질병정보를이용한딥러닝기반의자동진단건강검진알고리즘

Abstract

임상환경에서진료시간의대부분은환자의증상을듣고,추가증상을이끌어내는데사용된다.이러한활동을병력청취라고하는데,환자에게진행하는병력청취는정해진표준이없기때문에의사에따라서질문리스트가변경되기도하며진단결과에차이를발생시킨다.잘못된진단은환자를위험에빠뜨리고비용적인문제가발생할수있다.따라서의사가올바른진단을내릴수있도록환자와문진을진행하는것은중요하다. 최근진단과관련하여딥러닝기법을활용한연구가크게증가하였지만연구에주로사용되는dataset은환자의혈압이나혈당과같은수치정보가미리기입되어정제된상태로,환자와의문진을통해진단을이끌어나가기보다는환자의혈액과같은정보를토대로질병을예측하는연구가대다수이다.이러한연구는환자의수치정보를토대로이후의수치를예측하거나임상에서정의내려진질병진단기준에환자가부합하는지일차적으로점검하는방식이며,인공지능모델이예측한결과를설명할수없기때문에신뢰성보장이어렵다는단점이있어임상에서는활용이어려운상황이다. 실제임상에서활용되기위해서는진단이전에진행되어야하는병력청취에대한연구가필수적으로선행되어사용자의증상을수집하고,이를토대로질병을예측하여환자에게설명가능한정보를제공하여야한다.따라서본논문에서는질병과증상에대한데이터를수집하여질병과증상의연관관계를표현할수있는Symptom2Vec을제안하며,Symptom2Vec을환자의담화를분석할수있는BidirectionalEncoderRepresentationsfromTransformers(BERT)기반의AnsweringClassificationModel과결합하여실시간으로병력청취를진행하고질병을예측할수있는자동진단건강검진알고리즘을제안한다. 성능평가는Symptom2Vec과AnsweringClassificationModel을각각진행하였으며,Symptom2Vec의경우Word2Vec등과같은기존모델과비교하였을때평균증상유사도점수가0.983으로가장높은수치가도출된것을확인하였다.Symptom2Vec과결합하여사용자의담화를분석하는BERT기반의AnsweringClassificationModel은Accuracy가0.96으로다른모델에비해정확도가높은것을확인하였다.본연구를통해서적은양의데이터로도사용자와실시간문진이가능한것을확인하였으며논문에서제안하는알고리즘이환자와의사에게해석가능한진단과보조자료를제공할수있을것이라기대한다. |Inclinicalsettings,Doctorsspendmostoftheirtimelisteningtothesymptomsofthepatientandelicitingadditionalsymptoms.Thisactivityiscalledhistorytaking.However,thelistofquestionsdoesnothaveasetstandardsothediagnosisresultsmayvarydependingontheexperienceofthedoctor.Anincorrectdiagnosiscanendangerpatientsandcausecostlyproblems.Therefore,itisimportantforthedoctortoconductamedicalexaminationofthepatientsothatacorrectdiagnosiscanbemade. Recently,therehasbeenasignificantincreaseinresearchusingdeeplearningtechniquesinrelationtodiagnosis.Thedataset,whichisusedprimarilyinresearch,isarefinedstateinwhichnumericalinformationsuchasthebloodpressureofthepatientorbloodsugarhasbeenpre-filled.Mostofthestudiespredictdiseasesbasedoninformationsuchasthebloodofthepatientratherthanmakingadiagnosisthroughconversationswiththepatient.Thisisastudythatpredictssubsequentfiguresbasedonthecurrentnumericalinformationofthepatientorfirstcheckswhetherthepatientmeetsthediseasediagnosticcriteriadefinedinthecurrentclinicaltrial.Inaddition,thereisadisadvantagewhichisthedifficultyinensuringreliabilitybecausetheresultsofartificialintelligencecannotbeexplained.Therefore,itisdifficulttouseinclinicalpractice. Inordertobeusedinactualclinicalpractice,researchontakingamedicalhistory,whichmustbeconductedbeforediagnosis,isnecessarilyprecededbycollectingthesymptomsoftheusers,predictingdiseasebasedonthem,andprovidingexplanatoryinformationtopatients.Therefore,inthispaper,weproposeSymptom2Vec,whichcanexpresstherelationshipbetweendiseasesandsymptomsbycollectingdataondiseasesandsymptoms.BycombiningSymptom2VecwiththeBidirectionalEncoderRepresentationsfromTransformers(BERT)-basedAnsweringClassificationModelthatcananalyzethediscourseofthepatient,weproposeanautomaticdiagnosismedicalexaminationalgorithmthatcanpredictdiseasesbytakinghistoryinreal-time. PerformanceevaluationwasconductedforSymptom2VecandAnsweringClassificationModel.Comparedtotraditionalmodels,Symptom2Vechadthehighestmeansymptomsimilarityscoreof0.983.TheBERT-basedAnswerClassificationModel,whichanalyzesthediscourseoftheusersincombinationwithSymptom2Vec,confirmedthattheaccuracywas0.96,whichwashigherthanothermodels.Thisstudyconfirmedthatareal-timeinterviewwiththeuserispossibleevenwithasmallamountofdata.Weexpectthatthemedicalexaminationalgorithmisabletoprovideinterpretablediagnosisandassistancetopatientsanddoctors.