Machine learning-based automated sponge cytology for screening of oesophageal squamous cell carcinoma and adenocarcinoma of the oesophagogastric junction: a nationwide, multicohort, prospective study

Background Oesophageal squamous cell carcinoma and adenocarcinoma of the oesophagogastric junction have a dismal prognosis, and early detection is key to reduce mortality. However, early detection depends on upper gastrointestinal endoscopy, which is not feasible to implement at a population level. We aimed to develop and validate a fully automated machine learning-based prediction tool integrating a minimally invasive sponge cytology test and epidemiological risk factors for screening of oesophageal squamous cell carcinoma and adenocarcinoma of the oesophagogastric junction before endoscopy. Methods For this multicohort prospective study, we enrolled participants aged 40-75 years undergoing upper gastrointestinal endoscopy screening at 39 tertiary or secondary hospitals in China for model training and testing, and included community-based screening participants for further validation. All participants underwent questionnaire surveys, sponge cytology testing, and endoscopy in a sequential manner. We trained machine learning models to predict a composite outcome of high-grade lesions, defined as histology-confirmed high-grade intraepithelial neoplasia and carcinoma of the oesophagus and oesophagogastric junction. The predictive features included 105 cytological and 15 epidemiological features. Model performance was primarily measured with the area under the receiver operating characteristic curve (AUROC) and average precision. The performance measures for cytologists with AI assistance was also assessed. Findings Between Jan 1, 2021, and June 30, 2022, 17 498 eligible participants were involved in model training and validation. In the testing set, the AUROC of the final model was 0 center dot 960 (95% CI 0 center dot 937 to 0 center dot 977) and the average precision was 0 center dot 482 (0 center dot 470 to 0 center dot 494). The model achieved similar performance to consensus of cytologists with AI assistance (AUROC 0 center dot 955 [95% CI 0 center dot 933 to 0 center dot 975]; p=0 center dot 749; difference 0 center dot 005, 95% CI, -0 center dot 011 to 0 center dot 020). If the model-defined moderate-risk and high-risk groups were referred for endoscopy, the sensitivity was 94 center dot 5% (95% CI 88 center dot 8 to 97 center dot 5), specificity was 91 center dot 9% (91 center dot 2 to 92 center dot 5), and the predictive positive value was 18 center dot 4% (15 center dot 6 to 21 center dot 6), and 90 center dot 3% of endoscopies could be avoided. Further validation in community-based screening showed that the AUROC of the model was 0 center dot 964 (95% CI 0 center dot 920 to 0 center dot 990), and 92 center dot 8% of endoscopies could be avoided after risk stratification. Interpretation We developed a prediction tool with favourable performance for screening of oesophageal squamous cell carcinoma and adenocarcinoma of the oesophagogastric junction. This approach could prevent the need for endoscopy screening in many low-risk individuals and ensure resource optimisation by prioritising high-risk individuals. Funding Science and Technology Commission of Shanghai Municipality. Copyright (c) 2023 Elsevier Ltd. All rights reserved.