A Lightweight LSTM-Transformer Fusion Architecture for Real-Time Sepsis Mortality Prediction

BackgroundAccurate prediction of short-term mortality in sepsis patients is critical for timely clinical decision-making. However, existing deep learning models often focus on static physiological parameters while neglecting the dynamic response to medical interventions, leading to risk underestimation due to the "masking effect" of therapeutic measures.MethodsWe propose a lightweight hybrid deep learning framework that integrates dynamic intervention responses to predict 24-h all-cause mortality. Utilizing the MIMIC-IV v3.1 database, we included 13,788 adult sepsis patients. The model employs a dual-branch architecture: a Bidirectional LSTM to capture local temporal trends and a Transformer Encoder to extract global long-range dependencies. Crucially, we constructed a high-resolution feature set that includes vasopressor infusion rates and hourly urine output to quantify physiological feedback to resuscitation.ResultsThe proposed model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.8139, significantly outperforming seven mainstream baselines, including LightGBM (0.8015), Bi-LSTM (0.7870), and pure Transformer models (0.7704). Feature importance analysis revealed that indicators of treatment response, specifically urine output and norepinephrine dosage, were among the top predictive features, validating the clinical hypothesis that drug dependency and renal perfusion are sensitive markers of prognosis. Furthermore, external validation on the independent multi-center eICU Collaborative Research Database demonstrated robust generalizability: a zero-shot transfer yielded an AUROC of 0.6620, which improved to 0.7347 after lightweight domain adaptation fine-tuning, with a Negative Predictive Value (NPV) of 90.04%, confirming the model's cross-institutional applicability as a reliable rule-out tool.ConclusionOur LSTM-Transformer Fusion architecture effectively captures the complex "drug-physiology" interactions with low computational cost. By explicitly modeling the dynamic response to treatment and demonstrating cross-institutional generalizability through external validation on the eICU database, this lightweight model offers a robust and interpretable tool for early warning systems in resource-constrained intensive care environments.