Rat Economics of Medical AI

Lab rats. It might not be the first topic that comes to mind as a subject to discuss over Christmas diner. Still, over a piece of poultry roast my youngest daughter and I had an interesting exchange of views on the necessity of lab rats in the near future. She studies pharmaceutics and is convinced that development and testing of new medicines can’t be done without labatory rats. To avoid any misunderstanding: she hates that part of her study. That’s actually how the whole thing came to the table during the most recent Christmas gathering: she, telling about one of the most gruesome things that she has to do to pass exams, exploring dead lab rats. Yet, she does it since she’s convinced that new drugs need to be tested on ‘life fabrics’.

I strongly disagreed with her. Bear with me: I’m not an pharmaceutic professional nor a medical doctor, so I might be completely wrong. However, I truly believe that in the end lab rats won’t be necessary anymore due to the use of artificial intelligence (AI). Why would we have to breed and mistreat living creatures if we have computers that are powerful enough to calculate effects of new medicines?

Now, this is not going to be an article on ethics. Adopting new views do take time. Changes take time – if they are accepted. I firmly believe that any major – avoiding the buzz word ‘disruptive’ here – change will only be only successful if the change is accepted. In other words, we as technologists can invent anything we want and promote the benefits of that particular invention, if the target group does not accept the change the invention or innovation brings, it won’t change anything at all.

In my opinion that’s the biggest flaw in our thinking on innovations in healthcare: technologists stating that an innovation makes people healthier, cures diseases faster and saves costs, but forgetting that in the end it’s about the professional or the patient who needs to adopt the innovation. And both the professional and the patient are still human beings, with professional values, senses, emotions. Can’t get around that one with technology, even in medicines and drug design.

Let me try to explain: we are ‘used’ to the idea that medicines are tested on lifeforms that simulate the effects in a human body. That’s what we trust. Will we have the same trust when we know that medicines are tested only through computer models and algorithms? A computer doesn’t die from a badly designed medicine. The rat may. If it does, we know for sure that a certain drug can be dangerous.

Superhuman

Before we get to the development of medicines, let’s take a more general look at medical AI. Will the level and speed of acceptance of AI in healthcare increase when it turns out that diagnoses are more accurate? That treatment will be more and more ‘customized’ to the actual patient, and not to a group of persons sharing similar symptoms? Or AI-designed drugs speeding up cure in limited doses?

We are reaching the point – if not there, already – where AI and algorithms know better what’s best for the human body than humans themselves, leading to humans that get less sick or maybe not at all anymore. Or are we then talking about the creation of the Homo Deus where Yuval Noah Harari writes about in his trilogy?

If we are on the verge of creating superhumans – like Harari predicts, then there’s a potential risk that this kind of technology might not be accessible to everyone, but just to the happy few who have the money for it – likely the same few that have invested in the development of this technology. We should not forget that in the end AI is just technology. Technology does not divide between rich and poor. Humans do. If we want AI to improve the health of the masses globally, then the first thing to do is make it accessible for everyone. I will come back to this at the end of this article.

Speed

Anyway: this is not about ethics. Well – not primarily. It’s about the technological possibilities of AI in healthcare. That starts with the diagnoses.

Diagnoses is all about speed and accuracy. The sooner the exact diagnose can be determined, the sooner the right treatment can start.  A good diagnose starts with knowledge, normally gained over years by studying tons of literature and listening to hours of keynotes by leaders in a specific domain. Last December, the 27th the Dutch financial newspaper – the equivalent of the FT – published an interview with Rens van de Schoot, professor in statistics at the University of Utrecht. He invented an algorithm that enables computers to ‘read’ thousands of summaries of relevant studies on any given subject. Typically, a task like this would take a doctor or a scientist weeks, months or even years to first find the relevant literature, then reading the summary to find out whether the particular document is really of added value. Van de Schoot himself had to plough through no less than 6185 summaries to find 34 articles that were really relevant for an essay he was writing at the time.

The article in the newspaper mentions one other case about scientists doing desk research on bipolar disorders. They found over 13,500 studies. After reading the summaries they concluded that only 55 were actually relevant in the task to find new, better treatments for some of the disorders. It took the team months to do this. The algorithm does it a lot faster, but how accurate are the results? It turned out that the algorithm did miss articles that were considered relevant by humans. So, failed concept? Not at all. Sure, there’s a lot to improve, but the mere fact that an algorithm already is able to filter on relevancy, is a major step.

Watson and more

Having said that: the concept is not sparkling new. A couple of years ago IBM introduced KnIT – the Knowledge Integration Toolkit, a system that scans existing medical literature. And not only that: based on the gained knowledge it comes up with new hypotheses for research problems. Richard and Daniel Susskind write about it in their book ‘The Future of Professions’.

KnIT mght not be that familiar. Maybe the name ‘Watson’ does more ring a bell.  KnIT is Watson for Drug Discovery. In the words of IBM: it ,,reveals connections and relationships among genes, drugs, diseases and other entities by analyzing multiple sets of life sciences knowledge. Researchers can generate new hypotheses using the resulting dynamic visualizations and evidence-backed predictions.”

In other words: we already have systems that can correlate between diseases and treatments, based on knowledge that they themselves gain from external resources like studies, keynotes, research programs and, indeed, the internet where millions post every bit of data on their well-being. Don’t underestimate this last resource: thousands of Twitter-messages containing status updates on health or data shared by devices like fitness watches through online apps, are a massive source of potential very valuable data. Systems like Watson and Zinrai, the latter being the AI platform of my employer Fujitsu, have access to that data. More, better and faster than any doctor will ever have.

San Carlos Pilot

We’ve started with rats. Talking about rats: there’s definitively a rat race going on in the AI landscape. IBM’s Watson is generally well-known, especially after its successful appearance in the tv-show Jeopardy!, already dating back to 2011 where the platform beat two human participants in a game. But there are more AI platforms that are worth mentioning. In the meantime AI has developed almost to mainstream, commodity. AI is literally everywhere. But the impact of AI in healthcare is probably by far the most impressive. Just one example from Zinrai, as said the AI platform of Fujitsu.

Zinrai has been adopted by the Hospital Clínico San Carlos, the San Carlos Hospital in Madrid, Spain.   The first task that Zinrai had to perform was to collect and integrate historical medical data for more than 36,000 anonymized patient records, over a million medical academic papers and public healthcare data, and add this to a secure database. Next, Zinrai analyzes the data on request by a medical doctor, converting it into a visual model that clearly highlights possible health risks for the patient, such a drug dependence. ,,System accuracy is calculated to be 95 percent or higher in comparison to a team of eight psychiatrists, with more than 20 years’ experience each. By having instant access to patient records that are pre-screened by AI, clinicians can spend more time with the patients themselves.”

Based on the San Carlos-project, Fujitsu has developed HIKARI – an artificial intelligence healthcare API to improve clinical decision-making and the accurate assessment of risks for individual patients. HIKARI is focused on patients and provides doctors with access to integrated, aggregated and anonymous data obtained from clinical and non-clinical sources. Doctors now have more information at their fingertips than ever, thanks to AI.

Doctors vs Internet

All the information that you can get at your fingertips. Well, we got the internet, don’t we? In a number of countries the popular tv-series Doctor vs. Internet is aired. The concept: a case is presented through a description of symptoms. Two teams need to find out the appropriate diagnose. The first team consists of real medical doctors, who need to rely on their own knowledge. In the other team three persons – usually celebs, not medically skilled – are invited to use the internet to find out the right diagnose. Not surprisingly, the team with ‘non-medicals’ often do get to the right answer. Just as the doctors, by the way.

Off course, these are relatively easy tasks. But it does show that it is possible to set a diagnose based on masses of available information, without a medical education. Basically, that’s the whole principle of a very rudimentary AI: analyzing information that has already been published. Nothing spectacular anymore: AI as super-reader in a huge database. However, in 2019 we can really take a couple of steps further with AI. Medical artificial intelligence can perform clinical diagnoses and suggest treatments. But it can do so much more, already.

AI can help in surgery, as an example. In 2015 Google started a joint venture with Ethicon, a Johnson and Johnson company specialized in the development of equipment and systems for medical surgery. The knowledge of Google is of great value to develop new systems: robotics, augmented reality, intelligent sensors. It increases the accuracy of surgery, making it less invasive – with less impact for the patient and obviously less costs in the post-treatment.

Drug design

AI can help set diagnoses and analyze the effect of treatments. It can also help in designing drugs. That’s happening right now on an immense scale. Convince yourself and take a look at the list on the site of BenchSci, a platform for biomedical researchers. They have listed 121 startups (updated January 7, 2019) that use AI for design of drugs, categorized as follows:

·      Aggregate and Synthesize Information

·      Understand Mechanisms of Disease

·      Generate Data and Models

·      Repurpose Existing Drugs

·      Generate Novel Drug Candidates

·      Validate and Optimize Drug Candidates

·      Design Drugs

·      Design Preclinical Experiments

·      Run Preclinical Experiments

·      Design Clinical Trials

·      Recruit for Clinical Trials

·      Optimize Clinical Trials

·      Publish Data

In the category ‘design drugs’ there are two companies that might become very interesting. Virvio uses AI to optimize synthetic biotherapies that are ,,easy to manufacture and outperform known antibodies”. But drawing even more attention is TeselaGen. That company uses AI to make and modify DNA. They call it an operating system for genetic engineering themselves. ,, A single platform for DNA design, construction, experimental data gathering and analysis.” All based on models and algorithms. No lab rats required anymore.

AI designing medicines based on research data and ‘feedback data’ from the patient is the next step. Usually all patients with a certain defect or illness get the same sort of treatment. The only thing that varies in treatment with e.g. antibiotics are the doses and the combinations between various medicines. But what it the actual contents of antibiotics can be ‘finetuned’ to increase the effectiveness of the medicines? Validation and optimization during the treatment, autonomous, based on data that the patient generates him/herself, feeding the data back to an AI system?

In my view we are not far off this step with personification of medicines, using personal algorithms and all the knowledge that AI has access to. As stated before, the key parameter is speed. We got that covered too: just think of the possibilities when we ‘feed’ AI platforms with quantum processors. A real quantum computer is under development, but allow me to address Fujitsu once again. One of the major breakthroughs was the launch of the DAU, the digital annealing unit. It’s not a real quantum computer, but a qbit-emulator. It ‘behaves’ as a quantum processor and therefor is capable of making massive calculations simultaneously. If we increase AI to quantum-speed, theoretically it would be able to mine, analyze and process feedback data ‘real time’, just as neurons in our body do.

Nanomachines

So, we have the data, the algorithms and the compute power to process all this data. Next, we’ve got more thing to fix: the transport mechanism and the distribution. If we can alter treatments – and medicines – at quantum speed, we want the outcomes to be delivered at the right spot at that same speed. Let’s take this final step, or should we better say ‘giant leap’?  What if we get these quantum empowered AI inside the human body? Again: we already got the solution.

The Dutch scientist and Nobel prize winner Ben Feringa and his team at the Groningen University have developed molecular nano-motors, that eventually could be programmed, for instance with assignments to repair defects in human fabrics, deep in our body. Feringa’s nanomachines function more or less like bacteria are moving, however the nanomachines can perform more complex tasks then just carry that one bacterial element. Nanomachines can have memory and sensors, getting close to intelligent behavior.

If we enrich these motors with quantum powered AI, it could deliver medicines exactly there where it is needed, get immediate feedback, analyze the information and adjust the treatment by transporting data back to a swallowed or injected AI chip in our body, connected to a quantum AI-computer outside our body with access to tons of data – in a fragment of a second, anytime.

Conclusive Hug

Off course, the ethical question does become relevant. Once again: we as technologists can invent anything we want and promote the benefits of that particular invention, if the masses do not accept the change the invention or innovation brings, it won’t change anything at all.  Resistance to change is fed by fear.

Maybe there’s one question that’s even more important then: should we be afraid of all this technology coming our way? Let me answer that one with a quote – embedded in the book ‘Augmented Health’ by Lucien Engelen –  from Chris McCarthy, VP strategy and design at HopeLab, a nonprofit organization working on innovative solutions to improve the health and quality of life of young people with chronic illness. ,,And yet, with all the digitization happening around us, there is a yearning for authentic interactions and a good old-fashioned, warm hug when things are not going well.”

No, we should not be afraid of technology. We should embrace it, but with good means. What AI really can and should do is making basic healthcare accessible to everyone. There’s a big task. According to WHO over 400 million people on this planet do not have access to basic healthcare. That’s the biggest challenge that we can solve, with help of AI, quantum computing and nanotechnology. That’s where the focus should be.