Abstract:

For patients admitted to intensive care with acute ischemic stroke, machine-learning models offer a way to estimate mortality risk from routinely collected clinical information. Whether predictions remain dependable outside the development setting is less clear: hospitals differ in case mix, clinical practice, variable definitions, and the completeness of their records. This narrative review examines studies of mortality prediction in critically ill ischemic stroke patients through the connected issues of external validation, predictor stability, missing-data management, and model assessment. It considers primary investigations, existing reviews, and methodological guidance, distinguishing ischemic stroke from mixed stroke cohorts and in-hospital mortality from other outcome horizons. The review examines how study populations, datasets, predictor definitions, validation designs, and evaluation measures shape the interpretation of reported model performance. Discrimination, calibration, clinical utility, and reproducibility are treated as separate considerations when assessing whether a prediction model can be evaluated meaningfully across healthcare settings.

Keywords:

Ischemic stroke; mortality prediction; machine learning; intensive care unit; external validation; generalizability; calibration; missing data.

Aim/Objectives

Methods