人工智能食管细胞学风险预测模型在食管癌前病变中的构建和验证

Development and validation of an artificial intelligence‑assisted esophageal cytological risk prediction model for detecting esophageal precancerous lesions

摘要:

目的 利用人工智能辅助的食管细胞学构建筛查食管癌前病变的风险预测模型，并验证其诊断准确性。方法 本研究为食管癌筛查试验（esophageal cancer screening trial， EAST）队列研究数据的二次分析，纳入2021年1月1日至2022年6月30日位于我国食管鳞状细胞癌高发地区的39个三级或二级医院以及5个社区接受上消化道内镜检查和人工智能食管细胞学筛查的受试者共17 294例。14 415例在医院筛查的受试者构成医院机会性筛查队列，设定为训练集，构建基于轻量梯度提升机（light⁃gradient boosting machine， LightGBM）机器学习算法的人工智能食管细胞学风险预测模型（简称LightGBM模型）。在5个社区卫生服务中心进行筛查的受试者（n=2 879）构成社区筛查队列，设定为验证集，以内镜活检的病理结果为金标准，评价LightGBM模型在社区筛查队列中筛查食管癌前病变的诊断效能。结果 在机会性筛查队列中训练LightGBM模型，该模型预测癌前病变的受试者工作特征（receiver operator characteristic， ROC）曲线下面积为0.93（95%CI：0.91~0.95）。约登指数最大时，ROC曲线的cutoff值为0.08。以风险预测评分>0.08分作为筛查癌前病变的标准，LightGBM 模型在筛查癌前病变时的灵敏度和特异度分别为91.0%（95%CI：86.9%~95.1%）和86.2%（95%CI：85.7%~86.8%）。用社区筛查队列验证LightGBM模型，该模型在社区筛查队列中筛查癌前病变的灵敏度和特异度分别为95.2%（20/21）和87.5%（2 502/2 858），准确率为87.6%（2 522/2 879）。结论 人工智能食管细胞学风险预测模型筛查食管癌前病变灵敏度和特异度较高，在食管癌筛查中有推广应用价值。

Abstract:

Objective Artificial intelligence-assisted esophageal cytology was used to develop and validate a risk prediction model for screening esophageal precancerous lesions. Methods This study was a secondary analysis of data from the esophageal cancer screening trial (EAST). A total of 17 294 subjects were included who underwent upper gastrointestinal endoscopy and artificial intelligence-assisted esophageal cytology screening at 39 tertiary or secondary hospitals and 5 community service centers in areas with high incidence of esophageal squamous cell carcinoma in China from January 1, 2021 to June 30, 2022. Subjects (n=14 415) screened in the hospital constituted the hospital opportunistic screening cohort, which served as the training set. An artificial intelligence-assisted esophageal cytological risk prediction model (LightGBM model for short) was developed based on light-gradient boosting machine (LightGBM) machine learning algorithm. Subjects undergoing screening at 5 community health service centers (n=2 879) constituted a community screening cohort, which served as a validation set. The diagnostic efficacy of LightGBM model for esophageal precancerous lesions in the community screening cohort was evaluated by using pathological results of endoscopic biopsy as the golden standard. Results The LightGBM model, trained in the opportunistic screening cohort, exhibited an area under the receiver operator characteristic (ROC) curve of 0.93 (95%CI: 0.91-0.95) for detecting precancerous lesions. The cutoff value of the ROC curve was determined as 0.08 based on the maximum Youden index. The sensitivity and specificity of LightGBM model were 91.0% (95%CI: 86.9%-95.1%) and 86.2% (95%CI: 85.7%-86.8%), respectively, when the risk prediction score was >0.08 as the screening criterion for precancerous lesions. The sensitivity, specificity, and accuracy of LightGBM model for precancerous lesions in the community screening cohort were 95.2% (20/21), 87.5% (2 502/2 858), and 87.6% (2 522/2 879), respectively. Conclusion The artificial intelligence-assisted esophageal cytology risk prediction model showcased remarkable sensitivity and specificity in screening for esophageal precancerous lesions, underscoring its potential for widespread adoption and application in esophageal cancer screening.