Personalized Atrial Fibrillation Detection Based On Transfer Learning

Atrial fibrillation(AF)is one of the most common arrhythmias that can cause stroke,atrial thrombosis,heart failure and stroke.With the increase of age,the cardio-cerebrovascular system becomes more and more fragile,and the resistance to heart disease gradually decreases,leading to an increase in the prevalence of atrial fibrillation,which becomes a disease frequently occurring in the elderly and even adults.Therefore,accurate detection of atrial fibrillation and effective treatment measures are of great significance.However,the digital characteristics of atrial fibrillation electrocardiogram(ECG)signal are not clear,and there are great differences in morphology and time domain characteristics between different patients.Therefore,the accuracy may be dramatically reduced when the atrial fibrillation detection algorithm trained with partial patient data is applied to new patients,which makes personalized atrial fibrillation detection a challenging problem.The differences among patients are often ignored and there is a promotion for the accuracy and generalization ability in existing atrial fibrillation detection algorithms.In order to solve the above problems,this paper proposes a personalized automatic identification algorithm of atrial fibrillation based on information entropy and transfer learning.The main work of this paper is as follows:(1)A high performance AF detection algorithm based on empirical mode decomposition and fusion entropy is proposed.The empirical mode decomposition(EMD)method was used to decompose the heart rate variability signals into a set of eigenmode functions to obtain the frequencies of the original signals on different time scales.In order to obtain the characteristics that can characterize the HRV signal in local space and frequency domain,five kinds of information entropies are extracted from the eigenmode function,which are approximate entropy,sample entropy,fuzzy entropy,energy entropy and perturbation entropy.Finally,support vector machine was used to detect atrial fibrillation.The algorithm was verified on MIT-BIH-AF,MIT-BIH-NSR and MIT-BIH-Arr public databases.The results on MIT-BIH-AF yielded the best performance which were 99.01% sensitivity,98.77% specificity and 98.37% accuracy.The performance of the method has been improved compared with several atrial fibrillation detection algorithms,which verifies the effectiveness of the method in the detection of atrial fibrillation.(2)In order to solve the problem of large differences in ECG signals among different patients,a personalized AF detection algorithm based on transfer learning was proposed.Joint distribution adaptation combined with fusion entropy was introduced to reduce the distribution difference between subjects and other patient groups.First of all,the fusion entropy algorithm is utilized to extract the characteristics of the map to reproducing kernel hilbert space,and then the maximum average difference was adopted to measure the distance between the subjects and patients with other data distribution,marginal adaptation probability,conditional probability to obtain the best feature subspace;finally,k-nearest neighbor(KNN)classifier trained on different patient groups is applied to the participants to achieve good classification results.The sensitivity,specificity and accuracy of personalized detection of atrial fibrillation on MIT-BIH-AF were 98.69%,98.47% and 97.65%,respectively.In order to further verify the effectiveness of the joint distributed adaptation method in personalized atrial fibrillation identification,the results were compared with principal component analysis and migration component analysis,respectively.The results showed that the joint probabilistic adaptation method could effectively reduce the distribution differences among different patients,so as to achieve personalized atrial fibrillation detection.