[Evaluation of the application value of seven tumor-associated autoantibodies in non-small cell lung cancer based on machine learning algorithms].

Objective: Based on the diagnostic model established and validated by the machine learning algorithm, to investigate the value of seven tumor-associated autoantibodies (TAABs), namely anti-p53, PGP9.5, SOX2, GAGE7, GBU4-5, MAGEA1 and CAGE antibodies in the diagnosis of non-small cell lung cancer (NSCLC) and to differentiate between NSCLC and benign lung nodules. Methods: This was a retrospective study of clinical cases. Model building queue: a total of 227 primary patients who underwent radical lung cancer surgery in the Department of Thoracic Surgery, Shengjing Hospital of China Medical University, from November 2018 to June 2021 were collected as the NSCLC group, and 120 cases of benign lung nodules, 122 cases of pneumonia and 120 healthy individuals were selected as the control groups. External validation queue: a total of 100 primary patients who underwent radical lung cancer surgery in the Department of Thoracic Surgery, Shengjing Hospital of China Medical University, from May 2022 to December 2022 were collected as the NSCLC group, and 36 cases of benign lung nodules, 32 cases of pneumonia and 44 healthy individuals were selected as the control groups. In addition, NSCLC was divided into early (stage 0-ⅠB) and mid-to-late (stage ⅡA-ⅢB) subgroups. The levels of 7-TAABs were detected by enzyme immunoassay, and serum concentrations of CEA and CYFRA21-1 were detected by electrochemiluminescence. Four machine learning algorithms, XGBoost, Lasso logistic regression, Naïve Bayes, and Support Vector Machine are used to establish classification models. And the best performance model was chosen based on evaluation metrics and a multi-indicator combination model was established. In addition, an online risk evaluation tool was generated to assist clinical applications. Results: Except for p53, the levels of rest six TAABs, CEA and CYFRA21-1 were significantly higher in the NSCLC group (P<0.05). Serum levels of anti-SOX2 [1.50 (0.60, 10.85) U/ml vs. 0.8 (0.20, 2.10) U/ml, Z=2.630, P<0.05] and MAGEA1 antibodies [0.20 (0.10, 0.43) U/ml vs. 0.10 (0.10, 0.20) U/ml, Z=2.289, P<0.05], CEA [3.13 (2.12, 5.64) ng/ml vs. 2.11 (1.25, 3.09) ng/ml, Z=3.970, P<0.05] and CYFRA21-1 [4.31(2.37, 7.14) ng/ml vs. 2.53(1.92, 3.48) ng/ml, Z=3.959, P<0.05] were significantly higher in patients with mid-to late-stage NSCLC than in early stages. XGBoost model was used to establish a multi-indicator combined detection model (after removing p53). 6-TAABs combined with CYFRA21-1 was the best combination model for the diagnosis of NSCLC and early NSCLC. The optimal diagnostic thresholds were 0.410, 0.701 and 0.744, and the AUC was 0.828, 0.757 and 0.741, respectively (NSCLC vs. control, NSCLC vs. benign lung nodules, early NSCLC vs. benign lung nodules) in model building queue, and the AUC was 0.760, 0.710 and 0.660, respectively (NSCLC vs. control, NSCLC vs. benign lung nodules, early NSCLC vs. benign lung nodules) in external validation queue. Conclusion: In the diagnosis of NSCLC, 6-TAABs is superior to that of traditional tumor markers CEA and CYFRA21-1, and can compensate for the shortcomings of traditional tumor markers. For the differential diagnosis of NSCLC and benign lung nodule, "6-TAABs+CYFRA21-1" is the most cost-effective combination, and plays an important role in prevention and screening for early lung cancer.目的： 以机器学习算法建立并验证的诊断模型为依据，探讨7种肿瘤相关自身抗体（TAABs），即抗p53、PGP9.5、SOX2、GAGE7、GBU4-5、MAGEA1和CAGE抗体，在非小细胞肺癌（NSCLC）诊断及其与良性肺结节鉴别诊断中的应用价值。 方法： 本研究为临床病例回顾性研究。模型建立队列来自2018年11月至2021年6月于中国医科大学附属盛京医院胸外科进行肺癌根治术的227例初治NSCLC患者为NSCLC组，同时选择良性肺结节120例、肺炎122例及健康者120名作为对照组；外部验证队列来自2022年5月至12月，中国医科大学附属盛京医院胸外科行肺癌根治术的100例初治NSCLC患者为NSCLC组，同时选择良性肺结节36例、肺炎32例及健康者44名作为对照组。将NSCLC分成早期（0~ⅠB期）与中晚期（ⅡA~ⅢB期）亚组。采用酶联免疫法检测7种TAABs，电化学发光法检测癌胚抗原（CEA）和细胞角蛋白19片段（CYFRA21-1）在各组之间的血清浓度。采用4种机器学习算法，包括极限梯度提升（XGBoost）、Lasso逻辑回归（LR）、朴素贝叶斯（NB）、以及支持向量机（SVM）分别建立多指标联合检测模型，并选择XGBoost作为最佳算法建立了针对临床应用的患者在线风险评估工具。 结果： 除抗p53抗体外，其余6种TAABs及CEA、CYFRA21-1在NSCLC中血清浓度显著升高（P<0.05）；中晚期NSCLC患者血清抗SOX2［1.50（0.60，10.85）U/ml vs.0.8（0.20，2.10）U/ml，Z=2.630，P<0.05］和MAGEA1抗体［0.20（0.10，0.43）U/ml vs. 0.10（0.10，0.20）U/ml，Z=2.289，P<0.05］及CEA［3.13（2.12，5.64）ng/ml vs. 2.11（1.25，3.09）ng/ml，Z=3.970，P<0.05］和CYFRA21-1［4.31（2.37，7.14）ng/ml vs. 2.53（1.92，3.48）ng/ml，Z=3.959，P<0.05］浓度显著高于早期。采用机器学习算法XGBoost建立多指标联合检测模型（剔除p53后），6-TAABs联合CYFRA21-1均为诊断NSCLC及NSCLC早期的最佳组合模型，诊断最佳界值分别为0.410、0.701、0.744，AUC分别为0.828、0.757、0.741（NSCLC vs. 对照组，NSCLC vs. 良性肺结节组，早期NSCLC vs. 良性肺结节组）。模型的外部验证队列的AUC分别为0.760、0.710、0.660（NSCLC vs. 对照组，NSCLC vs. 良性肺结节组，早期NSCLC vs. 良性肺结节组）。 结论： 在NSCLC诊断中，6-TAABs诊断效能优于传统肿瘤标志物CEA和CYFRA21-1；6-TAABs+CYFRA21-1检测模型为诊断NSCLC最优的模型，其可有效地辅助临床用于NSCLC及NSCLC早期与良性肺结节的鉴别诊断，在肺癌预防和早期筛查中发挥重要作用。.