One of the big shifts in healthcare caused by so-called disruptive innovations in healthcare powered by AI and big data, is a shift from diagnostic and curative healthcare to predictive and preventive healthcare. While preventive healthcare is almost exclusively cloaked in positive attributes, we need to maintain semantic clarity about what it can and cannot deliver, so that patients are not misguided about its benefits and limitations and can make well-informed decisions regarding their healthcare. This paper zooms in on potential language confusion in this context when words such as “predict/prediction/predictive” and “(at) risk” are used in the context of AI algorithms and big data research, such as GWAS studies in genomics.

Although risks and probabilities were always a part of healthcare and although there may always have been some level of confusion about probabilities being interpreted as facts or truths, in curative healthcare these are side-effects of the goal of diagnosing and treating existing illnesses. In preventive healthcare, however, they are located centre stage, and following up on them is supposed to prevent illnesses from breaking the barrier between a possibility and a reality. However, we should be weary of forgetting that predictions and risks are (a) fictional constructs, expressing the state of our knowledge about realities, not expressing the realities themselves and are (b) dependent on arbitrary thresholds that are not value neutral. Thus, when a claim is made that an AI tool can predict whether sepsis will occur or whether someone will become depressed, we should not regard its outputs as factual claims, merely as a probability calculation that is (hopefully) better than a random guess and useful on a population level, but does not represent “the truth” about individual cases in any way. The term “variance explained” in genomics illustrates how uncareful wording also permeates very specialised scientific discourse, glossing over important limitations, in this case that the variance is not explained or understood at all, but rather known to be correlated with certain variations in the genome. 

In conclusion, we need more clarity and more humility in the claims we make in the context of predictive and preventive healthcare.