Data Processing of Single-Cell Impedance Profiles Using Three Deep Neural Networks of Feature-Level Joint/Marginal Fusions and Decision-Level Meta-Learning, a Comparative Study

This study presented data processing of singlecell impedance profiles using three information fusion methods of feature-level joint fusion, feature-level marginal fusion and decision-level meta-learning. Inherent features extracted from numerical simulation and recessive features extracted from recurrent neural network of single-cell impedance profiles were jointly fused by feedforward neural network or marginally fused by support vector machine at the featurelevel. Meanwhile, decision-level meta-learning was conducted to fuse prediction results of numerical simulation and recurrent neural network of single-cell impedance profiles based on feedforward neural network. In comparison to recurrent neural network only, these three types of neural networks of data fusion recorded a 50% decrease of classification inaccuracy in differentiating three leukemia cell lines of HL60, Jurkat, and K562.