Early warning for acute pyelonephritis with sepsis: risk prediction model for patients without renal medullary necrosis lesions.

Background: Sepsis is one of the key factors leading to mortality in hospitalized patients. Acute pyelonephritis, as a common upper urinary tract infection, serves as a significant source of infection, with some patients progressing to sepsis, which markedly increases mortality risk. This study aimed to utilize a large-scale public intensive care database (MIMIC-IV) to construct and validate a model specifically designed for predicting sepsis risk in such patients, thereby providing clinicians with accurate and practical early warning tools. Methods: This retrospective cohort study employed data from the MIMIC-IV database version 3.1, enrolling 663 patients. Demographic characteristics, comorbidities (Charlson Index), and multiple vital signs and laboratory parameters within the first 24 h of hospital admission were collected. Patients were randomly divided into a model development set (465 cases) and an internal validation set (198 cases) at a 7:3 ratio. Univariate logistic regression, cross-validation LASSO regression, and multivariate logistic regression models were performed. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC), while calibration and decision curve analysis (DCA) were employed to evaluate calibration accuracy and clinical Sepsis incidence rates utility. Results: The sepsis incidence rates in the development set and validation set were 12.90 and 12.63%, respectively. LASSO regression identified five predictive variables for the final model: Charlson Index, bicarbonate, calcium, mean corpuscular volume (MCV), and admission age. Multivariate analysis revealed that the Charlson index (OR = 1.10, 95% CI: 0.96-1.27), admission age (OR = 1.02, 95% CI: 1.00-1.04), and MCV (OR = 1.05, 95% CI: 1.00-1.10) were independent risk factors, while bicarbonate (OR = 0.90, 95% CI: 0.83-0.98) and calcium (OR = 0.27, 95% CI: 0.16-0.43) served as protective factors. The model demonstrated good apparent discrimination with AUCs of 0.820 (95% CI: 0.757-0.883) in the development set and 0.830 (95% CI, 0.737-0.924) in the validation set; internal bootstrap optimism correction yielded a bias-adjusted AUC of 0.805, confirming robustness against overfitting. Calibration was satisfactory, with decision curve analysis confirming net clinical benefit across threshold ranges of 1-80% (development) and 1-76% (validation). At the Youden-optimized cutoff of 0.194, the model achieved a sensitivity of 0.667 and specificity of 0.864 in the development set, and 0.600 and 0.803, respectively, in the validation set. Conclusion: This study successfully established a sepsis risk prediction model for acute pyelonephritis hospitalized patients without renal medullary necrosis lesions, based on five routine indicators. The model demonstrated excellent predictive performance and clinical utility in internal validation, providing a bedside quantitative tool for instantaneous risk stratification at hospital admission, facilitating early clinical vigilance and prioritized monitoring in this specific population.