MDCalc Statement on Race in Medical Calculators and Risk Estimates

Our Complicated Concerns over “Race”

• It is very obvious that "race" is a terrible surrogate for what we in medicine often care about: one's unique genetic makeup, and how that code impacts diseases we may or may not be at risk for, and even what medicines may or may not treat our bodies most effectively.

• It is difficult to determine why race ends up being a risk factor. Is it due to socioeconomic risk? A surrogate for geography? Access to healthcare? True genetic difference? Some may try to argue that the reason is inconsequential, so long as the correlation exists. But when we ask the clinician to apply these blunt instrument tools to the unique individual in front of them, it becomes crucially important that they understand the reasoning: how do my patient's race, geography, and socioeconomic background interact?

• This is the same core issue behind one of the main pitfalls of AI and Big Data, which many developers seem to ignore: when applying the blunt tools of evidence-based medicine to unique patients, correlation in a vacuum is not enough.

• We know that these broad, vague categories of "race" are a poor substitute for the diversity and complexity of human beings around the world. They require us to pigeonhole patients into particular boxes, when life exists so much more on a complex scale. The variability of individuals — and genetics — within a race is enormous.

• We know that the United States, for example, is very racially diverse, yet many studies only use "Black" or "Non-black" as their racial categories. Potential reasons for this over-simplification may be ignorance, racism, study cost, or even simple statistical mathematics of power and effect size; regardless, these studies are hard for physicians to interpret if a patient does not fit into one of those boxes.

• Similarly, "White/Caucasian/European ancestry" is often the "default" option in these studies, which makes sense given that the majority of people in the United States identify as Non-Hispanic white. However this "default" raises questions about normalcy as well.

• Importantly, it is hard to know how to handle these situations. Is it wrong, today, using the data currently available, to use race in these models? If we remove the race input from any one of the eGFR calculators, are we damning some patients to a "diseased" state by lowering their eGFR and recommending dialysis earlier than necessary, or properly diagnosing them so they are treated the same? (NephJC has an outstanding analysis on this topic.) Similarly with breast cancer's Gail Model: are there different incidences in breast cancer in women by race? How can we even know the "Truth"?

• Finally, the problem we are highlighting is not unique to race as an input, albeit more emotionally charged. Many inputs discovered through statistical approaches can have unclear causes, and the cause is critical to understand. Do patients in different countries express “shortness of breath” differently? Or, a classic often confusing input from the TIMI NSTEMI Score: taking aspirin turns out to be a risk factor for heart attack — because so many patients who take aspirin daily already have risk, or because aspirin "should" be preventing heart attack itself.