人工智能语音分析系统在帕金森病诊断中的一项探索性临床研究

Application of artificial intelligence voice analysis system in clinical diagnosis of PD

目的通过人工智能技术构建汉语的帕金森病(Parkinson’s disease,PD)语音辅助诊断系统,并应用临床上新收集的数据验证该系统在PD患者尤其是在轻症PD患者中的诊断效能。方法连续收集2018年1~12月解放军总医院PD专病门诊临床确诊的PD患者100例(PD1组),并以Hoehn-Yahr(H-Y)分级2.5级为界将PD1组分为轻症72例,中重症28例,另选择同期健康对照者100例(对照1组)。连续收集2019年1~10月解放军总医院PD专病门诊临床确诊的PD患者76例(PD2组),并以H-Y 2.5级为界分为轻症54例,中重症22例;另选择同期健康体检者77例(对照2组),所有研究对象分别采集持续5 s的元音[a]、[o]、[i]语音数据,并提取每种语音的特征参数及人工智能机器学习训练,最终优化建立PD语音识别模型,并对该模型的诊断效果进行评价。结果 PD1组轻症与中重症患者年龄和性别比较,差异无统计学意义(P>0.05)。在判断PD2组与对照2组发音[o]构建模型的AUC明显高于[a]和[i]构建模型的AUC(P<0.01)。在判断PD2组轻症与对照2组发音[o]构建模型的AUC明显高于发音[a]和[i]构建模型的AUC(P<0.01)。PD2组与对照2组发音[o]的准确性明显高于[a]和[i](86.3%vs 74.5%、75.2%,P<0.01);PD2组轻症与对照2组发音[o]的准确性明显高于[a]和[i](86.2%vs 71.5%、73.9%,P<0.01)。PD2组与对照2组发音[o]的敏感性明显高于[i](90.8%vs 75.0%,P<0.05);PD2组轻症与对照2组发音[o]的敏感性明显高于[i](92.5%vs 71.7%,P<0.05)。结论汉语人群中构建的PD人工智能语音辅助诊断系统,通过临床验证表明该系统有着很高的敏感性及较高的准确性,有助于在初级医疗机构全科医师的日常工作中开展PD的筛查和早期诊断。

Objective To verify the role of artificial intelligence voice analysis system in clinical diagnosis of PD.Methods One hundred PD patients admitted to our center from January 2018 to December 2018 were divided into mild PD group 1(n=72) and moderate-severe PD group 1(n=28) according to their H-Y classificaltion with 100 healthy subjects served as a control group 1.In addition,76 PD patients admitted to our center from January 2019 to October 2019 were divided into mild PD group 2(n=54) and moderate-severe PD group 2(n=22) according to their H-Y classification with 77 healthy subjects served as a control group 2.A PD voice-assisted diagnostic model was established and optimized by recording the data of vowel phonation [a],[o],[i]for 5 seconds and extracting the characteristics of each phonation and training the machine learning model.The value of machine learning model in diagnosis of PD was assessed.Results No significant difference was detected in age and sex between mild PD group 1 and moderate-severe PD group 1(P>0.05).The AUC for machine learning model of vowel phonation [o] was significantly larger than that for machine learning model in PD group 2 and control group 2(P<0.01).The accuracy of vowel phonation [o] was significantly higher than that of vowel phonation [a] and [i] in mild PD group 2 and control group 2(86.3% vs 74.5% and 75.2%,P<0.01).The sensitivity of vowel phonation [o] was significantly higher than that of vowel phonation [i] in PD group 2 and control group 2(90.8% vs 75.0%,P<0.05) than in mild PD group 2 and control group 2(92.5% vs 71.7%,P<0.05).Conclusion The sensitivity and accuracy of artificial intelligence voice analysis system are rather high in clinical diagnosis of PD.The artificial intelligence voice analysis system can thus contribute to the early screening and diagnosis of PD in primary hospitals.