Automated whole-brain MRI segmentation combined with clinical scores for the diagnosis of sepsis-associated encephalopathy: a case-control study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Sepsis-associated encephalopathy (SAE) is a serious complication of sepsis that can lead to high mortality and poor neurological outcomes. Current diagnostic methods, whether based on clinical scores or brain images, remain inaccurate. We aimed to formulate a multimodal model that incorporated three-dimensional (3D) magnetic resonance imaging (MRI) neuroanatomical characteristics with clinical scores to improve SAE detection. Methods In this retrospective case-control study, 35 healthy controls and 72 patients with sepsis (34 SAE, 38 non-SAE) were included and scanned using 3D T1-weighted MRI. Whole-brain segmentation was performed using FreeSurfer software. Clinical characteristics, including age, gender, and clinical scores (Acute Physiology and Chronic Health Evaluation II [APACHE II] and Sequential Organ Failure Assessment [SOFA]), were captured within the electronic health record data. Feature selection was performed using LASSO regression with 10-fold cross-validation exclusively on the training set. We trained eight machine learning models based on a combination of clinical scores and neuroimaging features, with hyperparameter tuning conducted via GridSearchCV using 3-fold stratified cross-validation. The performance of each model was evaluated based on stratified hold-out validation, and SHapley Additive exPlanations (SHAP) analysis was used to interpret the optimal model. Results Rigorous quality control procedures were applied, and 540 features were analyzed across the three groups. We identified 27 reliable indicators associated with SAE, including clinical scores (APACHE II and SOFA) and neuroanatomical characteristics (hippocampus, thalamus, amygdala, lateral ventricle, and globus pallidus). Among eight machine learning models trained and evaluated, the XGBoost model displayed significantly superior discrimination and calibration performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.92. Key predictors included APACHE II, SOFA, and volumetric changes in cerebellar cortex, hippocampus, and amygdala. SHAP analysis revealed nonlinear relationships wherein elevated clinical scores and neuroanatomical atrophy predicted SAE. Conclusions This multimodal model integrates clinical scores and neuroimaging features of SAE and demonstrates promising discriminative ability in differentiating healthy controls, septic patients without SAE, and SAE patients. The findings provide candidate neuroimaging metrics that may aid in diagnosis and offer potential insights into the structural correlates of SAE. However, these results are preliminary and require prospective validation in larger, multi-center cohorts before clinical translation.