Causality

Whether one thing actually causes another, as opposed to the two simply moving together in the data. Almost every AI model in healthcare today learns correlation: it finds patterns that predict an outcome without knowing why. That is often enough for triage or risk stratification, and it is not enough for a decision about what to do next, because acting on a correlation can change the very thing that made it predictive. The textbook case is the pneumonia model that learned asthma patients had lower mortality, when the real reason was that asthma patients were sent straight to intensive care. Causal questions need a study design that can answer them, not a bigger training set. When a vendor says its tool identifies the drivers of an outcome, ask whether that claim rests on a correlation found in retrospective data or on evidence that changing the input changes the result.

A readmission model ranks "number of prior admissions" as its top driver. A committee reads that as a lever and funds a program around it. The feature was a marker of how sick the patient already was, and the program moves nothing. The same model can be excellent at ranking who to call this week and useless for deciding what to change. Ask which of those two jobs a tool was built for before you plan an intervention around its output.

Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.