Learning Deep Representations from Heterogeneous Patient Data for Predictive Diagnosis
Chongyu Zhou, Yao Jia, Mehul Motani, Jingwei Chew · 2017
Predictive diagnosis benefits both patients and hospitals. Major challenges limiting the effectiveness of machine learning based predictive diagnosis include the lack of efficient feature selection methods and the heterogeneity of measured patient data (e.g., vital signs). In this paper, we propose DLFS, an efficient feature selection scheme based on deep learning that is applicable for heterogeneous data. DLFS is unsupervised in nature and can learn compact representations from patient data automatically for efficient prediction. In this paper, the specific problem of predicting the patients' length of stay in the hospital is investigated in a predictive diagnosis framework which uses DLFS for feature selection. Real patient data from the pneumonia database of the National University Health System (NUHS) in Singapore are collected to verify the effectiveness of DLFS. By running experiments on real-world patient data and comparing with several other commonly used feature selection methods, we demonstrate the advantage of the proposed DLFS scheme.