CT-based multiple instance and ensemble learning for lymph node metastasis prediction in esophageal squamous cell carcinoma: a multicentre, retrospective study

Accurately predicting lymph node metastasis (LNM) in esophageal squamous cell carcinoma (ESCC) is crucial for planning patient treatments. However, this task remains challenging, complicating treatment decision-making; this is particularly concerning for patients classified as pN0 owing to insufficient lymph node dissection (< 15 lymph nodes), as the effectiveness of postoperative adjuvant therapy (POAT) for these patients remains controversial. Therefore, we aimed to develop a CT-based predictive model to improve LNM detection in ESCC patients and identify pN0 patients who can benefit from POAT. We retrospectively enrolled 974 ESCC patients who underwent radical esophagectomy with adequate lymph node dissection (≥ 15 lymph nodes), dividing them into training (432 patients), internal validation (185 patients), and external validation cohorts (357 patients). To predict LNM, we developed a Stacking model using multiple instance and ensemble learning, leveraging the visible lymph nodes, primary tumor features, and clinical characteristics of each patient. Additionally, we separately enrolled 386 pN0 patients who underwent insufficient lymph node dissection and classified them into low-risk and high-risk groups on the basis of the optimal cut-off value from the Stacking model. Kaplan‒Meier and Cox models were used to assess the impact of POAT on patient survival. The Stacking model achieved area under the curve (AUC) values of 0.883, 0.834, and 0.819 in the training, internal validation, and external validation cohorts, respectively, outperforming traditional models based separately on tumor features, clinical characteristics, or the features of the largest lymph node. pN0 patients identified by the Stacking model as high risk had worse overall survival than low-risk patients did. POAT provided survival benefits for the high-risk patients but had no significant effect on the survival of low-risk patients. Our Stacking model achieved excellent LNM prediction in ESCC patients and holds promise for guiding personalized treatment strategies, particularly for pN0 patients with insufficient lymph node dissection.