Daily Sepsis Research Analysis

Sepsis prediction models trained on ICU data often fail to generalize under external validation because of distribution shift. Prior studies have focused on direct model deployment or conventional transfer learning methods (e.g., fine-tuning), yet systematic exploration of alternative strategies remains limited. We quantify shifts across three harmonized adult ICU cohorts (HiRID, MIMIC-IV, eICU; 216,536 stays) and compare five deployment strategies: generalization, fine-tuning/retraining, target training, supervised domain adaptation (DA), and fusion-training, across multiple deep learning architectures, and four target-data regimes (none; small ≤ 8k; medium 8-32k; large ≥ 32k stays). Fine-tuning consistently underperforms, even though it has been the go-to method in literature. Retraining and fusion perform best in small and large target data regimes, while DA yields the most stable gains with medium target data, improving AUROC and normalized AUPRC over other methods. These results argue for moving beyond routine fine-tuning for sepsis prediction and selecting strategies by target-data availability and operational context.

OBJECTIVES: To evaluate the diagnostic performance of a rapid point-of-care immunoassay measuring pancreatic stone protein (PSP) for early sepsis identification within the first three days of ICU admission. Subgroup analyses (sex, age, febrile status) were conducted, and the combined diagnostic value of PSP and C-reactive protein (CRP) was assessed. DESIGN: Multicenter, prospective, observational study. PATIENT: Four hundred sixty-six adults the ICU. SETTING: Six ICUs in the United States who were expected to required at least 24 hours of ICU care. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: We calculated the Youden Index to evaluate the clinical performance of the PSP assay, and the resulting threshold was used to identify patients with sepsis. Diagnostic performance metrics included sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (LR+), and negative likelihood ratio (LR-). Receiver operating characteristic analysis were performed for PSP and CRP. At the optimal PSP cutoff point of 117 ng/mL, PSP demonstrated a sensitivity of 74.2%, specificity of 67.8%, accuracy of 71.0%, PPV of 70.3%, NPV of 71.9%, and LR+ and LR- ratios of 2.30 and 0.38, respectively. Combining PSP and CRP improved diagnostic specificity to 95.2%. Subgroup analyses demonstrated consistent performance across sex, and higher specificity was observed in patients 18-60 years old. In febrile patients, PSP achieved high specificity (87.5%) but lower sensitivity (63.6%). In non-febrile patients, sensitivity and specificity were 67.7% and 76.6%, respectively. CONCLUSIONS: PSP can serve as a biomarker for the early identification of sepsis. Diagnostic performance across diverse ages, sex, and clinical presentation supports the assay's broad applicability. The combination of PSP and CRP enhances diagnostic specificity for sepsis detection, offering a complementary approach to improve sepsis detection and lead to earlier appropriate management.

Accurate identification of early pediatric abdominal sepsis (PAS) is essential to improving outcomes, yet most existing pediatric sepsis criteria and scoring tools primarily focus on cardiopulmonary dysfunction and overlook early intra-abdominal infections. To address this gap, we combined the real-world data with explainable machine learning to develop the Abdominal Sepsis Diagnosis model (ABSeD) for clinical decision support. The model construction used the retrospective data from 6566 pediatric patients who were admitted to the Children's Hospital, Zhejiang University School of Medicine from 2019 to 2023. Prospective data from 308 recruited patients across seven independent hospitals collected between January and March 2025 served as an external validation cohort. PAS status was determined through consensus or by reviewing laparoscopic surgery records. Multiple machine learning algorithms were compared, and the optimal model was further refined by hyper-parameter tuning. The ABSeD model, integrating nine routine clinical variables, demonstrated high diagnostic accuracy (training set: AUC = 0.934, 95% CI: [0.912, 0.950]; accuracy = 0.870, precision = 0.910), and robust multicenter generalizability (AUC = 0.928, 95% CI: [0.895, 0.961]; accuracy = 0.873, precision = 0.924). This model offers an explainable and practical digital tool for early detection of PAS, with potential to enhance timely intervention in hospitalized children with suspected or clinically identified intra-abdominal septic pathology.