摘要目的:建立血液分析检验结果人工智能（AI）审核规则。方法:纳入2019年8月1日至31日吉林大学第一医院住院患者血液分析数据18 474份，作为AI审核规则训练组，采集其对应的患者年龄、镜检结果、临床诊断信息，以血液分析报告参数、研究参数、报警信息等92个实验室参数作为AI审核规则的候选条件；依据镜检的人工审核作为审核标准，标注是否通过或拦截；采用决策树算法，通过高强度、多轮次及五折交叉验证，初步建立AI审核规则，通过设置重要病例必中，以优化AI审核规则。采用卡方检验比较AI审核规则与自动审核规则的假阴性率、精确率、召回率、F1分数、通过率等指标，评价AI审核规则性能。收集2023年11月1日至31日吉林大学第一医院检验科12 475份住院患者血液分析数据作为验证组，将AI审核规则模拟用于结果审核，分析AI审核规则的性能指标，验证AI审核规则的性能。结果:AI审核规则由15项规则、17条参数构成，能够分辨计数和形态异常。与自动审核规则相比，AI审核规则的训练组真阳性率、假阳性率、真阴性率、假阴性率、通过率、正确率、精确率、召回率、F1分数分别为22.7%、1.6%、74.5%、1.3%、75.7%、97.2%、93.5%、94.7%、94.1，均优于自动审核规则，差异具有统计学意义（ P<0.001），且无重要病例漏检。验证组真阳性率、假阳性率、真阴性率、假阴性率、通过率、正确率、精确率、召回率、F1分数分别为19.2%、8.2%、70.1%、2.5%、72.6%、89.2%、70.0%、88.3%、78.1，与自动审核规则相比，假阴性率较低、假阳性率和召回率稍高，差异具有统计学意义（ P<0.001）。 结论:利用机器学习的决策树算法建立并验证的血常规AI审核规则，能较稳定地识别、拦截与提示异常结果，与自动审核相比在血液分析检验结果报告中更加简便、高效、准确。

abstractsObjective:To establish a set of artificial intelligence (AI) verification rules for blood routine analysis.Methods:Blood routine analysis data of 18 474 hospitalized patients from the First Hospital of Jilin University during August 1st to 31st, 2019, were collected as training group for establishment of the AI verification rules,and the corresponding patient age, microscopic examination results, and clinical diagnosis information were collected. 92 laboratory parameters, including blood analysis report parameters, research parameters and alarm information, were used as candidate conditions for AI audit rules; manual verification combining microscopy was considered as standard, marked whether it was passed or blocked. Using decision tree algorithm, AI audit rules are initially established through high-intensity, multi-round and five-fold cross-validation and AI verification rules were optimized by setting important mandatory cases. The performance of AI verification rules was evaluated by comparing the false negative rate, precision rate, recall rate, F1 score, and pass rate with that of the current autoverification rules using Chi-square test. Another cohort of blood routine analysis data of 12 475 hospitalized patients in the First Hospital of Jilin University during November 1sr to 31st, 2023, were collected as validation group for validation of AI verification rules, which underwent simulated verification via the preliminary AI rules, thus performance of AI rules were analyzed by the above indicators. Results:AI verification rules consist of 15 rules and 17 parameters and do distinguish numeric and morphological abnormalities. Compared with auto-verification rules, the true positive rate, the false positive rate, the true negative rate, the false negative rate, the pass rate, the accuracy, the precision rate, the recall rate and F1 score of AI rules in training group were 22.7%, 1.6%, 74.5%, 1.3%, 75.7%, 97.2%, 93.5%, 94.7%, 94.1, respectively.All of them were better than auto-verification rules, and the difference was statistically significant ( P<0.001), and with no important case missed. In validation group, the true positive rate, the false positive rate, the true negative rate, the false negative rate, the pass rate, the accuracy, the precision rate, the recall rate and F1 score were 19.2%, 8.2%, 70.1%, 2.5%, 72.6%, 89.2%, 70.0%, 88.3%, 78.1, respectively, Compared with the auto-verification rules, The false negative rate was lower, the false positive rate and the recall rate were slightly higher, and the difference was statistically significant ( P<0.001). Conclusion:A set of the AI verification rules are established and verified by using decision tree algorithm of machine learning, which can identify, intercept and prompt abnormal results stably, and is moresimple, highly efficient and more accurate in the report of blood analysis test results compared with auto-vefication.