Model for predicting disease risk by applying convolutional neural network

Abstract

The invention discloses a model for predicting a disease risk by applying a convolutional neural network. The model is a four-layer convolutional neural network model which comprises a first layer, asecond layer, a third layer and a fourth layer; the first layer is composed of event matrixes built by EHRS, the second layer is a single side convolutional layer for extracting phenotypes the EHR matrixes in the first layer, the third layer is the maximum pooling layer with important phenotypes in the phenotypes extracted in the second layer, and the fourth layer is a softmax predicted whole connection layer. In the event matrix in the first layer, the horizontal dimension corresponds to a time stamp, and the vertical dimension corresponds to an event value; for combining patient EHR time smoothness, three different time fusion mechanisms are researched in the model; early fusion, late fusion and slow fusion are performed, and an advanced CNN framework with time fusion is built.

Description

technical field [0001] The invention relates to the field of predictive models, in particular to a model for predicting disease risk using a convolutional neural network. Background technique [0002] Healthcare systems worldwide are rapidly adopting electronic health records (EHRs), which are systematic collections of longitudinal patient health information (e.g., diagnoses, medications, laboratory tests, procedures, etc.) resulting from one or more encounters in any care delivery setting . This will greatly increase the availability of electronic clinical data. [0003] In recent years, interest in data analysis of patients' electronic medical records has exploded. Data-driven healthcare aims to effectively utilize big data representing learning collected in treating hundreds of millions of patients to provide the best and most personalized care, and is considered the most promising transformational healthcare. Electronic medical records are one of the main vehicles dri...

AI Technical Summary

Problems solved by technology

Although some computational models for EHR-based electronic phenotyping have been proposed recently (e.g., matrix-based methods and tensor-based algorithms), there are still many challenges such as temporality, sparsity, noise, bias, etc. Therefore, how to perform efficient feature extraction and phenotyping from patients' EHRs before further application is a critical step

Images