This Monday evening our next podcast becomes available. Dr. Gustavo Heudebert and I discuss another article about risk prediction. This topic has become a recurring theme on the Annals On Call podcast.
Why is risk prediction so important? In 2019 we make many decisions about prevention and testing based on risk prediction. In addition we also estimate harms and benefits. All these predictive models have advantages and flaws. In making a decision for statin use (another upcoming episode), we have to estimate the risk of cardiovascular events, how much taking a statin will decrease that risk, and the probability and type of side effects from taking a statin.
These predictions all come from mathematical modelling. Mathematical modelling is fraught with many hazards. One can apply the same modelling techniques to different databases and develop significantly different models. The episode with Dr. Rod Hayward – Improving Estimation of Cardiovascular Risk – gave a great example of how including different databases and even different time periods changes our risk prediction for cardiovascular events. In that episode, he also explained the problem of temporal trends. Cardiovascular risk has decreased over the past 30+ years. Thus, our predictions should change.
He also pointed out that we never have data on all the risk factors that one would want in a model. The cardiovascular risk prediction models do not take into consideration family history, renal disease or fitness – yet we know that all of these factors modify CV risk.
The same problems exist for benefit and harm prediction. The promise of big data assumes that we have complete data. Yet in medicine we never have complete data.
These predictive models can help our clinical decision making. We should always use them carefully and thoughtfully. Understanding that the numbers are estimates with confidence intervals can help us. They can help us start a conversation with our patients about risks and benefits. The numbers are not magic. We cannot enter them into a computer program to make clinical decisions. Resorting to algorithmic decision making will quickly become dangerous.
Many decisions occur in a “grey zone”. We should take into consideration that patient, his or her concerns, and the totality of their medical situation. Such decisions are complex and understanding these issues makes one wonder about many guidelines and performance measures. We should always remember HL Mencken’s quote, “there is always a well-known solution to every human problem—neat, plausible, and wrong”.
