Inpatient Deterioration & Risk Monitoring

AI that watches an admitted patient continuously and predicts an adverse event early enough to act on: sepsis and clinical decompensation, respiratory and hemodynamic instability, falls, and obstetric deterioration in labor. The input is a stream rather than a study, which separates this category from imaging triage, where a single scan is read once. The setting is acute inpatient, which separates it from remote patient monitoring in the home and from fall detection in residential senior living. Two failure modes define the category. The first is alert fatigue: here the failure can be the output being ignored rather than being wrong, because a model can be accurate and still clinically useless if it fires often enough that staff stop responding to it. The second is lead time, which trades against precision. An alert thirty minutes before an arrest is close to worthless and one six hours earlier is actionable, so a sensitivity figure means very little without the median lead time and the alert burden per patient day beside it. Ask for all three together, and ask separately whether any external party has validated the model on their own data, because independent revalidation of a widely deployed proprietary sepsis model is exactly how its weaknesses came to light.