Response to commentary on predictive modeling for cervical lymph node metastasis in esophageal squamous cell carcinoma

To the Editor, We sincerely appreciate the thoughtful comments regarding our recent study on a non-invasive approach using ultrasonography and machine learning for selective cervical lymphadenectomy in thoracic esophageal cancer[1]. The commentary highlights important limitations and future directions that resonate with our own reflections[2]. We welcome this opportunity to clarify our methodological rationale, acknowledge the constraints, and outline our vision for advancing this field. We recognize that certain aspects of our study could benefit from refinement to strengthen reliability and clinical applicability. The use of single-center data, the absence of morphological indicators such as the long/short diameter (L/S) ratio, and the selection of a CT-based nomogram as the baseline comparator may have influenced overall robustness and methodological depth[2]. While the predictive performance of our best-performing model is encouraging, further improvement is needed to achieve the level of predictability expected for routine clinical use. These considerations emphasize the importance of ongoing optimization and validation, while affirming that our initial findings provide a solid benchmark for future research. The constraints we faced were primarily practical. Our dataset originated from a single institution due to resource limitations, which ensured consistent imaging protocols and clinical annotation but reduced diversity; this was reasonable for proof‑of‑concept work. The baseline model was chosen because it is among the few available tools for evaluating cervical nodal involvement and was compatible with variables in our dataset. Broader comparisons with MRI‑, PET‑, or radiomics‑based models would be informative; however, these approaches rely on modalities and predictors not available in our retrospective cohort and will be prioritized in prospective studies. Similarly, the exclusion of the L/S ratio and other advanced ultrasound features reflected incomplete data documentation in the initial cohort. These decisions, made to maintain scientific rigor within the scope of available resources, help explain the modest performance and incremental gain over the baseline comparator. Nonetheless, they represent a meaningful step toward developing a cost-effective, non-invasive approach for individualized surgical planning and define clear targets for improvement. Looking ahead, we envision several strategies to address these gaps. Incorporating additional ultrasound features such as the L/S ratio, echogenicity, and vascularity patterns should enhance morphological characterization and improve predictive accuracy[3]. We also plan to explore convolutional neural networks to interpret ultrasound images holistically rather than relying solely on predefined criteria. Expanding the dataset through multicenter collaborations is essential to increase diversity and enable rigorous external validation for generalizability. Beyond imaging, integrating multimodal data, including clinical variables, molecular markers, and circulating tumor DNA, offers a promising opportunity to refine risk stratification and personalize treatment decisions[4]. On the modeling side, advanced algorithms such as deep neural networks and large language models, combined with multimodal inputs incorporating clinical notes, laboratory results, and radiologic data, may further improve performance while reducing manual feature extraction[5]. These innovations could enhance robustness and efficiency, bringing the approach closer to clinical adoption. Beyond technical considerations, it is important to situate this work within a broader clinical framework for managing cervical lymph node metastasis in esophageal squamous cell carcinoma. For patients with nodes exceeding a defined size threshold (e.g., ≥ 15 mm), current imaging modalities such as CT, PET, and ultrasonography are generally effective[6–8], and targeted interventions, including radiotherapy and neck dissection, remain appropriate and necessary. Conversely, for smaller nodes where detection is challenging and prophylactic dissection has not demonstrated population‑level benefit[9,10], current data favor a conservative or selective approach. The optimal management, whether observation, systemic therapy (e.g., immunotherapy, neoadjuvant or adjuvant regimens), or salvage intervention, remains context-dependent and is the subject of ongoing investigation. Our study represents an initial step toward refining this paradigm by improving preoperative risk stratification, thereby informing candidacy for selective cervical lymphadenectomy in high-risk patients. The overarching goal of our research is twofold: to offer a cost-effective approach for accurate detection in cases where current imaging already performs well, and, more importantly, to explore personalized strategies for detection and treatment in patients for whom conventional methods are insufficient. By demonstrating the feasibility of machine learning–based ultrasonography as a non-invasive tool, we hope to inspire further innovations that advance precision care for patients with esophageal cancer. Achieving this will require continued efforts to optimize models, expand datasets, and integrate multimodal information. In conclusion, we thank the authors for their constructive feedback, which underscores the importance of addressing these limitations and motivates us to pursue more impactful goals. We remain committed to refining our approach and fostering collaboration to help predictive modeling evolve into a clinically reliable and meaningful tool for guiding treatment decisions in esophageal cancer.