Research On Automatic Pulse Wave Classification Algorithm Under The Condition Of Few Shot

Remote portable wearable devices such as pulse wave sphygmomanometers and smart bracelets take into account safety and health functions,and can continuously monitor heart activity.The popularity of wearable devices makes family-oriented medical and health services more complete.These devices record a large amount of pulse wave data,which provides a data basis for classification and recognition using machine learning technology.Supervised learning is the process of extracting knowledge from a limited labeled data set,and then labeling unknown data.The core is to discover hidden laws from discrete category data,combine low-level features to form more abstract high-level representative attribute features,and discover data distribution feature representations.Among them,the key to effectively extracting the main features lies in the quality and quantity of the labeled data.However,the labeling of pulse waves is laborious and time-consuming,and the sample size of the existing publicly labeled pulse wave database is not enough to train a classification model with high accuracy and high generalization.Due to the few-shot characteristics of pulse wave,the feature expression ability of classification model trained by supervised learning is insufficient,and the risk of over fitting is high.From the perspective of network unit structure,model integration and model migration,this paper explores the method to improve the accuracy of pulse wave classification under the condition of few-shots.The main research contents are as follows:(1)In view of the complex structure of the long and short-term memory network unit and the large number of parameters,a gated memory unit structure suitable for small sample pulse waves is designed.Improved on the basis of long and short-term memory units,designed a forget gate to control the transmission of timing information,and reduce the model's demand for training samples by reducing the number of neuron parameters.(2)In view of the problem that the recursive network classification model based on gated memory unit is not effective for single-cycle pulse wave classification,a small sample pulse wave classification model based on multi-scale feature extraction is designed.Considering the rhythm characteristics,timing characteristics and spatial characteristics of the pulse wave together,the task of classifying pulse waves under different scale conditions can be satisfied.(3)In view of the problem that the training of complex models still requires a large number of labeled pulse waves,the ECG-pulse wave migration learning classification model based on the generation of confrontation network is designed,which can extract the shared characteristics of the ECG and pulse wave for classification,and realize the conversion from the ECG to the pulse wave.Cross-domain migration can still have a better classification effect when the target domain has a small number of labeled pulse waves.