Reproducible bicarbonate thresholds predict critically ill patient mortality in an international personalized survival study.

Bicarbonate abnormalities are common in intensive care unit (ICU) patients, but prior mortality studies neglected critical illness pathophysiology. This study aimed to identify optimal admission bicarbonate thresholds for mortality prediction and develop a deep learning survival model. Using data from 208,072 ICU patients across three independent cohorts (MIMIC-IV, eICU-CRD, YHD-HOSP) in the US and China, causal inference and an integrated framework quantified in-hospital mortality, with thresholds estimated and validated across cohorts. Thresholds of 24 mEq/L and 33 mEq/L were critical; patients with levels ≤24 or >33 mEq/L had significantly higher mortality risk (adjusted HR: 1.22; HR: 1.17). Causal SurvivalNet, a deep neural network, generated personalized survival curves (integrated Brier scores: 0.15, 0.076, 0.16) accessible via a web tool. The study identified reproducible, causal bicarbonate thresholds, with Causal SurvivalNet outperforming conventional approaches, underscoring admission bicarbonate's utility for early critical care risk stratification.