Classical cardiac risk prediction is based on Bayesian theorem and linear multivariate modeling. If a 42 year old woman comes to me with stable chest pain, the first thing I would do is to classify whether the chest pain is non-anginal, atypical, or typical. Using the age, gender, and character of the pain, we would come up with the pre-test probability of the chest pains being cardiac, using multivariate model (Diamond and Forester NEJM 1979). Then we may choose to perform stress test, and we will obtain the post-test probability, using Bayesian calculator. A risk stratification such as Duke Treadmill Score — a linear regression model based on exercise time, chest pain on exercise, and ST depression in mm– would be calculated. The decision of whether to proceed to cardiac catheterization would be mainly based on a combination of these results, according to ACC/AHA stable CAD guideline.
There are two major practical problems with this approach. First of all, the pre-test probability of chest pain in 1979 is far higher than what it is now, and these estimates may not be the same in different ethnic groups. Secondly, even after getting classified as low probability or low risk Duke score, the probability of CAD is still not zero. Thirdly, the model is based on linear regression modeling, when the real world is more complex and is based on non-linearity.
Recently I heard of this tragedy in my community: one young patient had a normal stress test and was told by his cardiologist to have a clean bill of health; while he was still in the waiting area getting ready to go home, he had a cardiac arrest and he could not be resuscitated. And then I had two young woman in the 40s who had normal stress test, Duke Treadmill score low risk, who ended up with 95% proximal LAD stenosis on cardiac cath.
So even if you have a low post-test probability low Duke treadmill score, the probability of cardiac event is still not zero. And you can say: well, that is uncommon. But to the family of the patient who died, your Bayesian reasoning is not going to console the grief of the family left behind.
I pondered this for a long time. I came to the conclusion: in the real world, patients are more complex than is suggested by linear statistical model. So I started searching for tools from the new field of complexity theory.
Cardiology uses tools that was used by financial institutions before 2007 crisis. Before the crisis, banks assume that financial variables (such as default rate) follow the normal curve. Therefore, they only have to plan for events that are with 2 standard deviation of the mean. In 2007, an extreme “black swan” event occurred, and default rate of the mortgage loan sky-rocketed. That resulted in the death of Lehman Brothers.
The explanation of quantitative finance thought leader, Nassim Taleb, is helpful: (1) people have been fooled by predictive model, which has great correlation to the past but has poor predictive abilities; (2) Black swan can happen, and there is a lot of “fat-tails” (i.e. rare events happen more than you think); (3) most risk model fails to account for the possibility of complete ruin (e.g. death).
My view is that we need to get rid of the linear methodology and adopt a non-linear risk assessment approach, taking into account of the complexity of the patient.
One way computer has tried to solve this kind of problems in the complexity theory is to perform machine learning using genetic algorithm. With trial and errors, the machine will evolve the best algorithm given the circumstance. The key point is that the algorithm evolves with additional feedback to optimize the survival of the complexity network. I also looked in the field of game theory to provide additional insight to the solution of this problem.
(1) As it turns out, the best neural network available is that of a brain whose survival depends on improving the outcome of its client patients. If the clinician brain is open to feedback of clinical outcome and the clinician’s survival depends on these outcome, our human brain will be better than any neural network derived by genetic algorithm. We can use decision aids such as linear clinical model, but our brain can integrate many different clinical inputs ignored by simple models. (The utility of clinical judgement has recently been demonstrated again in the literature!) But our brain has several known limitation such as anchoring bias, and knowing these cognitive limitations is important. Our clinical intuition is based on summation of previous clinical experience as applied to the current clinical scenario, and I search for the clinical decision that yields the least stress. My brain’s survival depends on me making the right decision. Thinking, Fast and Slow by leading cognitive scientist Daniel Kahneman (published 2011) presents important new cognitive science insight as to how to optimize this process in daily thinking. The fast clinical reaction has to be coolly analyzed.
(2) encountering patient is like a chess game:
it consists of multiple steps. Instead of treating clinical encounter as a one-time encounter, we need to plan our encounter with the patient like a chess game. Using our local experience, we need to make the clinical decision that makes the most sense with the resource available. Clinical experience is most important in this regards. Based on the experience we need to come up with potential clinical scenario and their probabilities as well as how to address them. To dumb down cardiology to a few simple risk score ignores the full complexity of medicine. We need to build in contingency planning: what if we made an error in our initial decision. The finding in game theory is important in this regards.