论文部分内容阅读
目的:建立人工智能（AI）模型，评价AI辅助多参数流式细胞术（MFC）诊断儿童急性B淋巴细胞白血病（B-ALL）微小残留病（MRD）的可行性。方法:以北京大学第一医院门诊B-ALL MRD随访病例227例标本进行AI应用探索性研究，以多维空间密度分布的非监督学习分群聚类算法为基础，对MFC数据进行聚类分析；同时参考人工逻辑设门思路和细胞的抗原表达水平，应用决策树和随机森林等监督学习算法判别细胞分类，并将AI分析B-ALL MRD结果以二维图、降维t-分布领域嵌入算法（t-SNE）和热图进行可视化呈现。通过与人工分析比较，评价AI辅助MFC诊断B-ALL MRD的性能和一致性。结果:AI分析可快速检测出骨髓中B细胞系各阶段淋巴细胞和B-ALL MRD细胞的数量和百分比；且在多次重复检测时各参数差值的变异系数（n CV）均为0。AI分析速度较传统人工分析提高4~5倍。与人工分析比较，AI分析B-ALL MRD细胞的查全率（敏感度）和阴性预测值为100%，准确度为82.44%。经人工审核后的AI分析结果再与人工分析比较，去除了假阳性，其查准率（阳性预测值）和准确度均可达100%。人工审核后的AI分析B细胞系各阶段淋巴细胞和B-ALL MRD细胞的百分比与人工分析结果差异无统计学意义（n P均>0.05）。n 结论:本研究初步建立的AI模型数据分析重复性好、分析速度快，具有较高的查全率（敏感度）和阴性预测值，可用于基于本研究的8色抗体组合方案辅助MFC诊断儿童B-ALL MRD。“,”Objective:To develop an artificial intelligence (AI) model to evaluate the feasibility of AI-assisted multiparameter flow cytometry (MFC) diagnosis of the childhood B-ALL minimal/measurable residual disease (MRD).Methods:227 cases of B-ALL MRD in the Peking University First Hospital were selected for our AI application of exploratory study. MFC data were automatically analyzed by unsupervised learning clustering algorithm with multi-dimensional spatial density distribution. Then, a boosted random forest supervised learning fitting model was developed based on manual gating strategy and antigen expression level, followed by automatical immunophenotyping for each cell population. Results of AI analysis were visualized by 2D-scatterplots, dimension-reduction t-distributed stochastic neighbor embedding (t-SNE) and heatmaps. The performance and consistency of AI-assisted diagnosis for B-ALL MRD were assessed compared with those of manual analysis.Results:Our AI model could rapidly detect the number and proportion of B-cells at each stage and B-ALL MRD cells in bone marrow, also the coefficient variation (n CV) of the difference of each parameter is 0 in the repeatability testing. The speed of AI analysis is 4 to 5 times faster than that of manual analysis. By comparison with manual analysis, the recall rate (sensitivity) and negative predictive value (NPV) of AI analysis for B-ALL MRD cells were 100%, and its accuracy was 82.44%. After manual confirmation by removing false positive results, the precision (positive predictive value) and accuracy of AI analysis for B-ALL MRD cells were up to 100%. In addition, no significant statistical difference was found in the proportion of B-cells at each stage and B-ALL MRD cells between AI analysis and manual analysis (n P>0.05).n Conclusion:We successfully established an AI model, which has high repeatability and speed with ideal recall rate (sensitivity) and NPV. Our AI model can be used as an assistant MFC diagnosis for childhood B-ALL MRD based on this study protocol with 8-color antibodies.