Comment on “Development and validation of a deep learning ultrasound radiomics model for predicting drug resistance in lymph node tuberculosis”

Dear Editor, We read with great interest the multicenter study by Xu et al entitled “Development and validation of adeep learning ultrasound radiomics model for predicting drug resistance in lymph node tuberculosis”[1]. The authors integrated radiomics and ensemble machine learning to develop a predictive model for drug resistance in lymph node tuberculosis (LNTB), providing novel insights into non-invasive diagnosis and treatment optimization. Several aspects, however, merit further clarification and discussion. First, although the cohort consisted of pathologically confirmed cervical tuberculous lymphadenitis, the inclusion and exclusion criteria were incompletely detailed. Clinical variables such as HIV infection, diabetes, immunodeficiency, or prior anti-tuberculosis therapy may influence imaging heterogeneity and resistance patterns, potentially introducing selection bias[2,3]. Secondly, although variance thresholds and Pearson correlation coefficients were utilized for feature screening, employing more robust feature-selection methodologies, such as least absolute shrinkage and selection operator (LASSO) regression, could have improved feature stability and model generalizability[4]. Third, subgroup analyses were not disclosed. Stratification based on lymph node size, presence of caseous necrosis, concurrent pulmonary tuberculosis, or disease duration could validate the model’s resilience across clinically pertinent subpopulations and enhance interpretability[5]. This research significantly advances the domain of LNTB resistance prediction in radiomics. In future multicenter research, it is essential to define explicit inclusion criteria, utilize stringent feature selection techniques, and execute thorough subgroup analysis. These measures are essential for improving the translational applicability and clinical dependability of these predictive models.