適切な特徴量を用いた高速な機械学習型iPSCs 由来自己拍動細胞塊の分化度評価手法

抄録

Regenerative medicine has attracted attention as a novel therapeutic approach for restoring damaged or impaired tissue and organ functions. The advancement of regenerative medicine requires efficient, safe, and high-quality induced pluripotent stem cell (iPSC) production, along with quantitative methods for differentiation status assessment. In this study, we aimed to identify characteristic features derived from the beating dynamics of iPSC-derived beating cardiomyocyte clusters as well as to propose and validate fast and highly accurate machine learning-based method for determining their differentiation status. The method we present in this study involves cell beating behavior acceleration vector calculation via optical flow analysis to construct a 3D beating point cloud defined by time frame, angle, and acceleration. Next, we extracted four features (periodicity, periodic variance, convergence, and convergence variance) and assessed the degree of differentiation using unsupervised (k-means＋＋, GMM) and supervised (SVM, decision tree) learning. We achieved the highest accuracy of 96.9% using a decision tree. In addition, by leveraging the structural properties of the 3D beating point cloud, we reduced the number of features from four to two while maintaining accuracy and decreasing computation time by approximately 80%. Moreover, in our study, we discuss why high differentiation accuracy is preserved with fewer features. Overall, the proposed method provides an efficient and highly accurate approach for iPSC-derived cardiomyocyte cluster differentiation status evaluation using a minimal feature set.