Abstract 17481: Automated Analysis of Action Potentials from Cardiac Cell Clusters Derived from Human Embryonic Stem Cells
Background: Cardiomyocytes (CMs) can be derived from human embryonic stem cells (hESCs) via embryoid body (EB) methods. EBs contain a heterogeneous population of CMs, which are typically identified as different phenotypes by their action potential (AP) morphology. Using techniques from signal processing and machine learning, we developed an automated algorithm to identify different phenotypes within a dataset of APs.
Methods: Optical APs were obtained from 9 beating cell clusters derived from hESCs (H9 line) after 20 days differentiation. Using a voltage-sensitive dye, APs were recorded over a 1.6 mm × 1.6 mm area at 16 micron spatial resolution, giving an average of 771±159 recordings per cluster. Signals were averaged by a 5×5 spatial filter to improve signal quality. APs were temporally aligned by their activation times, determined as the time of maximum upstroke slope. The similarity between APs of different recording sites was obtained by calculating the integral of the squared difference between signals. Spectral clustering (SC) was then applied to the similarities of the entire dataset to sort the recordings into groups. Segmentations were evaluated by two different metrics (SC cost and Davies-Bouldin index) that measured similarity within each group and dissimilarity among different groups. AP duration at 80% repolarization (APD80) and time from APD30 to APD90 (triangulation) were calculated for each recording.
Results: APD80 of 6940 total recordings (9 clusters paced at 90 pulses/min) was 128±29 ms (mean±SD). SC was performed on the entire dataset, assuming 2, 3, or 4 groups. Both metrics indicated that 2 groups best represented the dataset. Group 1 had longer APD80 and more triangulation than group 2 (147±16 ms vs 97±17 ms and 66±13 ms vs 58±18 ms, respectively). One group dominated over the other (occupying > 92% of the recorded area) in 4 of the clusters, while the two groups were more comparable in size in the other 5 clusters.
Conclusions: An algorithm has been developed to sort hESC-derived CMs into groups based on their AP morphology. When sorted into two groups, the percentages of each group varied greatly among the clusters. This work is a first step towards an automated framework to identify AP phenotypes of hESC-derived CMs.
- © 2012 by American Heart Association, Inc.