Medical entity vector conversion method based on attention mechanism

Abstract

The invention provides a medical entity vector conversion method based on an attention mechanism. A MedE2vec model is established, and the method specifically comprises the following steps: enabling an electronic medical record of the whole medical process of a patient to comprise multiple diagnosis and treatment events Event with scattered time distribution, wherein the single diagnosis and treatment event Event is composed of multiple medical entities Entity of the patient; inputting all medical entities of one clinical diagnosis and treatment of a patient, and initializing the medical entities into a diagnosis and treatment sequence represented by a vector by an initialization vector matrix W; capturing a relationship, namely an attention mechanism, between medical entities in the patient diagnosis and treatment event sequence V; capturing a relationship between different diagnosis and treatment events of the patient; obtaining a vector matrix W through iterative training, wherein the ith row in the W represents a vector in the medical entity set; and continuously optimizing the vector matrix W through a loss function to obtain a final medical entity vector. The invention discloses a deep learning model based on an attention mechanism. The MedE2vec can generate a more accurate medical entity vector.

Description

technical field [0001] The invention belongs to the field of computers, and in particular relates to a medical entity vector conversion method based on an attention mechanism. Background technique [0002] Language concept word vector conversion is an important method to extract information from data, especially for those unstructured text data. Only by accurately converting medical textual descriptions into numerical information, known as concept vectors, can they be analyzed using sophisticated machine learning and deep learning models. Although there are already some mature text-to-vector conversion tools, few of them are specially suitable for medical data, and most of them cannot capture the time series information in medical information, so the effect is not ideal. [0003] With the popularity of electronic medical records, the data accumulated by electronic medical record systems is also increasing, which makes it possible for us to mine valuable information from dat...

AI Technical Summary

Problems solved by technology

Although CNN can do parallel computing, it cannot capture long-distance feature relationships in natural language processing tasks. For example, in machine translation, CNN can only capture the relationship between words that are close to the word to be translated, and cannot get the distance The influence of more distant words on it, thus causing the problem of inaccurate translation

Images