Study On Sensorless Suction Detection Method Of Left Ventricular Assist Devices

The left ventricular assist device(LVAD)is one of the important solutions for thetreatment of heart failure.However,while using LVAD,there is a dangerous clinical phenomenon in which the LVAD pumps too much blood from the left ventricle and exceeds the normal blood supply,causing left ventricular collapse,called suction.In the related study on suction,the blood pump signals of LVAD are important tools for detecting suction.In this thesis,suction detection is studied using pump speed(PS),which is the intrinsic parameter of the LVAD.This study achieves the sensorless suction detection based on LVAD.Main contents and results are as follows:1.Simulation for PS and pump flow(PF)signals.According to the human cardiovascular system,combined with the LVAD,the circuit component was used to establish the human cardiovascular system-LVAD model to collect the pump speed signal for suction detection in the study,and the pump flow signal was collected for comparison.Nine types of pump speed and pump flow signals are collected in various conditions of heart failure and different activity levels of patients.Pre-process the original signals to establish PS and PF signals data set.After comparing the original signals of PS and PF with the signals filtered,it can be concluded that the influence on the two noisy blood pump signals could be basically eliminated.2.Feature extraction and selection of pump speed and pump flow signals.The PS and PF signals' features are extracted by the proposed algorithms with suction indices from SI1 to SI21.Using graphical results to visually illustrate the 21 feature separability.The results show that most features are intuitively highly separable.After extracting 21 types of features of PS and PF signals in 9 physiologic cases,they are classified according to No Suction and Suction states.The feature sets of PS and PF signals are established.The feature weights of every feature are calculated by Relief algorithm,and the feature selection is combined with the graphical results of feature extraction.Among the 21 features,the features with high weight are selected,and the redundant features and the bad features below the threshold are removed.After the feature SET-A(15 features)is selected,the threshold is further increased,and the number of features is reduced to the feature SET-B(9 features).The feature selection process reduces the data dimension of the proposed algorithm,improves the efficiency of the suction detection,and increases the accuracy of the suction detection.3.Analysis of classification and suction detection results.The SET-A and SET-B feature sets of PS and PF signals are classified and identified by 5 different classification methods.The results show:(1)The detection accuracy based on PS is higher than the detection results based on PF,indicating that the proposed suction method is feasible.(2)The suction detection accuracy based on PS signal with feature SET-A are higher than 97% for all five classification methods,and the BP neural network has the best detection result,which is98.75% as high detection accuracy;(3)In comparison,the accuracy of the five classification detectors based on the feature SET-B is higher than 96%,and the result of the discriminant classification algorithm has the best accuracy,which is 97.78%.Compared with SET-A,the accuracy with SET-B is slightly lower but no significant degradation in the detection performance.SET-A should be used for high suction detection accuracy.If it is necessary to reduce computational complexity while ensuring the accuracy,SET-B should be used.