Abstract
OBJECTIVE: This study aims to explore the capability of raw radiofrequency (RF) information in the diagnosis of the severity of Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD), especially its effectiveness in differentiating patients with varying severity of fatty liver disease. METHODS: The study included patients from the Beijing Friendship Hospital affiliated with the Capital Medical University, comprising 49 with mild fatty liver, 59 with moderate or severe fatty liver. categorized based on the magnetic resonance imaging (MRI) criteria. RF features were extracted from raw RF signal data using envelope statistics and spectral parameter calculation methods. In the data preprocessing stage, median imputation was used for handling missing values, and features were standardized. For feature selection, the Recursive Feature Elimination (RFE) method combined with the logistic regression model was used to select key features from all standardized features. During the model construction phase, a logistic regression model was used for training. For model evaluation and testing, a 100-times random sampling iteration method was used to calculate the average values and 95% confidence intervals of AUC, accuracy (ACC), sensitivity (SEN), and specificity (SPE), and ROC curves were plotted. RESULTS: During the feature selection process, M-std, m-attenuation-b-mean, Medf-mean, b-std were identified as the most influential features for the diagnosis of the severity of fatty liver. The classification results of the logistic regression model showed an average AUC value of 0.80 (95% CI: 0.75-0.86), as shown in Fig. 1, an average accuracy of 0.72 (95% CI: 0.66-0.78), average sensitivity of 0.81 (95% CI: 0.71-0.91), and average specificity of 0.64 (95% CI: 0.50-0.78). Fig. 1 Features selected and their level of importance CONCLUSION: This study indicates that combining RF information with logistic regression effectively diagnoses varying severities of fatty liver disease, offering a valuable clinical reference through careful feature selection and model evaluation.