Clinical AI

NHAMCS-ED looks almost ideal for clinical AI experiments: it combines structured emergency department visit variables with short reason-for-visit text. But it is a national probability survey, not a hospital EHR extract. Across my IV fluid utilization and hospital admission prediction projects, the biggest lesson was that the hard part is often not the model itself; it is deciding what the model is allowed to know, what the data actually represent, and what claims the analysis can support. ...

SHAP Explains the Model, Not the Patient In clinical AI papers, I often see SHAP plots used as if they explain the disease process itself. The figure may look technical and persuasive, but the interpretation often goes too far: “This variable has a high SHAP value, so it is an important clinical cause.” That is usually not justified. SHAP is useful. It can show what a trained model relied on when making a prediction. But its boundary should be clear: ...