This is a long uninformative title. I will try to make sense of it, let me know if I succeed.
Executive summary: We need more math (mathematical models of disease) and good thinking math in medicine, not “after the fact math” to deal with the mess we have created.
Early in the nineteenth century Laplace’s Demon was the first time someone articulated determinism. If someone knew the exact position and composition of all beings in the universe for a single moment and submitted these data to analysis, it would be possible to predict the future without fail, as “nothing becomes uncertain”.
Tversky and Kahneman studied the psychological underpinnings of human decision making. With their “prospect theory,” they argued that people are not as calculating as economic models assume. Instead, they said, people repeatedly make errors in judgment that can be predicted and categorized. Daniel Kahneman used a dual-system theoretical framework to explain why our judgments and decisions often do not conform to formal notions of rationality. System 1 consists of thinking processes that are intuitive, automatic, experience-based, and relatively unconscious. System 2 is more reflective, controlled, deliberative, and analytical. Judgments influenced by System 1 are rooted in impressions arising from mental content that is easily accessible. System 2, on the other hand, monitors or provides a check on mental operations and overt behavior—often unsuccessfully.
Ok, I am lost here… what is your point? My point is that even human behavior, an extremely complex system with many variables at play, can be studied and structured to predict judgements and choices.
Next on my list comes a movie… ok a movie based on a book “Moneyball” (2011). The central premise of Moneyball is that wisdom of baseball insiders is subjective and not accurate. Statistics used to evaluate players had an inherently old view of the game. Before new metrics were introduced to baseball, teams were dependent on the skills of their scouts to evaluate players. Rigorous statistical analysis demonstrated that on base percentage and slugging percentage are better indicators of offensive success, and these qualities were cheaper on the market than more historically valued qualities.
By re-evaluating the strategies that produce wins on the field, the 2002 Athletics were competitive with teams such as the NY Yankees. This approach brought the A’s to the playoffs in 2002 and 2003. Another example of how the effort to understand a system, and the ability to model it can produce better forecasts.
Let us go to another completely different field, commodities trading. Ray Dalio, the extremely successful hedge fund manager, started in the commodities business back in the 70s. He developed a model to link cattle, chickens, and hogs and their food (grain). He could project how much meat would come to the market and how much corn and soy meal would be consumed. Seemingly, by looking at how much acreage of corn was planted and applying the rain forecast we could guess how much grain would be available and at what time. This is a very smart and simple way to understand the production machine, with all its levers. Being able to understand how much output there will be with these specific set of inputs (rain, time, acreage planted) gives you a competitive advantage over your fellow traders. If you understand the machine, your model will be able to forecast the output with a specific set of inputs. Is this Laplace’s demon again?
How does it all tie together with the need for more math, reasonable and honest math, in medicine? The current paradigm is basically based on rolling a dice and praying (yes, scientists also pray… and they pray a lot before unblinding a trial) the only interference with the system is the drug under study. The key is to randomize patients to something or nothing, assuming “ceteris paribus” that if we see a difference in the outcome it is due to the drug under study. Though it is not a horrible logic it is fraught with faults. We assume the disease is homogeneous across all patients, but we are not sure so we run complex statistical models where we adjust for a myriad of variables…just to make sure. Even if we are successful we frequently find safety issues or interactions when the drug is used in real patients with more complex comorbidities. Where is the root cause of all these? Coming back to Donald Rumsfeld “… there are also unknown unknowns. There are things we do not know we don’t know.” Should we not try to diminish those? Are not these “unknown unknowns” the most dangerous ones?
My proposal is to start with a humble approach. We know some aspects of the disease but we ignore the inner workings of it. We invest countless hours and talent in describing specific aspects, steps of the disease but we never dare to put together an operational model of the disease. Our animal models focus on specific elements we think are relevant, but they are always a remote approximation to what happens in a real patient. That is why there are so many results in animal models that do not match what happens in real patients. However, we continue to hammer such models in our decision making process when we select a drug to pursue its clinical development.
We need to start building operational models of the disease at hand before we start any plan to advance cures, only understanding the relevant elements of such models will we be able to design better cures. By building operational models of the disease we will discover aspects of the disease that have not been addressed or are assumed to be true without any proof. Even though many articles and medical books contain beautiful graphic displays of the disease physiopathology, none of those models of the disease include any information on the relevant (“sine qua non”) pathways of the disease and epiphenomena of such pathological process. None of such models consider temporal sequences of events, its relevance or the strength of the associations in the different steps of the pathological processes. None of such models include any non-linearity in the description of such events, and we know by now that biological systems tend to be non-linear and only in severe pathological states turn into failed but linear predictable models.
Once we have a disease model, we can check if it can forecast results (outcomes) of certain treatments. If not, we should revisit the model and use such data to improve it. This will be a positive feedback process. If we have a sophisticated model we can evaluate outcomes by manipulating certain elements of it. That way we might be able to identify future treatments or design combination treatments.
It is worth the effort!
Reading list
The undoing project by Michael Lewis 2017
Principles by Ray Dalio 2017
Thinking, fast and slow by D Kahneman 2011
Money ball , sorry I just watched the movie
The logic of scientific discovery by Karl Popper 1959
Known and Unknown: A memoir by Donald Rumsfeld 2011
El Quijote, always a good read