Say “AI reads blood” and people picture a single machine: perhaps a robot doctor, perhaps a spreadsheet with a red flag on it. In practice, two different technologies are hiding under that phrase. One looks at blood. The other listens to its numbers change. Confusing them is where a lot of the hype—and a lot of the skepticism—begins.
Path one · Blood as a picture
The evidence is visible.
The first approach treats blood as an image. A drop is spread across a slide, a microscope photographs the field, and a computer-vision model examines cells or regions for a defined target: an abnormal morphology, an unusual cell, a parasite. Researchers have built object-detection systems that locate infected red cells in thin malaria smears.3
Its appeal is unusually concrete. A model can mark the exact region that triggered its output, letting a trained reviewer inspect the same evidence. That does not make the model perfectly explainable, and a box is not a diagnosis. But it gives the human something specific to accept, reject, or investigate.
It can also surface information that never appears as a row in a routine chemistry panel.Plasmodium falciparum, for example, can appear as delicate ring forms inside red blood cells on a thin smear.2That evidence has a shape. If the system receives only numerical lab results, the shape is not there for it to find.
Path two · Blood as numbers
The evidence is a trajectory.
The second approach never sees a picture. It works from structured measurements: a complete blood count, a metabolic panel, a biomarker drawn more than once. Far more patients generate tables of routine labs than digitized blood-smear images, so this is often the data already at hand.
Here, the important signal may not be one result but the direction across several results. In the synthetic example below, the first number has little meaning without its assay, units, and context. The sequence is the point: 12, then 48, then 190. A forecasting model can track many such movements together, looking for a pattern that precedes an outcome.
A single value is a location. Repeated values reveal a direction.
That is the real split. The image model identifies. The numerical model forecasts. A single-image system does not automatically know the patient’s direction over time, while a numerical system cannot inspect a form it was never shown. Neither is a bigger or smaller version of the other. They are built around different evidence.
The catch
A model can learn the hospital.
The numbers side carries a quiet risk: the model may learn the clinical workflow along with the biology. A lactate ordered at 3 a.m. is not merely a result. The decision and timing may reveal that a clinician was already worried. In a large retrospective study, whether and when lab tests were ordered sometimes predicted survival more strongly than the values themselves.4
A model can use those care patterns because they are genuinely predictive in its training data. Trouble arrives when the system moves to a hospital with different staffing, protocols, or ordering habits. What looked like knowledge of disease may partly have been knowledge of one institution’s routine.
That is not a reason to discard forecasting models. It is a reason to test what they learned, validate them across settings, and be precise about what their inputs contain.
Beyond the split
The fuller picture uses all three.
An image shows what is on the slide. Measurements show change. The clinical note may explain why the test was ordered, what the examination found, and what question the care team was asking. Medical-AI research is increasingly exploring models that combine images with clinical context rather than forcing one source to stand in for the rest.5
So when somebody says “AI reads blood,” ask which blood they mean: the picture, the numbers, or the patient story that gives both of them meaning.