The number that sells the model
Most clinical AI is bought on a single figure printed on a slide: accuracy. It sounds objective. It is usually the least useful number a vendor can give you.
Accuracy simply counts how often the model was right. If only 5 in 100 patients on a unit develop sepsis, a model that predicts "no sepsis" for everyone is 95% accurate and catches nobody. The brochure is technically true and clinically worthless.
Where the score came from matters more than the score
A model's results depend on the data it was tested on. Many vendors test on data that resembles what the model was trained on, sometimes from the same hospital and the same year. Performance drops, often sharply, when the model meets a different patient population, different devices, or different charting habits.
Before you trust any figure, ask three questions. Which patients was it measured on? Did the model ever see those patients during training? Can someone outside the vendor reproduce the result?
A score you cannot reproduce is a marketing claim, not evidence.
The numbers worth asking for
Sensitivity tells you how many truly sick patients the model catches. Precision tells you how many of its alarms are real, which is the number your nurses will feel first. AUC, short for area under the curve, summarizes how well the model separates sick from healthy patients across every alert threshold; 0.5 is a coin flip and 1.0 is perfect.
No single metric is enough. A model with high sensitivity and low precision floods the unit with false alarms. A model with high precision and low sensitivity stays quiet while patients deteriorate. Ask to see both, side by side.
Independent benchmarks change the conversation
The fix is the same one other high-stakes industries use: test every product against the same public, labeled data, scored the same way, by someone who is not selling it. When models are run head to head on shared benchmark datasets, the comparison stops being a battle of brochures.
Marketplaces built on this principle, including HealthMosaic, let buyers see measured results next to trusted reference models before signing anything. The best next step is still a short validation on your own de-identified data, because your patients are the benchmark that matters most.
