Novel insights into predicting the presence of micropapillary and solid components in stage IA lung adenocarcinoma using machine learning models of modifiable risk factors

Lung adenocarcinoma (LUAC) patients with micropapillary (MP) and/or solid (S) generally demonstrate a poorer survival prognosis. In the diagnosis and treatment of stage IA LUAC, precisely establishing personalized treatment strategies for patients is crucial for both clinical practice and scientific investigation. Our study aims to develop a novel prediction model based on machine learning (ML) to predict the probability of MP/S patterns in stage IA LUAC patients. Our retrospective analysis was conducted on 1,933 patients diagnosed with stage IA LUAC <i>via</i> postoperative pathological staging, focusing on evaluating MP/S pattern presence. MP/S-positive patients were matched with negative patients at a 1:2 ratio. Univariate logistic regression and Lasso regression were used to select variables with independent prognostic significance. The performance of the traditional logistic regression model was compared with nine ML models based on the identification and calibration. Nodule type, spiculation, Carcinoembryonic antigen level, maximum solid component diameter, median CT value, and CT value range were identified as independent influencing factors for predicting MP/S patterns. The K-Nearest Neighbors (KNN) model performed best among all ten models. The internal validation indicated an area under the curve (AUC) of 0.790, a Brier score of 0.167, and a Hosmer-Lemeshow (HL) test P value of 0.817, while external validation yielded an AUC of 0.790, a Brier score of 0.167, and a HL test P value of 0.120. Shapley additive explanation analysis revealed “nodule type” could alter the predicted probability of MP/S component presence by 13.6%, establishing it as a significant factor. An interpretable KNN model was successfully developed to predict the presence of MP/S components in stage IA LUAC patients, demonstrating superior predictive performance. Accurate evaluation of relevant tumor characteristics possesses substantial clinical significance, as it enables guidance on the optimization of surgical approaches to enhance patient prognosis.

伴有微乳头状（MP）和/或实性（S）成分的肺腺癌（LUAC）患者通常预后较差。在IA期肺腺癌的诊疗中，为患者精准制定个体化治疗策略，无论对临床实践还是科学研究均至关重要。本研究旨在构建一种基于机器学习（ML）的新型预测模型，以预测IA期肺腺癌患者的MP/S成分发生概率。本回顾性分析纳入1933例经术后病理分期确诊的IA期肺腺癌患者，重点评估MP/S成分的存在情况。以1:2的比例将MP/S阳性患者与阴性患者进行匹配。采用单因素logistic回归（univariate logistic regression）与Lasso回归（Lasso regression）筛选具有独立预后意义的变量。通过识别与校准，比较传统logistic回归模型与9种机器学习模型的性能。最终确定结节类型、毛刺征、癌胚抗原（Carcinoembryonic antigen）水平、最大实性成分直径、CT中位值及CT值范围为预测MP/S成分的独立影响因素。在全部10种模型中，K近邻（KNN）模型表现最优。内部验证结果显示，曲线下面积（AUC）为0.790，Brier评分为0.167，Hosmer-Lemeshow（HL）检验P值为0.817；外部验证则得到AUC为0.790，Brier评分为0.167，HL检验P值为0.120。夏普利可加解释（Shapley additive explanation）分析显示，“结节类型”可使MP/S成分存在的预测概率改变13.6%，证实其为关键影响因素。本研究成功构建了可解释性K近邻模型，用于预测IA期肺腺癌患者的MP/S成分存在情况，该模型展现出优异的预测性能。精准评估相关肿瘤特征具有重要临床价值，可指导优化手术方案以改善患者预后。