Method for deeply learning and predicting medical track based on medical records

Abstract

The invention discloses a method for deeply learning and predicting medical track based on medical records. The method comprises the following steps: S1, encoding diagnostic information and intervention information on admission through an encoding scheme and converting code into vector to acquire diagnostic information conversion vector (the formula is shown in the description) and intervention information conversion vector (the formula is shown in the description) separately, and converting the diagnostic information and intervention information on admission for one time into one 2M-dimensional vector [xt, pt]; S2, input the vector [xt, pt] into an LSTM model, and evaluating the current output value ht to obtain the current disease state; S3, predicting a diagnostic code dt+1 according tothe disease state ht and predicting the disease progression through the diagnostic code dt+1; S4, calculating an intervention code st of the time t, increasing a time structure in the LSTM model, collecting the historical disease states in multiple time ranges, collecting the state of each section of a horizontal time shaft, collecting all the diseases states, stacking into a vector (the formulais shown in the description), and feeding back the vector (the formula is shown in the description) into a nerve network to predict the future risk result Y.

Description

technical field [0001] The invention relates to the field of predictive models, in particular to a method for predicting medical trajectory based on deep learning of medical records. Background technique [0002] For predicting the future medical risk of a patient after admission, four open questions need to be faced: [0003] (1) Long-term medical dependence; [0004] (2) Representation of admission information; [0005] (3) Plot recording and irregular time; [0006] (4) Confounding the interaction between disease progression and intervention. [0007] Existing methods are poor at dealing with these problems. They do not adequately simulate variable lengths and ignore long-term dependencies. Time models based on Markov assumptions are limited to model time irregularities and have no memory, so they may completely forget the impact of previous major diseases on the body. . [0008] In recent years, deep learning (such as speech recognition, vision and computational li...

AI Technical Summary

Problems solved by technology

[0009] The purpose of the present invention is to solve the above problems and provide a method for predicting medical trajectory through deep learning that can solve the problem of medical risk prediction

Images