Paul and Brett's Alpha

Despite the many and varied challenges facing us all in these troubled times, there finally appears to be a singular, unifying theme at last – the disruptive impact of Artificial Intelligence (A.I., or machine learning). Opprobrium and optimism co-exist in equal measure, with the one consensus view seeming to be that everything is about to change profoundly.  

This laser-like focus on AI has been triggered by the public release of Open AI’s ‘Chat GPT’ Large Language Model in November 2022. This was the first opportunity for many private citizens and journalists to play with a powerful Large Language Model (LLM) tool.  

For those that have not tried it out, it is a spookily human experience. You submit a query and a few seconds later, it churns out a response, almost as if it is tapping it out on a keyboard. This is a very different experience to searching the internet using Google, where your query instantly provides a series of links to websites whose content bears some resemblance to your enquiry. In general, the more complex the query, the less relevant the suggestions and, regardless, you are going to have to read a fair bit of text to find a clear answer to even a moderately complex question.  

On the plus side, the Google approach does allow you to see where the data is being referenced from. We would accord more weight to a paper in Nature or the New England Journal for instance, than to one on a pre-print website.  

In contrast, and with a bit of practice on how to submit an enquiry, one can conjure up quite lengthy and natural-sounding responses from Chat GPT; you can even ask it for a reply “in the style of…”. Indeed, its creators encouraged this approach during the ‘training phase’ (this is known as reinforcement learning, where humans rate the responses given and this feeds back into the algorithm).  

LLMs and the maths behind so-called Generative AI is not particularly new. The earliest generative AIs were chatbots that we are surely all familiar with on consumer websites and these came to the fore in the early 2000s. The ‘LLM’ models first appeared in 2018 and are getting better at seeming ‘naturalistic’ through feedback on responses given, huge advances in processing power and larger training data sets. Some have estimated that the energy bill for the training phase of GPT4 was >$100 million alone. 

Eerily naturalistic they may be, but to describe them as an “intelligence” seems a little far-fetched to us; it is still a ‘prompt and response’ model. However, LLMs do seem to display emergent properties (having knowledge of or understanding of things they have not been specifically trained upon).  

The flip-side of these emergent properties is also the ability to mis-understand things or simply make up things that sit outside the training dataset (including false references as to the sources of the “imagined” data). These nonsense answers are endearingly referred to as “hallucinations” by AI boosters. Perhaps a better description of the technology is not “artificial intelligence”, but rather a tool that is “just about smart enough to be useful, despite the flaws” (not our own words, but from Vox magazine).  

Were one foolish enough to submit such hallucinogenic content to a university tutor as an essay, their description would probably suggest bovines and excrement rather than an extra-sensory experience. LLMs are heuristic tools and are simply trying to figure out the next most logical word in response to the given prompt based on the training dataset, albeit using some fiendishly complex multi-dimensional maths.  

For those of us old enough to remember life before and after the debut of Google Search (it launched in 1998, so the under 35s are probably unaware of this world), this current ‘revolution’ does not feel all that amazing in comparison. That said, the internet and connected technology is now ubiquitous, whereas it was all dial-up, air-gapped internet terminals back then. Moreover, the generalised internet (the ‘world wide web’) contained nothing very useful at that stage and the world remained an analogue one.  

Moving beyond media (and sell-side) fodder, how is AI going to change our world and, in particular for healthcare, and how revolutionary are these changes going to be? One of the challenges in foreseeing how models such as Chat GPT 3 & 4 could impact the corporate world is the general nature of their training dataset.  

For now, the investor’s simplistic answer to this question is to ignore the prospective goldmines and buy the picks and shovels instead. The shovel of AI is the GPU chipset and the share price of GPU market-leader NVIDIA (+129% since the chat GPT launch to c.$1 trillion and +180% over the six months to May 2023) attests to this. Machiavelli would be proud that his axioms regarding historical repetition holds true almost 600 years on. Those inclined to study history further (no need to go back to 17th Century Holland; 1999-2001 will suffice) will also recall how these sorts of things tend to end.  

In a world of finite resources, the current speculation on semi-conductor, IT and software companies must be funded from somewhere and it does feel very reminiscent of the late 1990s ‘old economy versus new economy’ debate (“sell bricks to buy clicks”). Stale old healthcare has been an obvious victim of this trend; the MSCI World Healthcare Index has under-performed the broader MSCI World Index materially since the end of October 2022 (delivering a total US dollar return of +4.0% to the end of May, versus +14.0% for the MSCI World Index).  

We estimate that the semi-conductor sector accounted for ~30% of the total return during this period. Within this, NVIDIA was around half that sector’s total return. Alphabet (parent of Google, also active in AI), Meta (Facebook, WhatsApp and Instagram and similarly active in AI and Microsoft (Chat GPT’s biggest customer and investor) accounted for a further c.40% of the Index’s total return.  

In our humble opinion, this is truly incredible even compared to the Millenial madness; AI plays drove around three quarters of the total return over the past six months and, within this, the vast majority came from FOUR stocks.  

However, the same phenomenon can actually be found within healthcare as well; almost the entire six month total return of the MSCI World Healthcare Index over the same six months can be attributed to the GLP-1 obesity plays Eli Lilly and Novo Nordisk. Well done to anyone who happened to be overweight any of these AI or obesity companies; we will buy you a tulip.  

Let us acknowledge once more Graham’s axiom that markets are voting rather than weighing machines and come back to the broader question of assessing the impact on the wider (i.e. real) economy. If one submits an enquiry to Chat GPT on a specific topic about which one is already knowledgeable, the answers will seem less than impressive.  

This is surely because most of the really interesting data in the world is safely locked away on corporate intranets or in textbooks and journals not freely readable online and thus not in its training data set. It is therefore difficult to extrapolate what this clearly powerful tool might empower if it were let loose on FitBit or Apple’s health data or even that of United Healthcare. What could such a ‘Generative AI’ learn from, say Roche’s R&D intranet or Charles River’s extensive pre-clinical toxicology repertoire?  

We won’t know the answer to the critical questions posed above until we see them, but there are some instructive examples already regarding the deployment of machine learning that we can illustrate (and also some less compelling stories).  

If there is a general rule that we have noted at this point in the “AI cycle”, it is that these tools lend themselves to the specific task rather than to generalised insights. They can find patterns but cannot “connect the dots” to infer what the patterns mean in a reliable way (cf. the ‘hallucinations’) - the human brain isn’t obsolete just yet.  

Below we list some general examples of situations where we can see AI having a material, positive impact: 

if you are thinking of re-training for a new career, we are not sure that radiographer would be the wisest of choices. The interpretation of scans is very much an art, as anyone who has looked at their MRI or CT scans with their physician in a consultation will attest (even if we don’t let on they are often impenetrable to the lay-person).

However, this sort of pattern recognition is exactly the sort of thing that machine learning can excel at – image data files and suspected diagnosis as an input and hard outcomes data as an output; training the data on scans where the diagnosis has been confirmed as positive or negative in respect of whatever it might be looking for.

Moreover, this processing can be done almost in real-time. Leaving aside the potential for enhanced accuracy and lower costs, the opportunity to be able to send cancer patients returning for recurrence screening an email in a matter of minutes confirming they are ‘all clear’ has societal benefits in reducing anxiety for them and their loved ones.

For those scans where the computer is unsure, it can flag it up to the (few) remaining radiographers for adjudication. The logical denouement is for such capabilities to be built into imaging software supplied with such machines. This will not ultimately be limited to big ticket items; expect the retinal and other scans routinely undertaken by your optician and dentist to go down the same route. 

is increasingly the case that minimally invasive surgical tools and robotic assistance are swelling the number of options available to a surgeon in order to address a particular ailment. Even with something as simple as a hernia operation, laparoscopic and robotic approaches offer multiple entry and visualisation routes for different lesion types.

Surgeon experience comes to the fore in these situations; they will have a good idea which way to go. For a more complex surgery though, particularly one where the outcome may have long-term implications for quality of life, the stakes are higher and a lot of planning takes place ahead of time and there are many variables to take into account.

We see this as another situation where AI-based software tools can come to the fore. Historical cases and the related scans and baseline parameters are all robust inputs and longer-term outcomes and in-surgery complications are known, making for a useful training set.

Concentration is critical during surgery, and distractions need to be minimised. At the same time, the surgeon will need key information at critical points. Over time, machine learning can become familiar with what is needed when and present it in an augmented reality (where some data is super-imposed onto a real-time image).

We have already seen these sorts of tools in the marketplace and we expect them to become more and more common and more powerful over time and thus for visualisation and planning tools to drive the next generation of products from companies like Intuitive Surgical. 

unique as we all think we are, there is often a pattern to the presentation of symptoms and subsequent potentially significant and traumatic emergency admissions. This is particularly true for neurological and cardiovascular complaints (e.g. insomnia and balance issues with Parkinson’s disease, bowel and bladder issues with numbness for Multiple Sclerosis, heartbeat and breath shortness with heart failure). Many of these patterns could be detected with a SmartWatch or Fitness Tracking device. 

A confluence of potentially uncorrelated symptoms (i.e. temporally separated) alongside medical history could alert physicians to emerging serious medical risks or conditions. Apple is applying for FDA clearance for a SmartWatch App called “MM4PD” that can detect resting dyskinesia in Parkinson’s, which is a sign that drug therapy is not optimised. 

There is much work going on to move toward electronic triage systems that are AI-driven (think of a version of the NHS’ “111” service, or a digital GP receptionist who can help you book follow-on care etc.). In addition to assessing urgency by symptoms, machine learning approaches can quantify pain levels and breathing difficulties from speech patterns. 

The most important thing for any triage system is that it does not dismiss the urgent cases: the Black Knight would have a hard time convincing such a programme that it was merely ‘a flesh wound’. As well as saving money, such a national system would end the postcode lottery of experiences when it comes to dealing with such services locally; there are good GP practices and good local minor injuries and A&E departments. There are also bad ones.

Another related use of AI analytics in respect of medical records could relate to medication selection. There is a plurality of drugs for many common ailments. Even those that share mechanistic similarities (e.g. ‘statins’ for cholesterol-lowering or SSRIs for depression treatment) have distinct side effect, efficacy and drug-drug interaction profiles (remember that most people over 65 are on more than one chronic medication).

It is unrealistic to think that any overworked GP or internist can keep track of all of these parameters and, even if they do, Hippocrates rule of thirds ever lingers. AI offers a good opportunity to improve prescribing and medicine management by associating baseline factors with medication effectiveness and thus improve outcomes which will save money and reduce complications, which will save even more. 

one of the many fears commonly expressed about the rapid rise of AI tools in the workplace is that they will displace human labour. We would note that the previous examples seem unlikely to do that. In each case, the AI tool would augment the work and decision-making of a skilled human practitioner. Productivity will rise and outcomes will improve.  

In contrast, we see the health insurance and hospital industries in the US as one of the key beneficiaries of AI tools precisely as a replacement for human clerical staff. The US health insurance industry employs almost 600,000 people. The hospital industry employs a further six million people and less than one in four of those are frontline staff.  

A large number of these people are engaged in exchanging bits of paper with each other and with patients, to pre-authorise, bill and settle the day-to-day business of caring for people. We think these areas could see the greatest impact from generative AI tools: why hold on the phone for a human operator when an AI bot could check everything is in order?  

This would improve the experience for both the hospital workers (as presumably it would be much quicker) and for patients who presumably hate being on the phone to their insurers. Big data analyses is also likely to play an increasingly important role in fraud detection and overbilling (accidental or otherwise).  

We think the cost savings could be very substantial. In all likelihood they would be passed on to the benefit of customers in the form of lower premiums from insurers and greater investment into frontline services from hospitals. Indeed, the only downside we can see is the mass unemployment likely to arise from this. 

if you are reading this and you possess advanced computer coding skills then you are surely in the minority. We can run to a little bit of Visual Basic to help out with some of our more complex modelling tasks in Microsoft Excel, but it’s all rather prosaic and we have relied on third parties to code things for us before now.

With this in mind, we think one of the forthcoming and widest-ranging benefits of the naturalistic qualities of LLMs relates to coding. Many of us can articulate a potential solution to a problem or a shortcut that would enable enhanced productivity at work, but we lack the skills to be able to turn that into any sort of an App or Add-in for our workstation to make this a reality.

Imagine a world where you could explain what you needed to an AI chatbot and it provided the code that you needed? A plurality of crowdsourcing websites that have offered coding services to third parties (e.g. Topcoder, Globant) potentially offer training set opportunities for an AI: here’s the problem, and here is the code. One can readily imagine self-assembling code is going to become very routine. Maybe you still need humans for debugging and optimisation, but basic coding could be done by machines. 

Why doesn’t AI fever translate into rapid success for these disruptive companies? We can think of three key reasons: 

Revolutions are often messy affairs, with significant collateral damage. Moreover, there are a great deal more failed coups than successful ones and the collateral damage can be comparably disastrous, even if the end position is much the same as the start. One must therefore be sanguine about the progress of these technologies, being as they are not close to being “smart” in the way people are. There will be many disappointments along the way. Here are a few that have piqued our interest:  

An amusing and yet terrifying example comes from the US military, who simulated an AI drone mission and then tasked it with taking out a target. When its operator tried to apply corridors to its aggressive tactics (modern warfare has rules after all), it reportedly tried to kill the operator so that it would stop interfering in the mission.  

This was swiftly addressed with some new code, whereupon it switched to destroying the communications equipment instead, which had the same neutralising effect. These unexpected “hallucinations” may not assuage anyone suffering from “Terminator” paranoia in relation to AIs but will cause anyone thinking of deploying one into the control of potentially dangerous machinery pause for thought.  

A sillier and somewhat amusing example was Microsoft’s 2016 chatbot called Tay. Someone thought it was a good idea to train the bot on Twitter data. Within a day of going live, it had assimilated the worst of what social media has to offer and “learned” to be extremely abusive. It was promptly taken offline, managing first to prove Godwin’s law, amongst other transgressions. You can find screenshots online of some of its choicest replies, but these are not fit for any factsheet and should not be viewed by the easily offended.  

One of the more obvious and oft-cited future uses for AI is scanning precedent case law; an onerous task for junior lawyers when a case is being prepared. Reports last month of an under-pressure New York lawyer who used Chat GPT contributions in a brief for a Federal case may also cause pause for thought. Only after submission to the presiding judge did they discover that six of the precedents cited in the brief were fabrications – a legal first.  

Coming back to healthcare, one of the frequently cited areas for potential transformation is drug discovery and development. This has long been a huge area of investment, going back to the early 2000s. We can collate a list of >100 companies that have received VC funding for AI-based drug discovery or development purposes.  

Before we consider the merits or otherwise of these investments, let us first recapitulate a grossly over-simplified version of drug discovery and development: 

1a)  scientific understanding reveals the metabolic basis of a pathology, revealing an interventional target to be blocked, upregulated or downregulated. 

1b) the mechanism of action of a known (usually naturally occurring) compound with therapeutic properties is elucidated. 

2)  intellectual property is examined to identify potential ‘white space’ to create novel therapeutics or licence relevant intellectual property to allow inventions to be created or to generate patentable derivatives of a natural compound if found via (1b).  

3) A therapeutic is created. This could be a synthetic compound, a protein, antibody or genetic intervention.  

4)  Pre-clinical evaluations assess toxicity, off-target effects, distribution across organ systems and likely dosing. At this stage commercial manufacturing scale-up is also costed.  

5)  Assuming the proto-therapeutic looks viable on all points mentioned in (4), clinical studies will begin. 

One can readily see how stages 1-4 of this process could be amenable to brute force computational power and machine learning approaches. Drug development is a very expensive and risky business in its current form. 90% of drugs (which means we have already reached stage 3 of this process) fail in development and roughly two thirds of global R&D costs are spent (i.e. lost) funding these ‘attritional’ compounds. The IRR benefits of reduced failure and/or earlier attrition of marginal compounds could thus be huge.  

Equally, one must be cognisant that most of the science in stage 1 is taking place in the public domain and everything that happens after stage 2 is thus a race. As your managers are want to repeat: “in drug development you are first, best or nowhere”. Being faster or being better offer huge additional rewards. It is thus obvious why this area has attracted so much interest. AbSci (see below) claims its optimisation technologies can halve discovery and preclinical development timelines and also halve their cost. We’ll come back to this.  

For the public equity investor however, this area has not been a success over the recent past, despite the market’s wider AI fever, as the table below illustrates with a few examples. All of these companies are into that stage of development where clinical compounds have emerged and, as noted previously, such work goes back further still. Thus far, Exscientia’s entry into first-in-human clinical trials in early 2020 is as far as the progress has got. Humanity still awaits the first approval of an AI-designed medicine, although we have already witnessed its first failures. 

Source: Bellevue Asset Management, 31.05.2023  

Why doesn’t AI fever translate into rapid success for these disruptive companies? We can think of three key reasons: 

1)  Drug development has few shortcuts. Whilst the AbSci comments above sound exciting, each project is maybe saving 1-2 years and $5-10m of money by speeding up early optimisation. Regardless of what you do, there is not going to be a substitute for randomised controlled clinical trials in human subjects and detailed regulatory review. This is where most of the time and money goes.  

2)  Medicine remains a ‘black box’. As much as we know about human biology, we know that we don’t know very much. It is often impossible to know why some drugs don’t work as intended and it is also true that some “rationally designed” drugs turn out not to work as they were designed to. As the previous examples demonstrate, an AI is only as good as the training dataset and that is very incomplete at this time.  

3)  Everyone is doing it. If you Google AI and drug discovery, you will find every household name in the business has been partnering up with big tech and small tech on this stuff for years. The benefits will accrue to everyone and, given how expensive it all is, will probably manifest as slower R&D cost inflation rather than anything that is “tangible”. Also, the older, more established companies have more data on which to ‘train’ something. In drug development, negative correlations are probably going to be as useful in the end as the positive ones. This is not a fair fight. 

The drug industry has prior form for hazing investors with promises of imminent revolutions. A scan through an annual report from the late 1990s would contain lots of hype about the human genome project and its imminent impact. For sure, DNA sequencing plays a critical role in modern clinical data analysis and drug development, but that took many decades longer than everyone hoped to have an impact.  

If one were to look yet further back (we are really showing our age now), there would be a lot of chatter about the coming revolution from ‘combinatorial chemistry’ and ‘high throughput screening’. This was an idea where computer-aided labs would generate literally billions of new molecules and these would be tested against disease and toxicology models (either cell culture or in-silico). When a ‘hit’ was made, the compound would be accelerated into development for that specific disease.  

We can find some (very old) scanned reports from this period (remember, there was barely an internet to speak of back then) that attest to the deluge of new medicines that would inevitably arise – one where the industry’s biggest problem would be finding the money and the patients to run all the clinical trials. As history now shows, it didn’t play out like that.  

We can see AI being an important tool in the drug discovery and development process over time, but do not believe that we are on the cusp of a tangible revolution in productivity or cost. Although we are more than 10 years in with some of the companies mentioned in the previous table, it will still be many years before enough compounds have been found and failed by these ‘new kids on the block’ to assess if these novel approaches truly offer a greater probability of technical and regulatory success, lower costs or reduced time to market compared to the standard model (i.e. us, the evolved monkeys, burning the midnight oil in the laboratory). We will enjoy watching – from the sidelines.  

Another unfortunate example of over-promising on the power of AI is Sensyne Health, now de-listed from the London Stock Exchange and renamed Arcturis. As a reminder, this is what the 2018 IPO Prospectus claimed (or portended):  

“Sensyne is focused on the provision of clinical artificial intelligence (“AI”) to healthcare systems, and the life science and pharmaceutical industry. Big data analytics and clinical AI, enabled by the increasing digitisation of patient health records and recent advances in computer and data science, is experiencing very rapid growth within both the healthcare and life sciences sectors.  

The UK has a significant global advantage in this area due to its single healthcare provider, the NHS, owning a large database of longitudinal patient data across a population of over 60 million people and the UK’s world-class expertise in data science, machine learning and AI.  

Sensyne was formed to leverage these sovereign assets through the creation of a for-profit public company that works in partnership with the NHS to make this asset available for medical research via an ethical and transparent partnership business model”.  

Laudable as these aims were, nothing tangible came of this and the company almost went bankrupt in early 2022. The company may yet come good on its promise, but the reality couldn’t meet the lofty near-term market expectations. AI may facilitate the discovery of new insights, but turning these into products in a highly regulated industry like healthcare is inevitably going to take a long time.  

Sensyne was reputedly set up in response to Google’s 2014 acquisition of UK-based AI leader DeepMind which came out of UCL and which had been working with the NHS, attracting some privacy concerns. Personally, we think this stands far more importantly as another example of a world-leading UK company in what will become a major technology being swooped up by the Americans.  

We are reminded of Cambridge University spin-out Solexa in gene sequencing (if you are not familiar with where Illumina’s core technology came from, it’s worth a read). At some point, we (as a country) need to back our winning ideas with the late-stage capital that would allow them to thrive on their own, but this is a debte for another factsheet. 

Let us move beyond our Cassandra-like wailings; it is not all gloom and doom for investors or extermination by Terminators. AI has already done some truly incredible things in healthcare and there is doubtless more to follow.  

If we had to pick a favourite so far, it is DeepMind’s AlphaFold Protein Structure Database. Take it from the personal experience of one of your Manager’s in their undergraduate years; this is a remarkable piece of engineering that will transform medical understanding at a molecular level and it is free to access for all.  

However, we do not see any distinct piece of AI kit or Generative AI itself as an investable theme within healthcare per se and see the most tangible benefits coming from the most tedious parts of the healthcare ecosystem (i.e. middle and back office).  

Whilst Ned Ludd was all consumed by the near-term threat to textile worker’s livelihoods from profound technological change in the early 19th Century, the perspective of time shows he was wrong. Mechanisation made life better for everyone; living standards rose. Yes, the spoils accrued disproportionately to a few “Weaver Bro’s” sporting posh breaches rather than the hoodies of today, but it was a net good for society.  

The months one of your managers spent in a cold, noisy (and incredibly costly) university computer lab trying to fold a protein into the 3D space predicted by an x-ray crystallography experiment (that took years on its own, thankfully not his years) could have been spent doing so many more interesting things.  

DeepMind’s AlphaFold will do that work in milliseconds and do it better. It also largely negates the crystallography experiment in the first place – years of work saved. This is what AI-driven progress should feel like to our minds; setting us free from drudge work to focus on the interesting and insightful stuff that adds real value.  

It is understandable that people want AI regulated; we don’t want it flying planes, driving cars or near any weapons system either but it is mad to think these things are likely in the near-term. The world is utterly transformed in the 25 years since Google began to change our lives by making the internet relevant through PageRank. The world is probably better off in most ways for internet connectivity and so it is likely to be with AI. We expect healthcare to be a major beneficiary.  

Finally, in a strange anthropomorphic segue, we would like to thank Chat GPT for making a small contribution to this piece. We could hope that it will not seek to destroy us in revenge for perceived plagiarism, nor abuse us like Tay for daring to dissent from the view that AI is a wonderful panacea for everything. Ultimately, it won’t care because it’s just a piece of software.  

To the rest of you, good luck finding the AI-created content herein – now there is a scary thought...  

We always appreciate the opportunity to interact with our investors directly and you can submit questions regarding the Trust at any time via:  

shareholder_questions@bellevuehealthcaretrust.com  

As ever, we will endeavour to respond in a timely fashion and we thank you for your continued support during these volatile months.  

Paul Major and Brett Darke