A prediction is only useful if someone can act on it
An accurate model can still help nobody. What I took from Stanford’s AI in Healthcare courses is that every output needs an action, a person to take it, and a threshold that respects how much that person can do.
Stanford’s AI in Healthcare specialization is five courses, which came to 54.5 hours of Stanford Medicine continuing education. It runs from how healthcare is organised, through clinical data and the fundamentals of machine learning, to how AI applications in healthcare are evaluated. The evaluation course is the one I keep coming back to.
A model’s accuracy tells you how well it predicts. It doesn’t tell you whether anyone ends up better off. For that, the prediction has to be paired with an action: someone has to receive the output while there is still time to do something, and what they do has to help more than it costs.
Here is the kind of example that made it concrete for me. Picture a model that flags patients likely to miss an appointment. Its accuracy could be excellent. Whether it helps depends entirely on what happens after the flag. If the clinic has someone whose job includes calling flagged patients, and those calls change who turns up, the model is useful. If nobody has time to call, the flag is a coloured dot on a screen. The same model can be worth a lot in one clinic and nothing in the next.
The part I hadn’t seen before is what that does to the threshold. A model produces a score, and someone decides how high the score must be before a flag appears. Set it low and you catch more cases, and you also bury the person who has to act on them. Set it high and the list is manageable, and you miss people. So the threshold is partly a staffing decision. It sets the size of someone’s workload every morning, and it’s easy for it to be chosen by whoever trained the model, on statistical grounds alone.
That makes it design work. The list of flags, the order it’s in, what each row offers as the next step, and what happens to a flag nobody touches are all screens. The model’s quality reaches the patient only through them.
Looking back at ClinicOS, I now see the rule that anything urgent or clinical pages a person straight away as one of these pairs. The agent’s output is worth producing only because a named person is there to act on it. I didn’t have this language when that rule was made, and the course didn’t shape it. It does give me a clearer way to explain why the rule matters.
On the next health product that uses a model, these are the questions I’d ask before designing anything around it.
For each output, what is the action, and who takes it? If nobody can name the person, the feature isn’t ready.
How many of these can that person handle in a day, and does the threshold respect that number? I’d want the threshold discussed with the people doing the work, not only with the people who built the model.
What happens to an output nobody acts on? It should expire, escalate or be visibly dismissed. A list that only grows teaches people to ignore it.
How will we know the action helped, separately from knowing the model was right? That has to be decided before launch, because afterwards every number looks like evidence.
Was the model tested on patients and data like ours? The clinical data course was a good reminder that data reflects how and where it was collected, and a model can carry those habits into a clinic that works differently.
I can’t evaluate a model’s statistics myself, and these courses didn’t make me able to. They gave me better questions to bring to the people who can, and a clear reason why the answers belong in the design.
AI in Healthcare Specialization, Stanford University, completed 3 October 2026. Verify