Fig. 4

Morphology-based prediction of MSC’s potencies. a–c Classification performance of T-cell proliferation inhibitory potency compared with modeling methods, type of using morphological parameters, and usage of time-course window size. Samples was classified into two categories: (Low-risk lots) Lots 1–7, (High-risk lots) Lots 8–11. kNN (a), LASSO (b), and RF (c) were compared. Mean + SD: using total 32 morphological parameters combining both mean- and SD-related parameters, Without SD: using mean-related 16 parameters, Only SD: using SD-related 16 parameters. The best model that exceeded the dotted line performance was selected. d Comparison of 5-day and 4-day models with each lot prediction results. e, f Prediction performance of growth rate compared with modeling methods, type of using morphological parameters, and usage of time-course window size. The growth rate (cells in the image at 138 h/cell in the image at 12 h) was used for training each sample. LASSO (e) and RF (f) were compared. Mean + SD: using a total of 32 morphological parameters combining both mean- and SD-related parameters, Without SD: using mean-related 16 parameters, Only SD: using SD-related 16 parameters. The best model showing RMSE lower than the dotted line performance was selected. g 5-day model performance of growth prediction. h 4-day model performance of growth prediction. i Top 15 parameters selected in the LASSO with only SD model. The number of selections indicates the total number of selections during the leave-lot-out cross-validation. j Weights for the top 15 parameters selected in the LASSO with the only SD model