses on ease of scheduling and notification for critical calendar changes... and as I found out it just worked fine in Slack, this could be very useful indeed:

Please do download the app if you're an iOS user and help the team with feedback; the insights back to you, and ease of scheduling your time, will be worth it.

Additionally, I've been using Holly Health recently and despite being one of those people that installs an app and then breaks the habits quickly, I was keen to give it a chance and allow me to track certain 'habits', in Holly parlance, that I'd set up for myself.

As a hater of notifications, as mentioned a few Nuggets ago, this had the potential to be a quick uninstall - but having been through a very pleasant onboarding experience, where even reading the T&Cs was both useful and understandable, it's a very impressive product.

Getting notifications right is hard and although I might confess to not performing all the things the habits tell me about, the reminder isn't the worst thing. The plan is to go from a passive month of watching how the app interacted, to using it more in anger, and then see how much it's actually helping.

It's not been the best year for products named after birds, but Grace Gimson and the team seem to be onto something with this one :)

Grace very kindly shared this project on LinkedIn earlier in the week too - something that's been at the back of the mind all year at Waracle where we believe sensor-laden devices such as smartphones, should be front and centre of the healthcare revolution given the data they can collect.

This time, there is clear evidence that by augmenting existing techniques with audio data, detection of diabetes in a patient could clearly be enhanced.

The material presented here reports a promising application of voice analysis for T2DM detection. Although the results are encouraging, further research with larger and more diverse cohorts is required to validate its effectiveness and generalizability.

Nevertheless, our findings highlight the potential of voice analysis as an accessible and cost-effective screening tool. An implementation of voice assessment could aid in early intervention and management of T2DM, and continued development could reduce the rising burden of the disease and improve health care outcomes.

The research paper is a long read but it's worth digging into - maybe you could tool up some AI to get a summary...

https://www.mcpdigitalhealth.org/article/S2949-7612(23)00073-1/fulltext

1800 words again. That'll do it for this week but as always if you see something worth a mention, or even want a mention yourself, please do drop me a DM on here.

Small print: This newsletter goes out to subscribers and across LinkedIn most Sunday nights around 7:30 pm. Feel free to contact me if you've seen or are creating something interesting in digital health. I work for?Waracle, but all opinions and content selections are my own. Anything in which I have a work or personal interest will be declared.

Cover photo was generated by Playground AI using the simple prompt "useless chatbot".