A predictive model answers the question: given what I can observe, what outcome is most likely? A causal model answers a different question: if I intervene in a specific way, what will happen? Both are valuable. They are not interchangeable. Health AI often conflates the two, which leads to misuse, overclaiming, and decisions made on grounds the model cannot actually support.
Predictive accuracy on historical data does not imply causal understanding. A model can achieve high predictive performance by learning associations that reflect confounding, selection bias, or measurement artifacts. When that model is applied to guide interventions, the associations break down because intervening changes the distribution that the model learned from.
What causal framing requires
Moving toward causal infrastructure means being explicit about what assumptions a model makes, what population it was derived from, and what interventional claims, if any, it supports. That requires careful study design, transparent documentation of causal assumptions, and honest reporting of what the evidence does and does not support. It is more demanding than predictive benchmarking, and that additional demand is appropriate given the stakes.
At Cytognosis, we distinguish between models designed for observation (understanding patterns in data), models designed for prediction (anticipating future states), and models designed to inform intervention (reasoning about what would change under different conditions). Each carries different validation requirements and different warnings about appropriate use.
Staying honest about the gap
Much of our current work is in the observational and predictive categories. Where we reason about mechanisms, we are explicit that we are generating hypotheses for experimental validation, not deriving actionable interventional conclusions. This page is part of our open notebook and will be updated as our methods and evidence evolve.