The scaled error rate indicates the median value of prediction errors among all the samples, normalized by the actual experimental values. Therefore, low scaled error rate indicates that prediction values are relevant to the experimental values. However, usage of the scaled error rate alone is vulnerable to accidental prediction noises. Therefore, we introduced the second criterion, correlation coefficient, which evaluates the combined correlations of experimentally determined values and predicted values among all samples. These criteria are complementary: scaled error reflects differences between plots, however discards information about overall plot accuracy, whereas correlation coefficient reflects the overall similarity of measurements and predictions, however is sensitive to outliers. The combination of low scaled error rate and high correlation coefficient indicates stable performance of a given model. Figure 6 depicts all prediction results. The data indicate that by using only the prior morphologies before the differentiation process, future collapses in all of differentiation potentials under continuous passage stresses can be predicted in advance. Comparisons of the transition patterns of the experimentally determined and predicted values revealed that all cellular properties were predicted with reasonable accuracy. Furthermore, in contrast to our previous study that used all morphological data from 14 days of differentiation culture period, the predictive performance was enhanced in this study using morphological data collected only from the first 4 days before the differentiation. For prediction of potential I, the best prediction accuracy was achieved by Model 3, which utilizes both morphological features and gene-expression profiles. Compared to the scaled error rate of the NULL model, the performance of Model 3 can be expressed as 2.6fold more accurate. In the sense of cost-efficiency of model construction, Model 8, which utilizes only the morphological features from 24 h, had a reasonably high predictive performance. For predicting potential II, morphology-based models such as Models 4, 6, 7, and 9 yielded extremely high predictive performance. Model 4 achieved the best accuracy, and Model 7 was the best model at the lowest cost. For predicting potential III, most of the models could not significantly outperform the NULL model. However, Model 9 had fairly accurate predictive performance. To replace human estimations of cell quality in the production of cells for cell-based PI-103 abmole bioscience therapies, we examined the performances of machine-learning models in predicting the quantitative rates of multi-lineage differentiation after long-term differentiation, using data from undifferentiated label-free images of hBMSCs. The novel advancing technological points achieved in this work are illustrated in Fig. S1. From images collected during the first 4 days of expansion culture before differentiation, the morphological features of each cell in the images were individually measured and converted into various morphological metrics that DAPT 208255-80-5 represented the statistical morphological profiles of the group of cells. These features were then used to train computational models that forecast the experimental results collected 2–4 weeks after the differentiation. Advancing from our previous success in predicting the single-lineage differentiation potentials of hBMSCs.