OBJECTIVE: To explore lung ultrasound radiomics features which related to extravascular lung water index (EVLWI), and to predict EVLWI in critically ill patients based on lung ultrasound radiomics combined with machine learning and validate its effectiveness. METHODS: A retrospective case-control study was conducted. The lung ultrasound videos and pulse indicated continuous cardiac output (PiCCO) monitoring results of critically ill patients admitted to the department of critical care medicine of the First Affiliated Hospital of Guangxi Medical University from November 2021 to October 2022 were collected, and randomly divided into training set and validation set at 8:2. The corresponding images from lung ultrasound videos were obtained to extract radiomics features. The EVLWI measured by PiCCO was regarded as the "gold standard", and the radiomics features of training set was filtered through statistical analysis and LASSO algorithm. Eight machine learning models were trained using filtered radiomics features including random forest (RF), extreme gradient boost (XGBoost), decision tree (DT), Naive Bayes (NB), multi-layer perceptron (MLP), K-nearest neighbor (KNN), support vector machine (SVM), and Logistic regression (LR). Receiver operator characteristic curve (ROC curve) was plotted to evaluate the predictive performance of models on EVLWI in the validation set. RESULTS: A total of 151 samples from 30 patients were enrolled (including 906 lung ultrasound videos and 151 PiCCO monitoring results), 120 in the training set, and 31 in the validation set. There were no statistically significant differences in main baseline data including gender, age, body mass index (BMI), mean arterial pressure (MAP), central venous pressure (CVP), heart rate (HR), cardiac index (CI), cardiac function index (CFI), stroke volume index (SVI), global end diastolic volume index (GEDVI), systemic vascular resistance index (SVRI), pulmonary vascular permeability index (PVPI) and EVLWI. The overall EVLWI range in 151 PiCCO monitoring results was 3.7-25.6 mL/kg. Layered analysis showed that both datasets had EVLWI in the 7-15 mL/kg interval, and there was no statistically significant difference in EVLWI distribution. Two radiomics features were selected by using LASSO algorithm, namely grayscale non-uniformity (weight was -0.006 464) and complexity (weight was -0.167 583), and they were used for modeling. ROC curve analysis showed that the MLP model had better predictive performance. The area under the ROC curve (AUC) of the prediction validation set EVLWI was higher than that of RF, XGBoost, DT, KNN, LR, SVM, NB models (0.682 vs. 0.658, 0.657, 0.614, 0.608, 0.596, 0.557, 0.472). CONCLUSIONS The gray level non-uniformity and complexity of lung ultrasound were the most correlated radiomics features with EVLWI monitored by PiCCO. The MLP model based on gray level non-uniformity and complexity of lung ultrasound can be used for semi-quantitative prediction of EVLWI in critically ill patients.
Humans ; Extravascular Lung Water/diagnostic imaging* ; Retrospective Studies ; Critical Illness ; Case-Control Studies ; Bayes Theorem ; China ; Lung/diagnostic imaging*

Adult ; Capillary Permeability ; Extracorporeal Membrane Oxygenation ; Extravascular Lung Water ; Humans ; Lung/diagnostic imaging* ; Prognosis ; Respiratory Distress Syndrome/therapy*