Learning High-dimensional Associations for Nonalcoholic Fatty Liver Disease Diagnosis Prediction

Nonalcoholic Fatty Liver Disease (NAFLD) is a major liver disease worldwide, and NAFLD diagnosis is the clinical foundation before healthcare strategies. NAFLD diagnosis predictive methods provide a promising way for intelligent diagnosis. Previous diagnoses mainly relied on liver biopsy and imaging, which are unsuitable for large-scale rapid screening. However, the observed quantitative data are high-dimensional and have missing values. To alleviate this issue, the stacking multi-scaled convolutional neural network (SMCNN) is proposed to predict NAFLD diagnosis. Firstly, the inputs are normalized and transformed. Secondly, the inputs are represented as multi-scaled feature maps. Finally, the scaled feature maps are stacked to connect with outputs. Several experiments have been done to validate twelve methods on a real dataset. The proposed SMCNN outperforms eleven methods in terms of five metrics.