AI Model Predicts 28-Day Death Risk in Septic Shock From Routine Admission Blood Tests

When a patient crashes into septic shock in the emergency department, clinicians have minutes to decide who needs the most aggressive resuscitation the hospital can offer. A research team at the First Affiliated Hospital of Wenzhou Medical University in China now reports an automated machine learning model that estimates a septic shock patient’s 28-day risk of death from just six routine admission blood parameters, calculated in real time at the bedside. In the study, published in Frontiers in Medicine, the model achieved an area under the ROC curve of 0.8911 on a locked independent test set and 0.861 in a separate temporal validation cohort of newly admitted patients, outperforming the traditional SOFA, APACHE II and qSOFA scoring systems on the same data. The work addresses a persistent gap in emergency critical care: the leading scoring tools were designed for ICU chart review, not for split-second triage decisions made before any resuscitation begins.

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection, and septic shock adds circulatory and metabolic collapse on top of it, making it one of the leading causes of mortality among intensive care patients worldwide. Despite the Surviving Sepsis Campaign and decades of diagnostic progress, both incidence and death rates remain stubbornly high, and sepsis stands as a primary risk factor for in-hospital mortality and 30-day readmission. Traditional prognostic instruments such as SOFA, APACHE II, qSOFA and the Modified Early Warning Score rely on static, single-time-point assessments with fixed linear weights, which cannot capture the nonlinear interactions that sepsis generates across the immune, coagulation and metabolic systems. Emerging biomarkers, meanwhile, often carry prohibitive costs that block implementation in developing regions, and single biomarkers show limited prognostic power against sepsis’s temporally dynamic host response. The Wenzhou team’s answer was to mine the cheap, universally available routine blood panel that every emergency department already draws within minutes of triage.

The study retrospectively enrolled 913 patients diagnosed with septic shock at the tertiary hospital between January 2019 and January 2025, splitting them 8:2 into a training and development set of 730 patients and a fully locked test set of 183. The cohort was grimly balanced: 458 patients, or 50.16 percent, died within 28 days, and stratified randomization kept mortality proportions nearly identical between training and testing groups. All laboratory indices were measured within two hours of admission, drawn immediately after emergency triage and before any fluid resuscitation, vasoactive agents or antibiotics, so the values reflected each patient’s baseline physiological state rather than a post-intervention snapshot. Stringent exclusions removed pregnant patients, those under 18, people with chronic liver or kidney dysfunction, immunologic disorders or malignancy, hematologic diseases, pre-admission trauma or active bleeding, pre-hospital blood transfusion, and hospital-acquired septic shock episodes, among others. Patients who withdrew from treatment before the 28-day outcome assessment were prospectively excluded during screening, which allowed complete 28-day mortality ascertainment through electronic record linkage and eliminated loss-to-follow-up bias.

The technical heart of the study is an AutoML framework driven by an Improved Rüppell’s Fox Optimizer, or IRFO, a metaheuristic algorithm that simulates fox foraging behavior. The researchers enhanced the original RFO with two strategies: chaotic mapping initialization, which spreads the initial candidate population more evenly across the search space, and dynamic Lévy flight steps, which balance exploration of new regions against exploitation of promising ones, reducing the local-optimum traps that plague metaheuristics in feature selection. During each iteration, IRFO performed a dual optimization task, simultaneously selecting a feature subset from a candidate pool in discrete space and tuning hyperparameters for XGBoost or LightGBM base learners in continuous space, guided by the objective of maximizing five-fold cross-validated AUC on the training fold. The framework was wrapped in a rigorous double-layer nested cross-validation pipeline: the inner loop optimized features and hyperparameters on each training fold, the outer loop evaluated performance on untouched validation folds, and the locked test set stayed isolated until the final model was frozen. Synthetic Minority Over-sampling Technique addressed class imbalance, and missing values, all below 5 percent, were imputed by distribution type before training.

On the training set, the AutoML model posted a ROC-AUC of 0.9127 and a precision-recall AUC of 0.9059, with an F1-score of 0.8388, edging out LightGBM at 0.9021, XGBoost at 0.8997, AdaBoost at 0.8443, logistic regression at 0.8042 and support vector machines at 0.7824. Applied to the independent test set, it maintained an ROC-AUC of 0.8911 and PR-AUC of 0.8790, with sensitivity of 0.9022 and specificity of 0.7143 at the default threshold of 0.5, and the lowest Brier score of 0.136 among all evaluated models. Calibration was excellent, with a Hosmer-Lemeshow chi-square of 3.29 and p-value of 0.915, indicating strong agreement between predicted probabilities and observed outcomes. An ablation experiment confirmed the optimizer mattered: replacing IRFO with Bayesian optimization, random search or grid search dropped test ROC-AUC to 0.8718, 0.8643 and 0.8567 respectively, while IRFO’s optimization superiority had been independently validated on all 12 CEC2022 benchmark functions against whale, grey wolf, particle swarm and genetic algorithm competitors.

The final model settled on six inputs: the neutrophil-to-lymphocyte ratio, hemoglobin, procalcitonin, D-dimer, mean platelet volume and age. LASSO regression with L1 regularization independently verified the feature screening, selecting seven key predictors under the sparse Lambda1SE criterion that encompassed all AutoML-identified features, adding C-reactive protein as a seventh. SHAP analysis, built on cooperative game theory, ranked NLR first in global importance, followed by hemoglobin, procalcitonin, D-dimer, MPV and age. The pathophysiology reads coherently: NLR dysregulation signals systemic hyperinflammation coupled with evolving immunoparalysis, where neutrophilia drives organ damage through NETosis while lymphopenia marks adaptive immune exhaustion. Hemoglobin reduction reflects not only hemorrhage but inflammation-driven erythropoiesis suppression, exacerbating tissue hypoxia, and SHAP interaction mapping showed hemoglobin deficiency potentiating NLR’s inflammatory impact through amplified hypoxic stress. Elevated D-dimer signals dysregulated immunothrombosis, and rising MPV indicates accelerated platelet turnover during thrombocytopenia, linking platelet hyperactivation with endothelial damage.

SHAP decision paths made the model’s logic tangible through individual case studies. One high-risk patient with NLR of 29.59, hemoglobin of 65 g/L, procalcitonin of 17.27 g/L, D-dimer of 1572 μg/L, MPV of 10.76 fL and age 71 received a predicted 28-day mortality of 99.6 percent, with NLR and hemoglobin contributing the largest SHAP values in a pattern the authors described as synergistic inflammation-anemia deterioration. A moderate-risk patient, whose procalcitonin of 36.59 and D-dimer of 2928 pushed risk up while a relatively favorable NLR of 9.84 pushed it down, landed at 50.8 percent predicted mortality with opposing contributions nearly canceling out. A low-risk patient with NLR 7.59, hemoglobin 109 g/L, procalcitonin 0.72 and D-dimer 505 received just 0.3 percent, reflecting a protective combination across inflammation, infection and coagulation indicators. The authors caution that SHAP values reflect model-conditional contributions within the current feature set and should not be read as independent clinical effect sizes or causal importance.

Decision Curve Analysis demonstrated superior net clinical benefit over conventional intervention strategies whenever the risk threshold exceeded 10 percent, and the team translated the continuous probability output into a two-stage clinical pathway. At the clinically recommended operating threshold of 0.35, sensitivity rose to 0.951, catching more than 95 percent of eventual fatalities, at the cost of specificity falling to 0.648, misclassifying some actual survivors as high-risk. Raising the threshold to 0.65 improved specificity to 0.890 but dropped sensitivity to 0.659, missing over 30 percent of fatal cases. The proposed framework therefore flags patients at predicted risk of 0.35 or above for maximum-level alerts, with all Sepsis Bundle interventions, including blood cultures, broad-spectrum antibiotics, hemodynamic monitoring and goal-directed fluid resuscitation, completed within one hour alongside ICU admission priority. Patients below the threshold enter an enhanced monitoring tier with sequential model-driven risk re-evaluations at 6, 12 and 24 hours. An interactive decision support interface allows clinicians to enter admission indicators and receive an instantaneous mortality probability from 0 to 100 percent.

Crucially, the researchers tested the frozen model on 169 consecutive new patients admitted from February to December 2025, a temporal validation cohort that played no role in development, retraining, feature reselection or threshold optimization. The two cohorts were comparable on age, NLR, MPV, hemoglobin, procalcitonin, D-dimer and CRP, and the frozen model achieved a ROC-AUC of 0.861 with a 95 percent confidence interval of 0.799 to 0.912, PR-AUC of 0.848, sensitivity of 0.847 and specificity of 0.750, with a Brier score of 0.153. The authors are candid about limitations: the single-center retrospective design risks regional selection bias, and the cohort’s near 1:1 death-to-survival composition, deliberately balanced for statistical power, differs sharply from the roughly 23 percent real-world 28-day mortality, introducing spectrum bias that inflates the reported test AUC of 0.891 into an upper-bound estimate. The static admission-based design also cannot track longitudinal disease evolution or guide decisions such as antibiotic de-escalation. A prospective multi-center external validation study is underway, and future iterations plan serial biomarker measurements at 0, 24 and 72 hours, deep learning for temporal features, and cluster-randomized trials of the decision support system in live emergency workflows.

Subject of Research: Automated machine learning prognostic model for emergency septic shock mortality prediction

Article Title: Development of an automated machine learning-based prognostic prediction model for emergency septic shock patients

Article References: Wei, S., Chen, H., Li, X., Ke, T., Chen, X., Lin, C., Qian, L., & Wang, X. (2026). Development of an automated machine learning-based prognostic prediction model for emergency septic shock patients. Frontiers in Medicine, 13, Article 1860425. https://doi.org/10.3389/fmed.2026.1860425

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1860425

Keywords: septic shock, machine learning, AutoML, emergency medicine, prognostic prediction, explainable AI, critical care, SOFA, APACHE II, SHAP, Development, automated