Charactering hESCs Organoids from Electrical Signals with Machine Learning

Abstract

Three-dimensional brain organoids offer an unprecedented opportunity to study human brain development and disease. Characterizing context-specific organoid behaviors from local field potential (LFP) signals presents many machine learning challenges yet to be adequately addressed. We present in this paper classical machine learning and deep learning solutions to identify the LFP signatures unique to CHD2-mutant hESCs organoids and characterize their changes over time. We showed that a support vector machine (SVM) approach selected five frequency band-power features for 6 and 9-month-old organoids and achieved respective 99.3% and 88.1% area under the curve (AUC) performances in differentiating CHD2-mutant organoids from control using top five band-power features. We also proposed a convolutional neural network (CNN) that successfully classified both month 6 and month 9 CHD2 mutant organoids from controlled organoids using raw LFP signals with AUCs of with 99.6% and 99.7% respectively. We also reconstructed the discriminate features learned by CNN and showed their similarity and uniqueness compared with features selected by SVM.