Integrating dynamic lactate trends and clinical variables using an LSTM deep learning model to predict acute kidney injury in ICU patients

Acute kidney injury (AKI) is common among critically ill patients and is associated with substantial morbidity and mortality. Serum lactate reflects global perfusion and metabolic disturbance; however, whether time-resolved lactate dynamics can improve early AKI risk prediction remains unclear. We conducted a retrospective study using the MIMIC-IV critical care database. An LSTM model integrated dynamic lactate trajectories from the week before ICU admission and clinical predictors available before ICU admission to predict AKI occurring from ICU admission to ICU discharge. All predictors were obtained before ICU admission, and no post‑admission or post‑AKI measurements were used. Patients were partitioned at the subject level into training ( n = 2,720), validation ( n = 778), and independent test ( n = 389) sets. Performance was compared with LightGBM, XGBoost, and Random Forest using the same information content. Discrimination, calibration, and clinical utility were evaluated using ROC-AUC, PR-AUC, Brier score, bootstrap 95% confidence intervals, calibration intercept and slope, calibration plots, and decision curve analysis. Among 3,887 ICU patients, 1,845 (47.5%) developed AKI from ICU admission to ICU discharge. In the independent test set, the LSTM achieved a ROC-AUC of 0.808 (95% CI, 0.764–0.851), a PR-AUC of 0.802 (95% CI, 0.747–0.855), and a Brier score of 0.187 (95% CI, 0.171–0.202). XGBoost had the highest point estimates for ROC-AUC (0.815) and PR-AUC (0.815) and the lowest Brier score (0.185), although the bootstrap confidence intervals overlapped those of the LSTM and Random Forest. At the validation-determined threshold of approximately 0.453, the LSTM sensitivity was 0.832 and specificity was 0.632. Subgroup analyses were exploratory and showed heterogeneity with wide, overlapping confidence intervals. For in-hospital mortality, the rechecked PR-AUC was 0.737. Modeling dynamic lactate trajectories together with pre-admission clinical variables provided useful discrimination for AKI occurring from ICU admission to ICU discharge. The LSTM performed comparably to strong tree-based baselines rather than consistently outperforming them. The findings support further evaluation, but temporal and external validation and prospective assessment of threshold-linked clinical actions are required before clinical deployment.