Abstract: Background　As a new auxiliary tool for diagnosis and analysis，artificial intelligence technology has been gradually applied to imaging medicine，medical big data analysis，diagnostics and medical expense budget. The prediction model for spontaneous ureteral calculus passage developed using artificial neural network by our hospital（First Affiliated Hospital，School of Medicine，Shihezi University）has proven to have satisfactory performance，but has not been tested by recognized multicenter clinical trials. Objective　To evaluate the accuracy and applicability of the prediction model of spontaneous ureteral calculus passage using artificial neural network developed by our hospital using a multicenter clinical trial，to provide evidence facilitating its promotion. Methods　From September 2017 to March 2020，we enrolled 1 620 urological patients with ureteral calculus who received conservative stone treatment in six hospitals（First Division Hospital，Fourth Division Hospital，Seventh Division Hospital，Ninth Division Hospital，and 13th Division Red Star Hospital of the Xinjiang Production and Construction Corps，and First Affiliated Hospital，School of Medicine，Shihezi University）. Nine parameters were collected，containing age，pain degree，calculus size，location，leucocyte count，neutrophil count，neutrophil percentage，lymphocyte count and C-reactive protein level，which were applied as predictive parameters included in the model developed using artificial neural network with the result of previous research for reference for predicting spontaneous ureteral calculus passage in patients with at least a small （≤1 cm）calculus，and large（>1 cm）calculus，and the prediction results were compared with actual conditions in a 4-week follow-up to estimate the accuracy of the model. Results　During the follow-up period，992 cases had spontaneous ureteral calculus passage，accounting for 61.23% of the total cases，who included 763 （68.06%）from the small calculus group（n=1 121），and 229 （45.89%）from the large calculus group（n=499）. The distribution of pain level，mean calculus diameter，leucocyte count，and C-reactive protein level differed obviously between those with and without spontaneous ureteral calculus passage（P<0.05），but distribution of calculus location，mean age，neutrophil percentage，and lymphocyte count did not（P>0.05）. For predicting the overall spontaneous ureteral calculus passage，the sensitivity，specificity and accuracy of the model were 87.10%，85.99% and 86.67%，respectively. For patients with at least one small calculus，the sensitivity，specificity and accuracy of the model were 84.28%，84.44% and 84.37，respectively. For patients with at least one large calculus，the sensitivity，specificity and accuracy of the model were 87.94%，87.15% and 87.69%，respectively. Conclusion　The accuracy of the model in predicting spontaneous passage of ureteral calculi，especially the calculus larger than 1 cm in diameter，has been recognized in our multicenter clinical trial.

Key words: Ureteral calculi, Artificial intelligence, Prediction model, Multicenter studies as topic

摘要： 背景　人工智能技术作为一种新型的诊断和分析工具已逐渐应用于影像技术、医学大数据分析、医学诊断和医疗保健预算等各个方面，目前有学者应用人工智能模型（AIM）建立了输尿管结石排出的预测模型并取得了较为满意的预测结果，但尚无获得临床一致认可的多中心验证报道。目的　评估输尿管结石自然排出人工智能预测模型在多中心临床应用中的普适性及准确性，并进一步推广运用。方法　选择2017年9月—2020年3月在新疆生产建设兵团一师医院、四师医院、七师医院、九师医院、十三师红星医院及石河子大学医学院第一附属医院泌尿外科收治的1 620例采取保守排石治疗的输尿管结石患者为研究对象，采集纳入患者的年龄、疼痛程度、结石直径、结石位置（上段、中段、下段）、白细胞计数、中性粒细胞计数、淋巴细胞计数、中性粒细胞分数、C反应蛋白（CRP）9项指标。参考前期研究结果，建立人工智能预测模型，并判断结石能否自然排出指导患者制定治疗决策。通过为期4周的随访，根据随访结果综合分析AIM的预测效能。根据结石大小进一步分层，将直径>1 cm输尿管结石归类为大结石组，≤1 cm输尿管结石归类为小结石组，进一步评估AIM对较大输尿管结石的预测效能。结果　随访期间992例患者排出结石，排出率为61.23%。两组年龄、结石位置、中心粒细胞计数、淋巴细胞计数、中性粒细胞分数比较，差异均无统计学意义（P>0.05）。两组疼痛程度、结石直径、白细胞计数、CRP比较，差异均有统计学意义（P<0.05）。AIM预测输尿管结石排出的灵敏度、特异度及准确率分别为87.10%、85.99%和86.67%；小结石组患者1 121例，763例（68.06%）排出结石，预测灵敏度、特异度和准确率分别为87.94%、87.15%和87.69%；大结石组499例，排石率为229（45.89%），预测灵敏度、特异度和准确率分别为84.28%、84.44%和84.37%。结论　预测输尿管结石能否自然排出的AIM在多中心临床评价中得到一致性认可。对于较大输尿管结石排出的预测效能准确，具有较强泛化能力。

关键词: 输尿管结石, 人工智能, 预测模型, 多中心研究（主题）