From calcium imaging to new functional biomarkers: machine learning unlocks disease-discriminating signatures of cardiac pathologies and arrhythmia syndromes
T Kaas, J Huettemeister, G Ramesh, N Goehringer, M Boegner, M Kirk, H Stachelscheid, F Heinzel, L Blatter, G Hindricks, F HohendannerAbstract
Background/Introduction
Despite the central role of calcium dysregulation across heart failure, ischaemic heart disease, and hereditary arrhythmia syndromes, no standardised framework exists for functional single-cell phenotyping across disease entities and species. Confocal calcium imaging captures these dynamics with high spatiotemporal resolution, yet analysis remains limited to few manually defined parameters, creating a bottleneck for drug screening and genotype variant adjudication.
Purpose
To establish and validate an automated machine-learning framework for classifying cardiac phenotypes solely from confocal calcium imaging alone across species, pathologies, and hereditary arrhythmia genotypes, and to assess cross-laboratory generalisability for translational deployment.
Methods
Recordings from 621 cardiomyocytes (5249 calcium transients; five species; heart failure, myocardial infarction, catecholaminergic polymorphic ventricular tachycardia [CPVT]) were analysed. Conventional ML models (Random Forest, SVM, XGBoost) were trained on engineered transient descriptors. A finetuned ResNet-18 was evaluated using three preprocessing strategies: whole-image downscaling, multi-crop aggregation, transient-centred cropping. Group-aware stratified cross-validation ensured strict leakage control. External validation used additional 191 scans from a separate laboratory and acquisition period.
Results
On training data, both conventional ML and ResNet-18 showed strong discriminative ability across all three tasks. Optimal preprocessing was task-dependent: transient-centred cropping was most effective for species and pathology, while multi-crop aggregation yielded strongest CPVT genotype discrimination, potentially reflecting marked distributed diastolic events benefiting from scan-wide sampling. Engineered features outperformed deep learning only for species while ResNet-18 was superior for pathology and genotype. On external validation, classical models generalised more robustly (Random Forest ROC-AUC 0.94±0.01 vs. best ResNet-18 0.86±0.01; p=0.013), confirming calcium signatures transfer across laboratories while image-based models show greater domain-shift vulnerability. Post-hoc calibration improved the reliability of CNN probability estimates (ECE 0.14 to 0.08). SHAP identified rise-time as most dominant; Grad-CAM confirmed biologically plausible attention.
Conclusion(s)
Calcium imaging contain sufficient information for automated classification across a biologically meaningful hierarchy. Deep learning surpassed conventional models for pathology and genotype; engineered features were more robust under domain shift. These findings establish the foundation for microscopy-based functional biomarker platforms, with next steps including multi-site validation for automated drug screening and genotype variant adjudication.