Time in range and glucose trajectory phenotype, rather than glycemic variability alone, improve risk discrimination in sepsis: derivation in MIMIC-IV and external validation in the eICU Collaborative Research Database

Background Hyperglycemia and glycemic variability are common in sepsis and are linked to poor outcomes, but whether dynamic glucose measures carry prognostic information beyond established severity scores remains unclear. We tested whether a strong association with mortality also implies incremental discriminative value. Methods Retrospective cohort study in MIMIC-IV (version 3.1) with external validation in the eICU Collaborative Research Database (eICU-CRD). Adults meeting Sepsis-3 criteria were included. Variability was quantified by the coefficient of variation (CV), standard deviation, mean amplitude of glycemic excursion (MAGE), time in range (TIR; 70–180 mg/dL) and range; SHR was admission glucose divided by HbA1c-derived estimated average glucose. GBTM used cubic polynomials. The primary outcome was 28-day mortality. Analyses included multivariable regression, restricted cubic splines, time-varying Cox models, propensity matching and incremental discrimination testing. Results Among 6,290 MIMIC-IV patients, 28-day mortality was 23.8%. All variability measures were independently associated with death: CV (adjusted OR 1.022 per 1%), MAGE (OR 1.003 per 1 mg/dL) and TIR (OR 0.970 per 1%), all P < 0.0001. GBTM identified five trajectories spanning 13.0% to 51.3% mortality (very high-stable vs low-declining adjusted OR 4.28, 95% CI 2.81–6.54), and the very high-stable group retained a two-fold hazard after propensity matching (HR 2.21, 95% CI 1.52–3.19). Association and discrimination nevertheless diverged: adding CV to a base severity model raised the AUC by 0.005 (95% CI 0.001 to 0.008; P = 0.004) against 0.030 for TIR and 0.023 for the trajectory group (both P < 0.0001) and did not improve the integrated discrimination improvement (0.009; P = 0.31); combining CV with TIR and MAGE did not improve on TIR alone (AUC 0.8323 versus 0.8320). In 15,545 eICU-CRD patients the five-class structure was recovered (average posterior probability 0.882), the MIMIC-IV model transported with a mean curve difference of 6.6 mg/dL, and TIR again outperformed CV. Conclusions Glycemic variability is a strong marker of illness severity in sepsis but adds almost nothing to severity-based risk prediction, whereas time in range and glucose trajectory phenotype carry independent prognostic information. Risk stratification should therefore favour time in range over variability indices.