Research On Diagnosing ECG Symptoms With The Limited Number Of Samples

Symptom recognition based on ECG is a common technique in clinical decision-making,which is generally a supervised learning problem.Therefore,sufficient high-quality labeled data is required to build a well-performing symptom recognition model.In practice,this condition is often not met,and only part of the labeled qualified data can be provided.The models built with the small amount of correctly labeled data often have poor performance and cannot satisfy the needs of clinical diagnosis.Aiming at the problem of the ECG diagnosis with only a limited number of sample,this paper proposes two solutions,a symptom recognition method based on transfer learning,and a method for symptom identification by means of causality analysis.The specific works are as follows.1.Aiming at the lack of high-quality labeled data in the symptom diagnosis,a symptom identification method based on transfer learning is proposed,which effectively utilizes manually labeled data with high-quality labels and weakly labeled data with partial or missing labels.It solves the problem of symptom identification under the condition of limited samples by increasing the training data size.In this method,the manually labeled data is used as the target domain,and the weakly labeled data is used as the source domain.Then,the feature migration algorithm is used to map the data of the two domains to the same space,so that the feature distributions of the two domains are close to each other.Finally,the instance migration algorithm is used to expand the data and build the model.This method not only alleviates the shortage of high-quality labeled data,but also effectively deals with the problem of data imbalance.2.Considering that the problem of symptom identification is a multi-label problem,and there is a certain correlation between labels,a causality-based correction method for symptom labels is proposed.Causality is a kind of knowledge that reveals the inner relationship of things.Knowledge will not change due to the amount of data,and it represents the essential information of things.From the essential connection between symptoms,the paper can reduce the amount of the required training data to a certain extent during the model training,thus effectively solving the problem of symptom identification under the condition of limited samples.The steps of the proposed method are as follows.First,the identification results of symptoms are divided into anchor set and candidate set;next,this paper performs correlation analysis on the symptoms in the anchor set,and add the symptoms that are highly correlated with the anchor symptoms but do not belong to the anchor set and the candidate set into the candidate set;a structural causal model is then constructed using the relationships between symptoms that are mined by the MMHC(Max-Min Hill-Climbing)algorithm;finally,the correction of the symptom label is completed with the help of the structural causal model.The main contributions are listed as follows.On one hand,the paper proposes the method for symptom identification based on transfer learning,which uses weakly labeled data to enrich the training data.It solves the problem of the lack of high-quality labeled data,and also deals with the problem of data imbalance,which can improve the accuracy of model recognition.On the other hand,the proposed causality-based symptom label correction method utilizes the correlation between symptoms,reduces the model’s dependence on the amount of data,and achieves the effect of improving model performance.Theoretical and experimental results demonstrate the effectiveness of the method.