A Novel Deep Neural Network Model for Multi-Label Chronic Disease Prediction

Conclusions concludes this work along with future work. Dataset and Data Preprocessing In the work, we mainly focus on multiple chronic disease classification. It can be formulated into a multi-label classification problem. There are three common chronic diseases are selected from the physical examination records: hypertension (H), diabetes (D), and fatty liver (FL). In the experiments, the physical examination datasets are collected from a local medical center, which contain 110,300 physical examination records from about 80,000 anonymous patients (Li et al., 2017a,b). Sixty-two feature items are selected from over 100 examination items based on medical expert experience and related literature in every physical examination record. These feature items contain 4 basic physical examination items, 26 blood routine items, 12 urine routine items, and 20 items from liver function. Two multi-label transformation methods consisting of binary relevance (BR) and label powerset (LP) method are used in this work. For BR method, the diagnosis of a given patient can be one of three possible results: all three chronic diseases, different combination of the chronic diseases, or no signs of any three chronic diseases, which means that there are totally eight different sets of diagnoses {000, 100, 010, 001, 110, 101, 011, 111}. Based on Label Powerset (LP) method, we get eight different prediction labels and can be represented by {0, 1, 2, 3, 4, 5, 6, 7}. In order to understand dataset be...
Source: Frontiers in Genetics - Category: Genetics & Stem Cells Source Type: research