Browse Prior Art Database

A Method for Risk Prediction from Electronic Health Record with Medical Feature Embedding

IP.com Disclosure Number: IPCOM000248190D
Publication Date: 2016-Nov-07
Document File: 4 page(s) / 301K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method to develop a general framework for extracting effective features from longitudinal patient data using word2vec. The method automatically extracts low-dimension patterns, automatically selects discriminative temporal features, and efficiently reduces feature dimension into a scalable size.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 01 of 4

A Method for Risk Prediction from Electronic Health Record with Medical Feature Embedding

Extracting effective efficient features from longitudinal patient data such as electronic health records (EHRs) and sensors, is of central importance for risk prediction in the healthcare domain. Challenges of risk prediction from EHRs include high-dimensionality of data sparsity, manual feature extraction, and temporality of data.

The current methodologies on computational prototyping are limited to a bag-of-words representation, which encounters high dimensional and sparsity problems. Traditional

dimension reduction methods cannot learn semantic meaningful features and consider the temporal information. In addition, current methods still need the manual feature extraction procedure.

A method is needed to produce low-dimension and meaningful representations of

words from the corpus.

The novel solution is a method to develop a general framework for extracting effective features from longitudinal patient data using word2vec. The method automatically extracts low-dimension patterns, automatically selects discriminative temporal features, and efficiently reduces feature dimension into a scalable size.

Figure 1: Summary of the solution

Figure 2: Word2vec for medical events


• Consider co-occurrence, correlation, comorbidity, etc. (e.g., essential hypertension and benign hypertension)

• Consider local window of each event
• Unsupervised training

1


Page 02 of 4

Figure 3: How word2vec is trained


• Given one event, maximiz...