A major bottleneck in human pluripotent stem cell (hPSC)-derived cardiac therapy is the approximately 10-day latency required to confirm functional beating, resulting in substantial resource loss when suboptimal cultures are maintained to end-stage without early indicators of failure. To enable earlier intervention, we developed a non-destructive AI pipeline that uses a vision–language model (VLM) to extract four morphological biomarkers—compaction, edge definition, core density, and background cleanliness—from brightfield images collected between Days 3 and 10 of spheroid development. Temporal features derived from these measurements were used to train a Random Forest classifier using 153 labeled spheroids. The model achieved an overall accuracy of 79.5% (AUC-ROC: 0.815), with 91.7% precision, 61.1% recall, 73.3% F1 score, and 95.2% specificity for predicting functional beating outcomes. Temporal ablation demonstrated predictive signal as early as Day 3, with performance improving by Day 5 (AUC-ROC = 0.767), while longitudinal background-cleanliness and core-density dynamics further enhanced performance when later timepoints were included. Although recall reached 61.1%, the system was intentionally optimized for high specificity to establish a robust quality-control gate, ensuring that cultures advancing through manufacturing are highly likely to be functional. This scalable, automated framework enables non-destructive quality assessment several days before conventional Day 10 functional evaluation without specialized reagents or destructive sampling, improving economic and scientific efficiency while reducing laboratory waste associated with maintaining non-functional cultures.
