Machine Learning Based Information Extraction for Diabetic Nephropathy in Clinical Text Documents

Abstract

Diabetic nephropathy is common complication of diabetes mellitus, it's important to intervene early. For building a predictive model for diabetic nephropathy. In order to extract relevant information as a prediction risk factor, we construct a golden standard corpus. 3422 admission summary notes from 2013 to 2017 in a tertiary hospital were included in the study. An information extraction method based on machine learning models is proposed to extract important information from unstructured medical record texts, in which Adaboost on Duration of Diabetes has best performance (F1=0.97), and Family history of heart disease extraction is most challenge, F1 value of best model result SVM is 0.73. The best performance of the other six types of information extraction model is between 0.85 and 0.96, and the practical application is feasible.