数字肺音人工智能分析与临床医师判断的一致性研究  

Consistency study on digital lung sounds between artificial intelligence analysis and clinician′s judgment

作　　者：刘国梁[1] 张玲[2] 卢清君[3] 许文兵 张运剑[5] 刘前桂 王英田 Liu Guoliang;Zhang Ling;Lu Qingjun;Xu Wenbing;Zhang Yunjian;Liu Qiangui;Wang Yingtian(Department of Pulmon and Critical Care Medicine,Center of Respiratory Medicine,China-Japan Friendship Hospital National Center for Respiratory Medicine Institute of Respiratory Medicine,Chinese Academy of Medical Sciences National Clinical Research Center for Respiratory Diseases,Beijing 100029,China;Department of Respiratory Medicine,Hebei Chest Hospital,Shijiazhuang 050047,China;National Center for Telemedicine and Internet Medicine,China-Japan Friendship Hospital,Beijing 100029,China;Department of Respiratory and Critical Care Medicine,Peking Union Hospital,Chinese Academy of Medical Sciences,Beijing 100005,China;Department of Respiratory and Critical Care,Beijing Jishuitan Hospital,Beijing 100035,China;Department of Respiratory Medicine,Beijing Geriatrics Hospital,Beijing 102488,China;Department of Respiratory Medicine,Beijing Airport Hospital,Beijing 101318,China)

机构地区：[1]中日友好医院呼吸中心呼吸与危重症医学科国家呼吸医学中心中国医学科学院呼吸病学研究院国家呼吸疾病临床医学研究中心,北京100029 [2]河北省胸科医院呼吸内科,石家庄050047 [3]中日友好医院国家远程医疗与互联网医学中心,北京100029 [4]中国医学科学院北京协和医院呼吸与危重症医学科,北京100005 [5]北京积水潭医院呼吸与危重症医学科,北京100035 [6]北京老年病医院呼吸内科,北京102488 [7]北京空港医院呼吸内科,北京101318

出　　处：《国际呼吸杂志》2022年第3期180-186,共7页International Journal of Respiration

基　　金：国家重点研发计划(2016YFC1304602)。

摘　　要：目的评估数字肺音人工智能分析与临床医师判断的一致性。方法本研究为诊断性实验。采用非随机抽样方法应用2737条选自《国家远程医疗与互联网医学中心数字听诊平台》的真实的临床数字肺音数据,应用Luntech®数字听诊人工智能分析结果判断,根据肺音的粗糙程度、是否有啰音、啰音的强度、是否有干啰音和是否有湿啰音5个方面进行分类整理后建立人工智能分析结果数据库。同时,制定统一分类定义,临床医师对数字肺音进行以上5个方面判断,对比分析数字听诊人工智能和医师听诊的肺音分析结果的一致性。应用描述性方法对肺音图变化特征进行整体描述;以医师判断结果为"金标准",应用混淆矩阵计算分类准确度、召回率、虚警率、精确度、kappa值评价两者的一致性。结果Luntech®数字听诊人工智能分析对于是否为异常肺音、有无呼吸音粗糙、有无啰音、有无干啰音和有无湿啰音的判断准确度分别为98.39%、95.14%、96.60%、97.84%和96.97%,召回率分别为96.60%、88.34%、91.65%、92.70%和86.68%,虚警率分别为3.48%、2.43%、1.03%、0.92%和0.63%,精确度分别为97.00%、92.86%、97.71%、97.08%和96.98%;并且一致性良好(kappa值分别为0.931、0.873、0.921、0.941和0.898);对于啰音强度的判断结果具有较好的一致性(湿啰音kappa值=0.790,干啰音kappa值=0.889)。肺音图具有明显形态特征,人工智能肺音分析指数及肺音图可敏感反应肺音性质的变化。结论数字肺音人工智能分析及肺音图与临床医师判断有较好一致性。Objective To evaluate the consistency on digital lung sounds between artificial intelligence(AI)analysis and clinician judgment.Methods This is a diagnostic experimental study.By non-random sampling method,totally 2737 pieces of real clinical digital lung sound data were enrolled from National Telemedicine and Internet Medical Center Digital Auscultation System.Luntech®digital lung sound artificial intelligence analysis system was used to analyze this results and judgment.The AI analysis results database is established based on roughness of lung sound,whether there is rale,the intensity of rale,whether there is dry rale and wet rale.Meanwhile,a unified classification definition was formulated,clinicians judged the digital lung sound in the above five aspects,aiming to compare the consistency between AI analysis of digital auscultation and auscultation results on lung sounds form clinicians,the change characteristics of phonopneumography as a whole were performed by descriptive methods.Clinician′s judgment results as"gold standard",the consistency of the two were evaluated by using confusion matrix calculating the classification accuracy,recall rate,false alarm rate,accuracy and kappa value.Results The application of digital auscultation AI analysis(Luntech®)judged whether it was abnormal lung sound,whether there was rough breath sound,whether there was rale,whether there was dry rale and whether there was wet rale,the results were as follows:the accuracy were 98.39%,95.14%,96.60%,97.84%and 96.97%respectively;the recall rates were 96.60%,88.34%,91.65%,92.70%and 86.68%respectively;the false alarm rates were 3.48%,2.43%,1.03%,0.92%and 0.63%respectively,and the accuracy was 97.00%,92.86%,97.71%,97.08%and 96.98%respectively;and good consistency(the kappa values were 0.931,0.873,0.921,0.941 and 0.898 respectively);the judgment results of rales intensity had a good consistency(wet rale kappa value=0.790,dry rale kappa value=0.889).Morphological characteristics were obvious in phonopneumography,lung sound index of AI

关 键 词：人工智能 数字肺音 肺音图 听诊 

分 类 号：R443.4[医药卫生—诊断学]