Estimating the dose-response relationship between first-24-h mean arterial pressure and ICU mortality in patients receiving norepinephrine: a GPU-assisted causalEGM analysis

초록

Norepinephrine is widely used in critically ill patients, but the relationship between early mean arterial pressure (MAP) exposure and mortality remains heterogeneous. We applied a GPU-assisted and parallelizable deep learning causal inference workflow to estimate the dose-response relationship between calculated average MAP within the first 24 h after intensive care unit (ICU) admission and ICU mortality among adult patients receiving norepinephrine. Using Medical Information Mart for Intensive Care-III data, 1373 patients were analyzed with causal inference by encoding generative modeling (causalEGM), treating calculated average MAP as a continuous treatment and ICU mortality as the outcome. Model performance, calibration, model-fixed bootstrap uncertainty, limited full-refit bootstrap sensitivity, subgroup conditional dose-response curves, and spline-based sensitivity analyses were evaluated. The fitted causalEGM model showed moderate test-set discrimination (area under the receiver operating characteristic, 0.717; Brier score, 0.186). The overall average dose-response curve suggested a nonlinear MAP-mortality relationship, with a model-derived statistical nadir at 93.9 mmHg (95% model-fixed bootstrap interval, 93.3-94.9 mmHg). The 1% practical threshold was lower, at 87.3 mmHg (95% interval, 86.2-88.4 mmHg). Some subgroup estimates were boundary-dominated, and limited full-refit bootstrap analyses repeatedly placed the nadir near the upper empirical support boundary, indicating sensitivity to model retraining and empirical support. This GPU-assisted, parallelizable causalEGM analysis provides hypothesis-generating evidence of a nonlinear association between calculated average MAP within the first 24 h and ICU mortality. The fitted nadir should not be interpreted as a direct clinical MAP target. External and prospective validation is required before clinical implementation.

키워드