From cellular calcium imaging to precision diagnostics: deep learning-based classification of hereditary 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
Hereditary arrhythmia syndromes such as catecholaminergic polymorphic ventricular tachycardia (CPVT) are rare but potentially fatal causes of sudden cardiac death in the young. Despite advances in genetic testing, many patients remain undiagnosed due to variants of uncertain significance and inconclusive stress testing. iPSC-derived cardiomyocytes enable direct visualization of patient-specific calcium-handling abnormalities, but manual analysis is time-consuming and limits clinical translation.
Purpose
To address these limitations, this project developed and validated an automated machine-learning framework using confocal calcium-imaging data from patient-derived iPSC-cardiomyocytes. A hierarchical diagnostic approach—from species classification to major cardiac pathologies and genotype identification—demonstrated model sensitivity to interspecies, disease-related, and mutation-specific calcium-handling differences.
Methods
A curated multi-species dataset (>600 line-scan recordings from human, rabbit, rat, mouse, and pig cardiomyocytes with heart failure, myocardial infarction, or CPVT3) was analyzed. Conventional models (Random Forest, SVM, XGBoost) were trained on manually extracted calcium-transient features, while a ResNet-18 network was fine-tuned on raw microscopy data from whole line-scan images using transfer learning. Performance was evaluated by weighted F1-score (wF1) with stratified 10-fold cross-validation and bootstrapped 95% confidence intervals. Model interpretability was assessed using SHAP values and Grad-CAM.
Results
Conventional machine-learning models using manually extracted calcium-transient features achieved wF1 = 0.79 ± 0.02 (species), 0.83 ± 0.01 (pathology), and 0.83 ± 0.04 (genotype). The ResNet-based approach applied to compressed microscopy data reached comparable or higher performance, improving species (wF1 = 0.68 ± 0.03 → 0.90 ± 0.08) and pathology (wF1 = 0.80 ± 0.02 → 0.87 ± 0.12) classification using alternative heads, while genotype prediction rose from wF1 = 0.87 ± 0.05 to 1.00 ± 0.00, suggesting potential overfitting. External validation on a limited dataset for species yielded wF1 = 0.50 ± 0.02 (random wF1 = 0.32, internal wF1 = 0.68 ± 0.03). Grad-CAM indicated model focus on calcium transients, and SHAP analyses highlighted the decay phase as most informative, suggesting that targeted analysis may further increase performance.
Conclusion
Deep-learning models trained directly on raw microscopy images were non-inferior to, and partly outperformed, conventional feature-based algorithms, distinguishing biological and pathological states and detecting CPVT mutations via calcium-signaling patterns. This marks a first step toward an automated and scalable diagnostic pipeline for hereditary arrhythmia syndromes. External validation on a subset of species data confirmed robustness across laboratories, while CPVT and genotype results remain exploratory pending independent validation datasets.