Answers are grounded in this paper and cite their sources.
Acute kidney injury (AKI) is common in intensive care units (ICUs), and class imbalance and heterogeneous data complicate model development. The value of combining longitudinal physiology with static patient context, relative to imbalance-handling and conventional fusion strategies, remains uncertain. To develop and evaluate a Dual-Stream Context-Aware Network (DSCAN) that uses static patient characteristics to modulate temporal representations for AKI risk prediction in three ICU cohorts. We retrospectively constructed cohorts from MIMIC-IV, eICU-CRD, and West China Tianfu Hospital (TF). AKI was defined using KDIGO criteria, and samples within 24 h before first AKI onset were labeled as positive warning windows. Each cohort was independently divided into training, validation, and test sets. DSCAN combined a Transformer-based temporal stream with a static-feature stream through a sigmoid gating mechanism. It was compared with dynamic-only models, class-imbalance strategies, and early-fusion baselines using AUROC and threshold-dependent classification metrics. Additional analyses assessed patient-independent partitioning in MIMIC-IV, external evaluation of the fixed eICU-trained model, and validation-based alarm-threshold adjustment. The study included 67,797 MIMIC-IV, 77,150 eICU-CRD, and 5,131 TF patients. DSCAN achieved AUROCs of 0.874, 0.857, and 0.676, respectively. AUROC gains over the strongest comparator were small in MIMIC-IV and eICU-CRD, and DSCAN was not the best-performing model by AUROC in TF. Across cohorts and threshold strategies, improvements in sensitivity at selected operating points were accompanied by variable specificity, precision, and F1-score. Predicted risks exceeded observed event rates across cohorts, and decision-curve net benefit was limited beyond low threshold probabilities. The fixed eICU-trained model achieved external AUROCs of 0.752 in MIMIC-IV and 0.787 in TF. Validation-based threshold adjustment increased specificity at lower sensitivity. SHAP attribution identified serum creatinine-related variables, urine output, and hemodynamic indicators as influential predictors. Static-conditioned temporal modeling achieved cohort-dependent discrimination comparable to strong baseline methods. Risk overestimation and limited decision-curve net benefit constrain its clinical application. Further evaluation should address probability recalibration, clinically defined alarm thresholds, and prospective clinical utility.
No takes yet. Share an insight, caveat, or question.
Tao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: