Research Of Heart Failure Detection Algorithm Based On Multi-modal Cardiac Physiological Signal And Deep Neural Network

Cardiovascular disease is the leading cause of death in humans,with heart failure being the primary cause of death among cardiovascular diseases.It severely affects patients’ quality of life and life expectancy,imposing a significant medical burden on society.Heart failure is a complex clinical syndrome characterized by abnormal heart structure or function,leading to symptoms and signs caused by ventricular contraction or filling disorders.Among them,left ventricular dysfunction is one of the more severe forms of heart failure.The main methods for diagnosing heart failure in clinical practice are transthoracic echocardiography,B-type natriuretic peptide,and N-terminal pro-brain natriuretic peptide.However,these methods rely on the expertise and level of specialists and have limitations such as high cost,slow speed,or invasiveness.Early intervention can effectively reduce the mortality and hospitalization rates of heart failure patients,but many patients are still not diagnosed in time.Timely screening and diagnosis of heart failure are of great significance for disease prevention and prognosis improvement.Heart sounds and electrocardiographic signals reflect the electrical and mechanical activity characteristics of the heart.Combining the two can comprehensively record information on cardiac electromechanical activity.Both can be obtained using noninvasive methods,with low equipment costs and simple collection conditions,making it easy to conduct medical examinations.Therefore,analysis methods based on these twosignals provide the possibility for early non-invasive detection of coronary heart disease.Existing research on clinical diagnosis of heart failure based on heart sounds and electrocardiography relies on manual measurement and manual analysis of signals and cannot achieve low-cost universal screening.Existing research based on machine learning only analyzes single-modality signals and fails to utilize the complementary relationship between dual-modality signals.Existing heart failure-related datasets have problems such as single data sources,small data volume,and low data quality.This paper establishes a dataset of synchronized electrocardiogram and heart sound data collected clinically for heart failure.It studies the effectiveness of multimodal cardiac physiological signals in detecting heart failure and explores the application value of multimodal signal feature fusion and large-scale datasets in early non-invasive detection of heart failure.The main work and innovations of this paper are as follows:1.A synchronized heart sound and electrocardiogram dataset of patients with left ventricular dysfunction was established.The data consists of synchronized heart sounds and electrocardiograms in the time dimension,with the label value being the left ventricular ejection fraction measured by the biplane Simpson method.The baseline information of the patients includes gender,age,systolic blood pressure,diastolic blood pressure,and underlying diseases.The dataset is a medium-sized multimodal cardiac physiological signal dataset,containing a total of 1046 samples from 806 patients.2.Based on the synchronized heart sound and electrocardiogram dataset of patients with left ventricular dysfunction,a multimodal deep neural network was proposed.It aims to extract features between modalities through a temporal neural network and then analyze heart sounds and electrocardiograms simultaneously to screen for left ventricular dysfunction.The model is a two-stage multimodal fusion neural network,consisting of a feature extraction and fusion segment and a feature classification segment.It can learn information between modalities simultaneously in the frequency and time domains and perform fusion analysis.3.Based on multi-head self-attention,the signal feature extraction layer in the neural network was improved on the basis of the above research,and a multimodal left ventricular dysfunction detection model based on Transformer Encoder was proposed.Compared with traditional sequential models,it has a shorter maximum path length and can compute each element in the sequence in parallel,improving the computational speed of the model while slightly increasing the detection accuracy.4.An ECG-TTE data pair dataset was established.The collected data is a standard 12-lead electrocardiogram.The annotation information includes the left ventricular ejection fraction measured by the biplane Simpson method,basic omics information,clinical diagnosis,laboratory diagnosis,electrocardiogram examination results,past medical history,etc.After data cleaning,a total of 56,817 12-lead electrocardiogram data from 44,730 patients were obtained.5.Based on the ECG-TTE data pair dataset,a variable-size heart failure detection algorithm based on channel fusion and mobile inverted bottleneck convolution was constructed.This algorithm can identify patients with left ventricular dysfunction from 12-lead electrocardiogram data in an end-to-end manner.Through Grad Saliency Map method to conduct interpretability experiments on the model,it was found that the gradient mainly concentrates on the QT interval,QRS wave group and T wave of the electrocardiogram and is significant in V1-V5 leads and AVF leads.At the same time,this model can flexibly change the number of input leads and model size to adapt to devices with different computing resources.