Hierarchy-Aware AI Nuclear Phenotyping in Human iPSC-Derived Cardiomyocytes: A Preclinical Framework for Imaging Biomarker Development

Background/Objectives: Nuclear morphology and chromatin organisation provide quantifiable imaging features that may serve as candidate biomarkers of cellular state. However, many cellular imaging studies analyse multiple cells from the same microscopy field without explicitly accounting for their hierarchical dependence, which may lead to optimistic estimates of model performance. This study developed a hierarchy-aware AI framework to investigate DAPI-derived nuclear phenotypes in human iPSC-derived cardiomyocytes (iPSC-CMs). Methods: A six-day fluorescence microscopy dataset was processed using marker-controlled watershed segmentation, automated quality control, morphology and masked chromatin-texture analysis. Multiple deep-learning, transfer-learning, multiple-instance learning, and self-supervised approaches were evaluated. The final dataset contained 154 quality-controlled nuclei from 51 source FOVs. The primary prediction task was the binary classification of operationally defined Early (Days 1–3) versus Late (Days 4–6) experimental timepoint groups from DAPI-derived nuclear images. Primary evaluation used nested five-fold FOV-grouped cross-validation with three random seeds, preventing nuclei from the same FOV from appearing in both training and test sets. Results: For the primary Early-versus-Late timepoint classification task, the prespecified ViT-B/16 model achieved 52.9% accuracy, an F1-score of 0.571, and a ROC-AUC of 0.586. Among secondary full-time-course models, hybrid morphology-gated attention multiple-instance learning achieved the highest accuracy (58.8%) and F1-score (0.644), while frozen-ViT neural pooling achieved the highest ROC-AUC (0.614). In the secondary Days 1–2 versus Days 5–6 extreme-timepoint analysis, the DINOv2 multi-view model achieved 76.7% accuracy, an F1-score of 0.741, and a ROC-AUC of 0.765. FOV-balanced resampling produced similar ROC-AUC estimates for the primary ViT (0.583) and DINOv2 (0.750) analyses. Nuclear area showed the largest FOV-level effect (d=0.59, p=0.0406, q=0.442), with none of the 12 imaging descriptors significant after Benjamini–Hochberg correction. Conclusions: DAPI-derived nuclear phenotypes showed a detectable but heterogeneous time-associated imaging signal, with stronger discrimination between more widely separated timepoints. These findings represent candidate imaging signatures rather than validated diagnostic biomarkers. The framework provides a hierarchy-aware approach for prioritising nuclear imaging features for future biological and translational validation.