While We Are Awake

In “While We Were Sleeping” David Hemenway describes the impact of innovations in safety and injury prevention that have saved millions of lives by preventing people from wandering into death’s neighborhood in the first place rather than clawing them back from death’s door at the last gasp. In effect, these technologies and policies help us while we are sleeping and Hemenway argues that as such we take them for granted. After all, it is inherently difficult to appreciate avoiding something that will never happen.

This all came into sharp focus a few weeks ago when my uncle died of a heart attack. I remember when I was young, I used to glue his cigarettes together to stop him smoking. It probably foretold a career in preventive medicine but it must also have been incredibly annoying. My uncle, over his lifetime tried a number of ventures and unsurprisingly, not all of them worked. I remember a couple of days after he died as my brother and I filtered his remaining possessions what became painfully clear was that he had been making plans right up until the end: extensive plans, vividly imagined and meticulously documented. In the end, as we all do, he just ran out of time.

Sithamparanathan Sivagnanam

One morning I went shopping; on that day, like any, the mall hummed with the back and forth of shoppers against the piped atmosphere and inter-changeable storefronts of any town, anywhere. I headed purposefully to the pharmacy to buy some Aspirin for I had decided that I should do something to prevent what seemed otherwise inevitable. In public health terms this is called primary prevention and it is arguably the holy grail: preventing a disease before it manifests. This proactive approach is in sharp contrast to the more reactive (and costly in all senses) norm: preventing a recurrence or decompensation of a disease that is already manifest (known as secondary prevention) or treating a disease once it decompensates (known as tertiary prevention). In this particular case, there is evidence that, for those over 50, a humble little Aspirin taken every day has a god-like ability to prevent heart attacks and other potentially fatal diseases including colorectal cancer. I am not over 50 however I was born in Sri Lanka to Sri Lankan parents and as such my risk of heart disease is about double (estimates vary) that of someone of northern european descent. As such interventions that are otherwise indicated for people older or with established heart disease than me probably make sense for me.

Reassuringly, using the line of argument above — if I take this little pill and it works, there is a heart attack that a future me should be having that he will not have. This hopefully means that there are some more years of life that I will have without knowing where they came from. Maybe in those extra years I will get to see our girls shape their world a little more, have babies of their own or win an olympic gold medal (admittedly with these genes — unlikely.)

So why is the level of promotion around this “god pill” or other things that can help us “while we are sleeping” not at least as zealous as that for all the other things at the mall on any given day? Or more broadly why is it so difficult to find out what long term health conditions I am at risk of and what I can do to prevent them? Is there a hidden agenda to avoid bolting costly years of “coffin dodging” to the end of life? Or is it just that is practically impossible to know what health is or specifically what one needs to do on any given day to make the most of one’s health? Is it just that it is difficult to make money from preventing health conditions and it is difficult for us as human beings to act differently now to prevent potential diseases in the future?

Before I start on what is going to be an extended answer, I suspect some qualification is needed. For the last few years i have spent a lot of time around engineers and this has profoundly changed how I think about medicine. The great thing about medicine, probably the greatest thing about medicine, is that it is such a broad church. I now believe that in many ways medicine is an engineering discipline, for it shares some of it’s core themes with engineering: reconstituting a complex natural system as discrete processes, building a descriptive model of said processes, identifying the constraints and using the tools of science to build solutions. Typically in order to make natural systems tractable as discrete processes reality needs to be simplified — creating an abstraction — a more simple version that is easier to model thereby fashioning a set of problems that are easier to solve. This is one such exercise. I appreciate, the clinician’s instinct is necessarily to think about each individual and manage edge cases. With this in mind I apologize in advance for the significant edge cases that are totally ignored in service of narrative fidelity.

So, qualifications aside, how could things be different? To answer this question I would like to first take a detour. I was recently talking to my good friend Gopal Ramachandran (who as well as his numerous personal qualities and surprising penchant for KFC has a PhD in mathematics and a medical degree — in my experience a unique combination) about chess grandmaster Garry Kasparov’s new book “Deep Thinking.”


The book focuses on the ultimate man vs machine match up of one of the greatest ever chess world champions and Deep Blue — chess playing software written by IBM running on a custom made super-computer. During his illustrious career, Kasparov had apparently played around 2400 serious games of chess against the world’s best chess players and lost just 170. In fact of those 170 defeats one had been to Deep Blue a year previously though ultimately he won the match on aggregate. But in 1997 as Gary sat down on the other side of those 64 squares from what would become his silicon nemesis, things were about to change. Over 6 games, Deep Blue won 3–2 with one draw and a previously unimaginable scenario was realised — a computer system had beaten the reigning world champion.

At this moment, if we could reduce the universe to just the world of chess, we all became gorillas. What I mean is that just as human beings developed the cerebral cortex and inevitably gained dominion over our primate cousins, the wheels began turning (or turning faster?) on mankind’s eventual subjugation. Of course the universe is not just chess and the super-intelligent AI overlord might never be created but what is of greater near term interest is what might happen in the next few years while we are awake.

Immediately, following this fateful match and for a number of years afterwards Kasparov threatened recourse to the courts for accusations of cheating (accusations he later retracted) then through a process of study reached acceptance of the dominance of machines in chess and other areas of cognition and ratification of the belief that human computer hybrids — humans and computers working together will be able to create a much better world than either in isolation.

In this time, human beings have continued to play chess for fun and computers have helped develop the next generation of players. In fact, the burgeoning field of Centaur chess — a human player and a computer chess program playing as a team against other such pairs was championed by none other than grandmaster Gary Kasparov.

Kasparov has become an advocate of machine or artificial intelligence in general arguing appropriately that humans are typically bad at processing large quantities of data and we tend to find patterns where none exist. He believes, like I do, that the potential of computers is that they can help us be more objective and amplify our intelligence through augmenting it with computational intelligence. As Steve Jobs famously said “,computers are like bicycles for our minds.”

Mankind on a bike is much more energy efficient than jets, horses or errrm… salmon

So what has any of this got to do with preventive medicine? To answer this question another detour is needed into the murky waters of “value based” healthcare.

Based on the research of Professor Michael Porter, “Value-Based Health Care Delivery” is “a framework for restructuring health care systems around the globe with the overarching goal of value for patients — not access, cost containment, convenience, or customer service.” Announced to much fanfare and championed by the great and the good and other less significant characters (including yours truly), it is debatable how many health systems around the world have successfully followed the call and are in fact oriented around value for the patient. In fact “value” has been appropriated as a buzzword for any generic initiative to share risk between the providers of care (individual clinicians, clinics and hospitals) and the institutional payors for care (here in the US private insurers and in most other places governments) through any mechanism. However, I think the original simple value framework and the manufacturing thinking of lean are together useful in conceptualising health systems as interlocking processes supposedly designed to deliver value to patients and analyse said processes mechanistically.

Within this framework, the value for preventive healthcare in the near term is the reduction in probability of a disease manifesting or if it is already manifest, a complication of a disease occurring. For most conditions, including most existentially the condition of life itself, definitive cure is not possible and as such the purpose of the health system is to prevent something worse than the current state occurring. As such, even for something as seemingly nebulous as prevention, a production system can be conceived — a series of actions that reduce the probability of something worse happening. In such a system each individual person exists in a continuous stream of health states. Each state is determined as a point in a multidimensional space where each dimension pertains to different information regarding a patient at a point in time (physiological state, pathological state and emergent states such as medication compliance, social isolation or self efficacy). Based on a patient’s health status, it is possible to associate interventions that aim to reduce the probability of the manifestation of or a deterioration of a particular disease(at a high level this is what clinical guidelines are) with a pair of states. As such a model is created for the health of an individual patient over any period of time with discrete health states as nodes and preventive medicine interventions as edges. The optimal interventions can be mapped entirely deterministically on average across a whole population. This is essentially what clinical guidelines do now and it is what Deep Blue did with chess.

Such a system can be calibrated according to patient state but also policy imperatives such as optimising flow through the preventive medicine system to ensure optimal use of resources. As such, rather than clinics or hospitals being the center of the patient journey, it could just become one stop amongst many. Similar to the strategies of more modern AI based game systems (e.g. AlphaGo or later incarnations of Deep Blue), moves that seem surprising to the human observer could be optimal in aggregate. In addition, the promise exists for the local creation of evidence-based medicine from the sum of individual chronic disease management experiments undertaken. This would of course accelerate the translation of large-scale clinical data from the byzantine approach of today into dynamically generated and tested clinical insights available to all.

This is a general approach, a class so to speak, and over the last five years at Wellframe I have been working on a specific instance of such a system. Specifically we have made a condition-agnostic mobile health platform that delivers dynamically-generated daily multimedia health checklists to patients on a mobile device. The mobile application is married to a clinician dashboard where patients are filtered and prioritised according to common bundles of clinical work and clinical priority. Furthermore, clinician intervention is itself protocolised with a series of tasks to do on any given day dependent on a patient’s ‘health state’. Patient data and integrated decision support tools enable communication between the clinician and the patient through a secure multimedia messaging channel. In addition, a clinician can add, subtract, or end a care program; call or visit a patient; or refer them to resources or appointments elsewhere in the healthcare system. All of these actions can and will be automated, but we started by designing the system to support actions taken at a clinician’s discretion, rather than to generate autonomous actions. Like Kasparov, we are interested in creating Centaurs.

At the heart of the Wellframe model is an identification of the common logic structure in clinical protocols and an approach to interpolating clinical protocols to resolve for the tasks a patient has to do in between visits. As a recovering clinician it has been edifying that there is a pretty consistent logic structure to preventive medicine protocols (of course intuitively there has to be or it would be practically impossible to teach or learn medicine. )

Through defining a common logic structure it is possible, as with all programs, to break clinical protocols into atomic units and to administer them in a rules-based manner in clinical practice (according to preshared deterministic pathways co-created with the clinical delivery partner and based on practice standards and clinical research findings). I believe this is ground zero.

We have been working with health plans in the US to help their most complex patients with this approach. The beauty of software, the realisation that made me give up clinical practice and work on both the creation of software and in turn software companies for the rest of my days is first their extraordinary scalability and the way in which through creation of a data asset the more patients one serves the more value can be delivered to the next patient without commensurate increase in costs. This means that the greater the scale at which care is provided, if it is provided digitally, the economics of preventive care delivery tend towards the asymptote of product companies such as Apple rather than that of hospitals. Two and two really can and will make five.

I believe that us apes can build software in healthcare that not just mimics the feats of Deep Blue in chess but actually goes much further. It is important to remember how much has changed since Deep Blue. Deep Blue focused on a problem that was even then considered computationally tractable if very hard. The way chess works is that if one were to look at a chess board at any one time it would be possible to describe all possible potential future moves and as such pick the optimal one. Of course no human player does this, instead as humans we rely on the extraordinary parallel processing and pattern recognition capacities of the human brain to pick the move that “feels right”. If a computer is able to process all possible options and apply a fairly rudimentary decision filter it would not need to learn to beat even the best humans. And so it proved.

Since that fateful day in ‘97, advances in sensing, processing and algorithms have meant that established techniques that had been forgotten have shown immense promise in problems that are not computationally tractable without learning — such as the game Go, massively more complex than chess where London and now Google’s Deep Mind defeated the world champion, Lee Sedol in 2016. Such learning algorithms and the hardware on which they were run are incredibly becoming available to anyone with an internet connection. If you have not I would highly recommend watching this week’s Google IO for more…

In such a world where preventive care can be delivered digitally and then improved by computational learning on a data substrate derived passively from the delivery of care itself, I believe previously unimaginable advances in preventive medicine will be realized. In fact, I don’t think that the impediments are in hardware or software but in the organisation of human systems and in incentive networks.

Realistically preventive medicine will remain about working with patients in intervals of time around teachable moments rather than continuously over a lifetime. It will be possible to predict those teachable moments and then target specific patients at the right time, not just with the right interventions but through the right outreach channel and with the right message that appeals just to them. Furthermore at one of these teachable moments, the likely cost, benefit and propensity of a patient to engage over a treatment course can be predicted and this calculation used to organise the human and physical resources of the health system to where they can be most effective and then subsequently to make a seemingly infinite chain of adjustments to ensure the optimal allocation as patient’s needs change. In fact, much more preventive care will be social i.e peer to peer as once a certain data threshold is met it is possible to match people to other people like them, either within their community or outside who have similar goals and who are experiencing similar challenges who can work with each other — possibly obviating the need for clinicians altogether.

In summary I believe it is possible now for every person to know what they need to do and look out for to make the most of the aspects of health matter to them and for those responsible for the preventive health system to have real time insight into who needs what level of support and when as well as the ability to deliver the right thing to the right person at the right time at national scale.

In such a system will human beings become redundant? Will we go the way of the gorillas — beholden ultimately to our better adapted cousins? In the very long run, the answer is “possibly, yes,” but I am confident that we will first enjoy an unprecedented renaissance of discovery and productivity.

Reducing the universe to healthcare again, Marvin Minsky wrote in Society of Mind “in general, we’re least aware of what our minds do best,” and added “we’re more aware of simple processes that don’t work well than of complex ones that work flawlessly.” I have abstracted preventive healthcare into a stream of information exchanges which of course is only part of the story. A sense of companionship and of connectedness give both direction and relevance to all actions in health (as in life) and for the foreseeable future there will remain a need for caring folks working to help others in need. It is just that they will be able to focus more of their time on caring rather than algorithmic information dissemination, collection and inference that computers can do much better.

Secondly, in what has become known as “Kasparov’s Principle” it is stated that a human with a weak machine and a strong process will necessarily outperform a human with a strong machine and a weak processes. Having, seen where software can and cannot be successfully applied in complex human systems I firmly believe that in the short term it will be in the design of processes for humans and computers to work together that the gains of artificial intelligence are going to be realised and only humans can do that.

Grandparents :)