This proof-of-concept retrospective study, utilizing the MIMIC-IV v2.2 database (2008–2019), demonstrates that an interpretable machine learning model based on routinely collected electronic health record data from ICU patients can effectively predict the initiation of continuous vasopressor infusions in sepsis cases.

Researchers identified adult Sepsis-3 ICU stays, defining cases as those starting vasopressors 6–48 hours post-admission and controls as sepsis patients with ≥48-hour stays without vasopressors.

Through propensity score matching on factors like age, sex, Charlson comorbidity index, modified SOFA score, weight, and early lactate/hematocrit, they analyzed 1,539 cases and 1,431 controls, with an independent validation set of 751 stays.

Key findings highlight the model’s ability to provide clinically interpretable predictions with actionable lead time, identifying risks such as declining blood pressure, elevated lactate, and abnormal hematocrit. However, results are limited to internal validation, warranting caution; external multi-center validation and prospective silent-mode trials are essential to establish generalizability and evaluate impacts on time-to-vasopressor, fluid management, and overall patient outcomes.

Read More