一种基于深度特征融合的可解释性12导联心电图自动诊断模型研究

打开文本图片集

Abstract： ObjectiveToenhance theaccuracyandreliabilityof12-leadelectrocardiogram（ECG）automaticdiagnosis.Methods Hereinweproposea12-leadECGautomaticdiagnosismodelbasedondeepfeaturefusion （MRHL-ECGNet），whichconsists ofamulti-aleatureetractionfrotdResNet34，globalatureingodulendaimeseiesalysisule. TheHyenaHierarchyConvolutionOperatorwasappliedtothe12-leadECGautomaticdiagnosistaskformoreefficient captureoflong-rangedependencies whilereducingcomputationalcomplexity.Itegrated Gradients （IG）-based interpretability analysis technologywasused toachieve visualizationofthedecision-makingbasisofMRHL-ECGNet.TheCPSC2018dataset was used to trainand testMRHL-ECGNet，andits performance wasasessedusing multiple quantitative evaluationindicators andevaluationexperiments.ResultsInthe9-class ECGclasificationtask onthe testset，MRHL-ECGNet achievedan acuracyof0.972anAUCof0.983，anF1scoreof0864，aprecisionof0.873andarecallof0.857，allsurpaingther comparative models. Thismodel only took 0.007s to output a diagnosis for a single sample on a GPU and 0.156 son a CPU， witha memoryfootprintof67.196MB.ConclusionTheproposedMRHL-ECGNetmodeldemonstratesexcellentclasification performancein12-leadECGautomaticdiagnosiswithalightweightdesignandgoodinterpretabilityandthushasgreat potential for clinical application in ECG-aided diagnosis.

Keywords： electrocardiogramautomaticdiagnosis;deeplearning;HyenaHierarchyConvolutionOperator;interpretabilityof model

心电图（ECG）作为一种记录心脏电活动的无创检测手段，能够提供关于心脏节律、传导系统以及心肌状态等重要信息，对于多种心血管疾病（CVD）的早期筛查和监测具有不可替代的作用[12。（剩余18552字）