Acceptability and accuracy of artificial intelligence–assisted sponge cytology for screening of esophageal squamous cell carcinoma and adenocarcinoma of the esophagogastric junction: a multi-center cohort study

<h2>Abstract</h2><h3>Background</h3> Esophageal squamous cell carcinoma (ESCC) and adenocarcinoma of the esophagogastric junction (AEJ) present significant health challenges in China, often diagnosed at advanced stages with poor prognoses. Screening is the pivotal strategy to relieve the burden of ESCC and AEJ in high-risk areas. Even though endoscopy has proven effective in the early detection of both cancers in high-prevalence regions, its invasiveness and resource-intensiveness make it impractical for large-scale screening. Therefore, developing a less invasive and readily accessible method with good diagnostic accuracy to identify high-risk individuals before endoscopy is urgently needed. We aim to evaluate the acceptability and accuracy of artificial intelligence (AI)-assisted sponge cytology tests using a novel cell collection device for ESCC and AEJ screening in Chinese high-risk regions. <h3>Methods</h3> Participants aged 50 years or older were recruited in five high-risk regions of ESCC and AEJ. Cells from esophagus and esophagogastric junction were collected using a novel and minimally invasive capsule sponge, and cytology slides were scanned by a trained AI system. The qualitative outcomes (indicating the location of abnormal cells) and quantitative outcomes (counts of total scanned cells, potentially abnormal cells and 105 cytological features) were reported. Participants scored acceptability immediately following the procedure on a scale of 0 (least) to 10 (most acceptable). Endoscopy was performed subsequently with biopsy as needed. Feature selection was performed using Boruta algorithm. Lasso logistic regression model was developed to predict a composite outcome of high-grade lesions (ESCC, AEJ and high-grade intraepithelial neoplasia), with cytological features and epidemiological features as the predictive features. Model performance was primarily measured with the area under the receiver operating characteristic curve (AUC). Internal validation of the prediction models was performed using the 1000-bootstrap resample. <h3>Findings</h3> A total of 1852 participants were enrolled and completed the study procedure. No serious adverse events were documented during the cell collection process, and acceptability scores were 10 (72.1%), 9 (19.8%), 8 (3.8%), 7 (1.5%) and 6 (0.9%). 30 (1.6%) participants were diagnosed with high-grade lesions confirmed by endoscopic biopsy. The lasso logistic model achieved an AUC of 0.902 (95%<i>CI</i>: 0.851, 0.952) for detecting high-grade lesions, outperforming that of the cytological diagnosis and the sole use of abnormal cell counts. Internal validation of the model by bootstrap analysis was used, and the mean AUC of the model was 0.9 (95%<i>CI</i>: 0.845, 0.944). <h3>Interpretation</h3> We demonstrate the safety and acceptability of AI-assisted sponge cytology in high-risk regions, with high accuracy for detecting ESCC, AEJ and their precursor lesions. Our results pave the way for innovative etiology and early-detection research.