PROBLEM: At present a routine identification of Adverse Drug Events (ADEs) is not available for Dutch Intensive Care Units (ICU). This hampers timely ADE detection. Especially drug-induced acute kidney injury (DAKI) and therapy failure due to underdosing of antibiotics (TFUA) present serious concerns for ICU. Algorithms obtained by advanced machine learning (ML) techniques may offer real-time DAKI/TFUA detection, allowing timely corrective actions and promoting a continuous learning cycle.
MAIN OBJECTIVE: To provide and assess automated DAKI/TFUA detection algorithms obtained by advanced ML techniques in the ICU setting.
STUDY DESIGN: Retrospective cohort study.
STUDY POPULATION: More than 100,000 multicenter ICU admissions from 14 Dutch ICUs with a complete medication history, demographic, diagnostic, and progression of physiological parameters over time.
OUTCOME MEASURES: (1) Predictive performance of ML models in terms of discrimination, accuracy and calibration. (2) The associations between medication and AKI/TFUA, in terms of adjusted odds ratios. (3) Positive predictive value of the DAKI/TFUA detection algorithms.
SAMPLE SIZE/DATA-ANALYSIS: DAKI affects around 7% of ICU patients, and TFUA 8,5%. Therefore, to detect at least 100 DAKI/TFUA cases, at least 1428/1177 ICU admissions are needed. Using various advanced ML approaches, prediction and causal models for DAKI/TFUA will be developed and subgroup discovery analyses will be conducted.