Study on a 28-day prognostic model for ICU patients with AB infection based on machine learning

This study aims to establish 28-day prognostic model for intensive care unit (ICU) patients with Acinetobacter baumannii (AB) infection based on machine learning. This retrospective study collected clinical data of patients admitted to the Intensive Care Medicine Department (ICU) of the Second Hospital of Nanjing from March 2021 to October 2023. The data underwent univariate and multivariate COX survival analysis, Lasso regression analysis, survival curve analysis, and model construction with SPSS and R language. The model was evaluated with area under the curves, calibration curves, and decision curve analysis curves to assess its clinical utility. Univariate COX survival analysis revealed that 11 variables: hormones, sequential organ failure assessment (SOFA), CD4 T, ThTsCD4CD8, hospital-acquired pneumonia, ventilator-associated pneumonia, pro-procalcitonin, pro-interleukin-6 (IL-6), procalcitonin, IL-6, and duration of mechanical ventilation (DurationMV) were significant risk factors for mortality in AB-infected patients (P < .05). In the Lasso regression analysis, lambda.1se (0.147) was selected as the optimal λ value. Variable selection was performed with the K-fold cross-validation method in the glmnet package, involving 9 influencing factors including albumin, acute physiology and chronic health evaluation II, CD4.T, SOFA, hospital-acquired pneumonia, ventilator-associated pneumonia, pro-IL6, IL-6, and DurationMV. An epwise Cox regression model was established with a global Schoenfeld test value of 0.1629. The mean square difference between survival probabilities and actual results was consistently <0.25, indicating accurate model predictions. Based on above-mentioned factors, multivariate Cox analysis showed that SOFA and DurationMV were independent risk factors for fatal outcomes, with no correlation between variables and variance inflation factor values <5. Grouping based on SOFA score showed statistically significant differences in mortality rates (P < .05), with higher scores associated with shorter survival times. DurationMV had greater prognostic value than the SOFA "score," enhancing academic accuracy. Using DurationMV and SOFA variables, a survival analysis line chart model for 7, 14, and 28 days was constructed. The model was validated on 147 patients randomly divided into a training set and a validation set, achieving an area under the curve of over 80%, with calibration curves close to the diagonal line, indicating relative stability in predicting the 28-day prognosis of AB-infected patients. A 28-day prognostic model for ICU patients with AB infection based on machine learning was successfully developed and its value validated.