A long short-term memory network with SHAP interpretability for dynamic prediction of ICU delirium: development and...

ICU delirium is a common and serious complication in critically ill patients, yet existing prediction models rely predominantly on static data and lack interpretable decision logic, limiting their clinical utility. This retrospective cohort study enrolled 464 ICU patients from the First Affiliated Hospital of Zunyi Medical University as the internal cohort and 70 patients from the Second Affiliated Hospital as an external test set. A variable system comprising nine baseline characteristics and 24 dynamic variables was established through systematic evidence synthesis and a clinical pilot investigation. Dynamic variables were organized into 24-hour rolling observation windows (T1-T7), and a two-layer bidirectional long short-term memory (LSTM) network incorporating a masking mechanism and class-weight correction was developed to predict delirium occurrence in each window based on all preceding sequential inputs. The model achieved an area under the receiver operating characteristic curve (AUCROC) of 0.835 (95% CI: 0.724-0.946), sensitivity of 0.696, and specificity of 0.894 on the internal test set, with satisfactory calibration and clinical net benefit on this set across a broad probability threshold range. External validation yielded an AUCROC of 0.830 (95% CI: 0.716-0.944), with calibration and decision curve analysis in the external cohort further supporting reliable probability estimates and clinical net benefit, indicating stable cross-institutional generalizability. SHapley Additive exPlanations (SHAP) analysis revealed a clinically coherent, phase-dependent transition in dominant predictors: invasive mechanical ventilation predominated during T1-T3 (peak mean |SHAP| = 0.286 at T3), arterial pH and Sequential Organ Failure Assessment (SOFA) score emerged as principal drivers during T4-T6, and SOFA score retained the highest attribution at T7, collectively providing an evidence-based rationale for temporally differentiated monitoring and intervention strategies.