ZeCardioAI: combining zebrafish, AI, and xAI for an in-depth cardiac phenotyping platform

Abstract

Artificial intelligence (AI) integrated with high-throughput assays offers a powerful route to accelerate discovery in relevant biological models. Functional cardiac imaging is a prime application, where deep learning (DL) and explainable AI (xAI) can overcome limitations of traditional phenotyping methods, such as manual analysis, subjective interpretation, and low scalability. In cardiovascular research, the zebrafish model is highly valuable due to its translational relevance and accessibility for high-throughput applications. Here, we present ZeCardioAI, a computational platform combining zebrafish experimental advantages with DL and xAI methodologies. The platform automatically extracts comprehensive cardiac phenotypes from live imaging, achieving high precision while maintaining interpretability, critical for mechanistic insight and translational validation. ZeCardioAI, when applied to zebrafish models of dilated and hypertrophic cardiomyopathy (CM), detected subtle yet clinically relevant phenotypic differences. Machine learning classifiers achieved robust separation of disease from healthy phenotypes, and xAI revealed discriminative features aligning with established clinical markers. Our developments should prove valuable in addressing the unmet medical need in CMs to find new, specific treatments. The platform’s modular architecture supports future adaptation to diverse disease contexts beyond CMs, enabling large-scale, fully automated phenotyping at a throughput unattainable by manual approaches. ZeCardioAI establishes a new standard for AI-powered biological research, offering transformative potential for accelerating drug discovery, advancing precision medicine approaches, and deepening fundamental understanding of complex biological systems across multiple therapeutic areas.

Graphical Abstract

Graphical Abstract.

Introduction

Cardiomyopathies (CMs) are diseases of the heart muscle causing structural and functional abnormalities and are a leading global cause of heart failure. Affecting 3.73 million people worldwide, their prevalence rose by 41.5% between 1990 and 2019 [1], imposing a major healthcare and socioeconomic burden. The two most common forms are dilated cardiomyopathy (DCM) and hypertrophic cardiomyopathy (HCM).

DCM is characterized by left ventricle dilation with wall stretching and thinning, resulting in systolic dysfunction due to weakened cardiac contraction [2]. HCM presents excessive left ventricular wall thickening, decreasing chamber volume and impairing diastolic function [3].

CM disease biology is complex due to (i) the complex genetic architecture; over 20 sarcomere and sarcomere-associated genes are known to cause HCM, while DCM involves >40 genes affecting multiple cellular structures, including the sarcomere, cytoskeleton, and nuclear envelope. Even within the same gene, different mutations can lead to distinct disease patterns, and the same mutation can manifest differently among individuals and families. These effects are further shaped by environmental modifiers, making genotype–phenotype relationships difficult to predict [4–7]. (ii) the impact of disease progression; How CMs develop over time can be highly variable, influenced by additional genetic background, environmental modifiers, and other yet-unknown contributors [8]. Suitable biological models and analytical frameworks to enable precision medicine approaches for CM patients are currently missing.

Current therapeutic solutions for CMs are often limited to surgical interventions and palliative heart failure management rather than addressing specific underlying disease mechanisms [9, 10]. Two emerging targeted therapies show promise: Mavacamten, recently approved for HCM treatment [11], and Levosimendan, currently in clinical trials for DCM [12]. However, their full effects and mechanisms are still being elucidated. This therapeutic gap highlights an urgent need for more effective, mechanism-based treatments for CMs.

Zebrafish has emerged as a powerful translational animal model for cardiovascular research and high-throughput drug discovery. Its small size and optical transparency enable sophisticated live microscopy, while high progeny rates support extensive screening [13]. Furthermore, zebrafish share 82% genetic homology with humans for orthologous disease-related genes, ensuring biological translatability and clinical relevance [14]. The genome is readily editable via CRISPR/ Cas9 for precise modeling of genetic diseases, and fluorescent transgenic lines specifically label cardiovascular components for detailed in vivo analysis [15]. In vivo imaging of the zebrafish heart has proven particularly useful to address off-target cardiotoxic effects of drugs [16].

Traditional zebrafish cardiac analysis relies on fractional shortening, a 1D proxy for chamber volume variation derived from kymographs along manually drawn axes [16]. While valuable, its subjectivity and limited dimensional scope cannot report subtle functional abnormalities. Artificial intelligence (AI)-assisted image processing offers a means to overcome these limitations, providing objective and scalable quantification of cardiac dynamics.

As such, AI has become a valuable tool in clinical imaging, enabling automated volumetric analyses that extract detailed cardiac metrics from large datasets while reducing experimenter bias [17–19]. However, limited research applies these methodologies to zebrafish, and critical parameters like myocardial strain, an early detector of DCM and HCM in humans [20], remain unmeasured in zebrafish larvae.

AI approaches are valuable in cardiac phenotyping, and to ensure transparency and validate model decisions from a biological perspective, explainable AI (xAI) has become essential [21]. xAI methods like SHAP (SHapley Additive exPlanations) quantify individual feature contributions to predictions, enabling a clearer understanding of model decisions and identifying consistent biological patterns and clinical biomarkers [22].

Despite technological advances, identifying precise CM therapies remains challenging due to limited understanding of disease progression. Current diagnostic approaches (echocardiography, MRI), while providing valuable clinical metrics, may not sufficiently detect early CM stages [23]. This shortcoming, coupled with the complex genetic architecture and variable clinical presentations, necessitates innovative approaches for drug discovery and target identification.

To address these challenges, we developed ZeCardioAI, a fully automated framework for CM phenotyping that supports accurate model characterization, high-throughput drug screening, and target discovery. ZeCardioAI builds on the ZeCardio Screening Platform (ZeCardio1.0) [16], a semi-automated tool for cardiotoxicity prediction from in vivo imaging. While ZeCardio1.0 relied on manual annotation and was limited in scope, ZeCardioAI now enables the computation of key physiological metrics such as AV delay, contraction and relaxation times, and myocardial strain, in addition to standard cardiac parameters. It operates without user interaction and is suitable for both local and cloud-based execution. In addition, it provides automated organization, indexing, and large-scale analysis of experimental datasets, which allows to obtain high-quality datasets for model training.

The aim of this work was to develop ZeCardioAI, an in-depth cardiac phenotyping platform based on zebrafish CM models. The platform integrates fluorescence microscopy for real-time cardiac imaging with deep learning(DL)-based chamber segmentation and border tracking, enabling computation of both conventional and advanced functional parameters, including strain. These quantitative metrics were used to train machine learning (ML) classifiers that distinguish dilated and hypertrophic CM models from wild-type (WT) larvae, while xAI identified the features most relevant to classification. By combining automated imaging, advanced cardiac quantification, ML, and interpretability, ZeCardioAI provides a systematic framework for in vivo CM characterization and high-throughput drug screening, advancing precision medicine in CMs and offering translational potential for other disease areas.

Materials and methods

Zebrafish husbandry and mutant generation

Zebrafish were maintained under standard conditions at 28.51C with a 14 h light/10 h dark cycle. Embryos were obtained by natural mating and maintained in E3 medium at 28C until 120 hours past fertilization (hpf). Mutant (MUT) lines for nexn (DCM) and mybpc3 (HCM) were generated in the background Tg[myl7:GFP] [24] using CRISPR/Cas9 with two sgRNAs per gene. Stable lines were established by outcrossing and validated via PCR and gel electrophoresis. See Supplementary Materials.

Zebrafish larvae preparation and imaging

At 120 hpf, larvae were treated with PTU and immobilized in tricaine for imaging. High-throughput cardiac imaging was performed using the ZeCardio Screening platform previously described [16], combining a VAST BioImager (Union Biometrica) [25] and Leica DM6B microscope. Larvae were oriented at 144 (0 being the dorsal position and rotation always towards the left body side), and time-lapse videos of beating hearts were recorded at 100 fps for 20 s using 10 magnification. See Supplementary Materials.

Video quality control and preprocessing

Experiment quality control involved checking metadata for completeness and integrity and then subjecting the videos to a quality check using an ML binary classifier. The quality check was performed on a single frame sampled near the end of each video to enable rapid assessment and capture potential late-acquisition artifacts (e.g. a larva drifting out of frame). This classifier was trained via five-fold cross-validation on a balanced dataset of manually selected suitable and unsuitable images (namely out-of-focus or overexposed, cropped, not detected, badly oriented, deformed, corrupted, or pigmented), also with balanced representation across defect types. To ensure generalization, a 10% random, balanced validation set was held out, maintaining a balanced selection of the classes (including subclasses within the defective class), and selected from different batches generated in experiments performed on different days. Evaluated models included RegNetY_400MF, ResNet18, SqueezeNet 1.1, xResNet18, and XResNet34 [26–28].

Videos that passed the quality control were cropped to pixels. A binary heart mask was generated by the Sobel filter [29]. The mask was used to compute the centroid of the heart, which was then served as the cropping reference point. Heart orientation was standardized (anterior to the right and dorsal to the top) by image flipping. Videos were downscaled to 8 bits and saved as uncompressed.avi files. See Table S1 in Supplementary Materials for the number of experiments and videos that passed the quality control.

Annotation and feature extraction for cardiac function analysis

The following metrics were extracted from zebrafish heart videos: beats per minute (BPM), ejection fraction, arrhythmias, contraction and relaxation times, atrioventricular (AV) delay (time between atrial and ventricular beats), chamber diameters and volumes, and ventricle strain (the deformation of the cardiac tissue during contraction and relaxation) and strain rate. Ventricle and atrium semantic segmentation was performed, saving the mask coordinates at each frame. From the signal of the mask areas over time, the BPM was computed via its discrete Fourier transform. The AV delay was computed as the cross-correlation between the atrium and ventricle signals. Chamber diameters were derived from mask diameters, and volumes approximated using the biplane method of disks [30]. End-diastolic and end-systolic points, identified by signal peaks/valleys, enabled extraction of all cardiac metrics except strain.

For strain analysis, since the data consist of 2D fluorescence microscopy projections rather than Doppler or speckle-tracking echocardiography commonly used for strain analysis [30], two new types of strain were defined: boundary segmental strain, and boundary radial strain. The ventricle border was divided into four 5-marker point regions, with markers automatically tracked. Boundary segmental strain was calculated as the Lagrangian strain [30] (lengthening/shortening relative to maximum length) of each 5-point segment. That is, for each segment, its Lagrangian strain is , where is the segment’s length at time and is the reference maximum segment length (end-diastole). For the boundary radial strain, we first defined a fixed center point C in the ventricle. For each point P in one of the aforementioned 5-point segments, we calculate the Lagrangian strain of the segment P-C. The boundary radial strain for that segment is the average of these individual strain values.

Further details and custom Python code to extract the clinical metrics from the zebrafish heart videos is provided in the Supplementary Materials.

Cardiac chamber segmentation

The DeepLabV3 convolutional neural network was employed for automatic segmentation of the zebrafish heart chambers [31]. A training dataset was built of manually annotated key frames extracted from videos capturing different cardiac cycle phases. Three trained scientists independently annotated the dataset, split into three subsets, and a consensus protocol developed with cardiac biology experts was followed to ensure consistency.

To render the segmentation model invariant to rotation, translation, and deformation, data augmentation was performed including random translations in the range of pixels along x–y axes, random rotations within degrees, and random elastic deformations [32]. Transfer learning was applied by initializing DeepLabV3 with weights pretrained on a subset of the COCO train2017 dataset [33], specifically on the 20 categories included in the Pascal VOC dataset [34].

The model was trained for nine epochs with an Adam optimizer, learning rate of , mean squared error loss, and a batch size of four. In total,20% of the data was used for validation, ensuring that all frames from the same video were assigned to the same split to prevent data leakage. Performance was evaluated using Dice similarity coefficient, Hausdorff distance, and balanced average Hausdorff distance [35], with Euclidean distance as the ground distance for Hausdorff metrics.

Border local marker tracking

To compute heart strain, 20 equidistant ventricle border markers (a density selected to balance spatial resolution with annotation efficiency) were automatically tracked across video frames using DeepLabCut (DLC), a DL-based pose estimation software [36] (version 2.3.4). Annotation was performed on videos by two trained scientists by selecting representative frames capturing the full cardiac cycle.

A ResNet-50 model was used as backbone network [26, 37], using default DeepLabCut parameters and training for 100 000 iterations. Performance was evaluated by comparing DLC labels to manual annotations (average pixel distance).

Feature selection and machine learning classification

Univariate analysis and feature selection

Data preprocessing was done in the following order to ensure biological validity and cross-experiment comparability: automated removal of physiologically implausible samples (e.g. negative ejection fraction), normalization of contraction/relaxation times and AV delay by heart rate, batch-wise normalization of all features relative to WT controls to reduce inter-experiment variability, and final z-score standardization, with scaler parameters saved for consistent application to future data.

Feature distributions and independent t-tests identified inconsistencies and non-significant variables (-value >.05). However, subsequent analyses were performed both with and without non-significant variables to assess whether they contributed predictive signal. From feature pairs with a Pearson correlation coefficient > 0.8, one feature was removed.

Machine learning classification

For each gene, ML algorithms (logistic regression, random forest, support vector machine, multi-layer perceptron, XGBoost, AdaBoost, and Gaussian Naive Bayes) were evaluated to classify WT or MUT data derived from the imaging (a list of variables with mean values is provided in Table 1).

Table 1.

Baseline characteristics of heart metrics in dilated and hypertrophic models (nexn and mybpc3), reporting mean standard deviation and independent samples t-test P-values for mutant versus wild-type comparisons, with only global strain metrics shown (except top SHAP strain variables) and bold values indicating features used in each classifier.

Heart metric	nexn ()	nexn − / − ()	P-value	mybpc3 ()	mybpc3 − / − ()	P-value
Atrial beats per minute (1/min)
Ventricular beats per minute (1/min)
Ventricular mean ejection fraction (%)
Ventricular mean contraction time (s)
Ventricular mean relaxation time (s)
Atrial mean contraction time (s)
Atrial mean relaxation time (s)
AV delay (s)
Atrial mean mask area at end-systole (m)
Ventricular mean mask area at end-systole (m)
Atrial mean mask area at end-diastole (m)
Ventricular mean mask area at end-diastole (m)
Atrium maximum diameter (m)
Ventricle maximum diameter (m)
Atrium minimum diameter (m)
Ventricle minimum diameter (m)
Global mean boundary segmental strain (%)
Global peak boundary segmental strain (%)
Global mean segmental strain rate (%/s)
Global peak boundary segmental strain rate (%/s)
Left-right wall segmental strain delay (s)
Global mean boundary radial strain (%)
Global peak boundary radial strain (%)
Global mean boundary radial strain rate (%/s)
Global peak boundary radial strain rate (%/s)
Left-right wall radial strain delay (s)
Ventricular peak left boundary segmental strain (%)
Ventricular mean bottom-left boundary segmental strain (%)
Ventricular mean left boundary radial strain (%)
Ventricular peak left boundary radial strain rate (%/s)
Ventricular mean bottom-left boundary radial strain (%)
Ventricular mean right boundary radial strain (%)
Ventricular peak right boundary radial strain rate (%/s)

A nested -fold cross-validation, maintaining class distribution, was used for model selection, with inner folds for hyperparameter tuning and outer for evaluation. Performance was assessed via precision, recall, F1 score, and area under the receiver operating characteristic curve (AUROC), averaged across folds.

Explainable artificial intelligence and phenotypic clustering

To ensure interpretability and biological plausibility of the classifiers, we used SHAP, an xAI technique used to quantify feature contributions [22]. The computed SHAP values were visualized using beeswarm plots, highlighting the most influential features for classification. The top five SHAP-ranked features were further used as input for Uniform Manifold Approximation and Projection (UMAP) [38] to visualize clustering in a 2D representation (parameters: 50 neighbors, 0 minimum distance, and Euclidean distance). Abnormal data points (outliers in UMAP, defined as points lying in regions of overlap on the genotype-specific clusters) were manually selected for study. They were compared with baseline ranges using the top 10 SHAP features, and verified with data ranges and ML prediction scores/probabilities. Outlier analysis enabled biological interpretation of genotype–phenotype mismatches, technical artifacts, or natural variability, reinforcing the robustness and interpretability of the classification framework.

Institutional review board statement

The animal study was approved by Generalitat de Catalunya “Direcció General de Polítiques Ambientals i Medi Natural” under Reial Decret 53/2013, d’1 de febrer, referential procedure number 11925, under the title “Evaluación de la función cardíaca en pez cebra.” The study was conducted in accordance with the local legislation and institutional requirements.

Results

The goal of this research was to develop an AI-based phenotyping platform which would enable a deep characterization of zebrafish cardiac disease models, allow the validation of new therapeutic targets, and the assessment of the efficacy of new potential drugs, using HCM and DCM as a case study for platform development and proof of concept. For reference, Table S5 in Supplementary Materials provides an overview of the experimental setup and analysis pipeline of ZeCardioAI, alongside a comparison with recent zebrafish cardiac analysis frameworks.

Nexilin1 and Myosin binding protein C3 zebrafish knock-out models facilitate the study of cardiomyopathies in vivo

To model DCM and HCM, zebrafish transgenic lines carrying nexn and mybpc3 mutations were used, respectively. These lines were selected from a broader set of generated lines based on (i) the strength of their cardiac phenotypes, (ii) their viability and (iii) their suitability for high-throughput analysis (data not shown). NEXILIN encodes an actin binding protein and component of the Z-disk, essential for calcium handling and Z-disk and sarcomere stabilization, thereby ensuring optimal muscle contraction. Dysfunctions in Nexn have recently been strongly linked to DCM [39, 40]. MYBPC3 encodes cardiac myosin-binding protein C, a key sarcomeric regulator, and its mutations, especially truncating variants, are a major cause of HCM by disrupting contractile function and promoting myocardial thickening, thereby impairing cardiac relaxation [41, 42].

Homozygous MUT fish were created and grown to adulthood for nexn and mybpc3. Those MUT fish (and their WT cousins) were incrossed to produce the respective MUT and WT progeny, containing either both +/+ or no − / − functional allele for respective gene. The use of − / − parents ensured a complete loss of function. All zebrafish models were created in the transgenic background of Tg[myl7:GFP] [24], which expresses GFP in cardiomyocytes (Fig. 1A) allowing for in vivo live imaging to assess heart morphology and function. Neither reduced viability nor macroscopic morphological defects were observed in homozygous mutants of nexn and mybpc3 (Fig. 1B). Using the previously developed ZeCardio HTS platform [16], we observed expected alterations, namely increased ventricular size at end systole for nexn, decreased ventricular size at end diastole for mybpc3, and reduced cardiac output in both models. To enable deeper phenotypic characterization, we developed ZeCardioAI as a successor platform, extending the analytical capabilities of the original pipeline.

Figure 1.

Hypertrophic and dilated CM models and datasets. (A) GFP expression in the zebrafish heart at 5 dpf. The orientation for optimal view of the heart is quasi lateral, at 144 (0 = dorsal (D)). Note how both of the heart chambers lie approximately in the same imaging plane. (B) Lateral view of dilated (nexn) and (C) Hypertrophic (mybpc3) CM zebrafish knockout models (− / −) and their controls (+/+) at 5dpf in lateral (L) view imaged with the VAST onboard camera. dpf: days post-fertilization, scale bar 500µm.

ZeCardioAI platform: an innovative artificial intelligence framework for in vivo cardiac phenotyping and disease classification

We developed ZeCardioAI, a DL-powered analysis platform to achieve detailed assessment of heart metrics.

Hearts of zebrafish larvae were imaged and the datasets were subjected to ZeCardioAI platform’s first component addressing data quality assessment. A ResNet18 classifier was selected as the best performing model to discard poorly imaged hearts (Fig. S1 in Supplementary Materials). It was trained on 800 manually selected high- and low-quality images (400 each, from which 56 corresponded to out-of-focus, 60 to overexposed, 50 to cropped, 50 to not detected, 50 to badly oriented, 41 to deformed, 50 to corrupted, and 43 to pigmented), and achieved an accuracy of 0.98 0.02 (validation accuracy: 0.99). This ensured using high-quality datasets for subsequent model training. Tables S2 and S1 in Supplementary Materials show performance metrics for all models tested, and the total amount of zebrafish imaged and how many passed the initial quality check.

Cardiac chamber segmentation allows for refined chronotropic assessment

To enable unbiased and detailed characterization of myocardial contraction dynamics, we developed a DeepLabV3-based segmentation model for automatic tracking of full cardiac chamber contours (Fig. 2B). The model was trained on 494 manually annotated images from 107 videos, ensuring generalization across diverse cardiac phenotypes (see Fig. S2 in Supplementary Materials).

Figure 2.

Cardiac chamber segmentation model and ventricle border marker tracking model. (A) Zebrafish heart at 120 hpf, imaged via brightfield fluorescence microscopy. (B) atrium (top chamber) and ventricle (bottom chamber) outlined as marked in the annotation tool. Note how the ventricle outline does not encompass the bulbus arteriosus (arrowhead). (C) Semantic segmentation of a nexn and nexn larva heart. A still from a movie shows the heart and segmented chambers and the area change over time of each chamber mask (green, atrium; blue, ventricle). Note the comparable beat rate but reduced area changes in both chambers in the nexn larva. (D) 20 equidistant points along the ventricle border were annotated excluding the bulbus arteriosus. (E) Boundary segmental strain was computed for four segments (each made up of five equidistant dots). For each segment, the boundary segmental strain was measured by computing their Lagrangian strain. (F) Boundary radial strain was computed for the same four segments as the segmental strain by calculating the Lagrangian strain for each point to the center point of the chamber (marked as a dot). The individual strains were then averaged to denote the boundary radial strain of each segment. (G) Line plot of the mean boundary radial strain over time (three heartbeats) in representative nexn and nexn larvae. Note that the boundary radial strain of different segments is different and is reduced about half in the nexn larva.

DeepLabV3 was chosen for its ability to capture multiscale contextual information while maintaining high-resolution feature maps, improving boundary precision, a key aspect to obtain smooth transitions across frames. To improve generalization and reduce the need for large amounts of data and long training times, we applied transfer learning and data augmentation. In particular, elastic deformations proved effective in capturing the diversity of cardiac morphologies in zebrafish larvae.

The model achieved strong segmentation accuracy for both chambers, with test set Dice scores of 0.98 0.003. Hausdorff metrics showed a maximum deviation of 3.5 0.9 pixels, suggesting only minor localized errors. The balanced average Hausdorff distance further confirmed strong boundary alignment with low variability (0.028 0.007). Overall, the model produced reliable and smooth segmentations frame-to-frame (Fig. 2C, data for mybpc3 not shown). From chamber masks and their area changes over time, several metrics were extracted: heart rate, chamber diameter and approximate volume at end-systole and end-diastole, AV delay, and percentage of arrhythmic beats and parameters such as ejection fraction and contraction and relaxation times could be derived (Table 1).

For reference, automated chamber segmentation required 2 min per heart (2000 frames), whereas manual segmentation of only end-systolic and end-diastolic frames (300 frames) typically required 2.5 h per heart. Measuring was done using NVIDIA Tesla T4 GPU acceleration (CPU-only execution is 10 slower).

Ventricle border marker tracking allows for assessment of heart strain parameters

To measure heart strain, we used DeepLabCut (DLC) to track 20 equidistant markers along the ventricle border across frames (Fig. 2D). The network, trained on 10 videos with 15 annotated frames each, achieved an accuracy of 1.44 pixels (mean distance between DLC and manual labels).

Two types of strain were defined for this case study: boundary segmental strain and boundary radial strain. Both, nexn compared with nexn and mybpc3 compared with mybpc3 showed a reduction in strain parameters (Fig. 2G, data for mybpc3 not shown), consistent with findings in human patients [43, 44] and underlining the value of capturing local contractility impairments in biological models for DCM and HCM.

Automated ventricular border marker tracking required 11 s per heart video, encompassing 75 cardiac cycles (with NVIDIA Tesla T4 GPU acceleration, while 10 slower on CPU), compared with 8 min for manual tracking of one single cardiac cycle.

Univariate analysis confirms known human phenotypes in zebrafish cardiomyopathy models

To enable classification, we generated two balanced datasets: DCM-nexn (427 nexn, 427 nexn) and HCM-mybpc3 (186 mybpc3, 186 mybpc3). Data preprocessing involved removing outliers, with thresholds for heart rate (<50 or >350 BPM) and absolute strain (>40%), filtering physiologically implausible values.

For the DCM-nexn dataset, after feature filtering, the dataset retained 15 features, encompassing ejection fraction, contraction times, AV delay, ventricle size, and strain. These variables reflected hallmark DCM phenotypes, including enlarged ventricles, reduced systolic function, and impaired myocardial strain (Fig. 3A and B).

Figure 3.

Univariate analyses and strain profiles for DCM and HCM models. (A) Boxplots of the normalized features: minimum ventricle mask area, ventricular mean ejection fraction (EjF), and ventricular peak radial strain for nexn and nexn. (B) Polar plot of the mean radial boundary strain for the 20 tracked markers along the ventricle border for nexn and nexn. (C) Boxplots of the normalized features: maximum ventricle mask area, ventricular mean ejection fraction, and ventricular peak radial strain for mybpc3 and mybpc3. (D) Polar plot of the mean radial boundary strain for the 20 tracked markers along the ventricle border for mybpc3 and mybpc3. Statistical significance (independent samples t-test): ****-value<.0001. WT: wild-type; MUT: mutant.

For the HCM-mybpc3 dataset, the same preprocessing and filtering pipeline resulted in 23 retained features, including ejection fraction, ventricle relaxation time, AV delay, chamber sizes at end-systole and end-diastole, and strain. Retained features captured known HCM traits such as reduced diastolic ventricular size, lower ejection fraction, and diminished strain, consistent with patient phenotypes (Fig. 3C and D).

Summary statistics for both datasets are presented in Table 1.

Cardiomyopathy classification models separate conditions and explainable artificial intelligence identifies key phenotypic features

We trained and evaluated classifiers to distinguish nexn (WT) from nexn (DCM) larvae and mybpc3 (WT) from mybpc3 (HCM) larvae. Below we present a breakdown of model performance, feature contributions, and data visualization techniques used to further explore the classification results.

For DCM-nexn classification, XGBoost was selected as the winning model due to performance (Supplementary Material, Table S3) and interpretability advantage. XGBoost achieved the highest average accuracy (0.7727 0.0342) and F1-score (0.7620 0.0349), along with strong and balanced precision and recall, and high mean AUROC across the five splits (0.8404 0.0394). Train, test, and validation ROC curves are provided in the Supplementary Materials (Fig. S3). XGBoost showed robust performance in distinguishing nexn from nexn zebrafish, with 85% accuracy for nexn and 77% for nexn on the testing dataset, and mirroring the results on the validation set, indicating no overfitting. Additionally, its tree-based nature allows for more efficient SHAP analysis [45], an xAI technique to interpret the model’s decisions.

SHAP analysis identified ventricular mask area at end-systole, radial strain at right and bottom-left boundaries, AV delay, and average atrial contraction time as top predictors (Fig. 4B), consistent with DCM phenotypes, validating the animal model and the XGBoost classifier. Consistently, retraining the classifier using only the top 10 SHAP-ranked features yielded performance metrics comparable to those obtained with the full feature set (Supplementary Material, Table S3). UMAP embeddings of those features (Fig. 4A) showed good WT/MUT clustering with only small overlap, which could indicate some phenotypic heterogeneity. Outlier analysis revealed genotype–phenotype mismatches and borderline cases near the decision boundary. Comparison with XGBoost prediction scores confirmed this ambiguity, as borderline cases showed scores near 0.5.

Figure 4.

Explainability of the DCM-nexn and HCM-mybpc3 WT versus MUT classifiers. (A) and (C) UMAP embeddings in 2D of the five most influential features identified by SHAP for nexn (Top, XGBoost classifier) and mybpc3 (Bottom, SVM classifier). The density plot of the UMAP points visualizes the distribution, highlighting the densest clusters and illustrating the spread of variability within each class. Density marginal distributions are shown on the corresponding axes. (B) and (D) SHAP beeswarm plots illustrating model decision-making for the 10 most influential features. Positive SHAP values indicate a stronger contribution toward the MUT classification, whereas negative values favor the WT class.

For HCM-mybpc3 classification, a Support Vector Machine (SVM) with radial basis function kernel was chosen as the best model (Supplementary Material, Table S4), with the highest average accuracy (0.773 0.071), precision (0.794 0.068), and mean AUROC across the five splits (0.854 0.044), along with competitive recall (0.732 0.099). Train, test, and validation ROC curves are provided in the Supplementary Materials (Fig. S4). SVM achieved 91% accuracy for mybpc3 and 86% for mybpc3 on the testing dataset, with similar validation results, indicating no overfitting or bias. SHAP analysis (Fig. 4D) revealed key features: AV delay, ventricular and atrial mask areas at end-diastole and end-systole, and ventricular ejection fraction, strain, and strain rate, all aligning with HCM phenotypes in patients and reinforcing the biological relevance. Similarly, retraining the classifier using only the top 10 SHAP-ranked features resulted in performance comparable with that obtained with the full feature set (Supplementary Material, Table S4). UMAP embeddings showed clear clusters with limited overlap (Fig. 4C). Outliers again corresponded to phenotype–genotype mismatches or inherent biological variability, supported by SVM decision function scores, with borderline cases showing values near zero. Notably, including non-significant variables did not affect the classification outcome and these did not rank among the top features in the SHAP analysis.

Overall, both classifiers effectively distinguished WT from MUT larvae using biologically relevant cardiac features, supporting their potential for assessing therapeutic effects in zebrafish CM models.

Discussion

We developed ZeCardioAI, a platform for high-throughput imaging and automated phenotypic analysis of in vivo cardiac function in zebrafish larval hearts. ZeCardioAI allows for characterization of zebrafish disease models, representing an important step towards precision medicine and safety understanding, since not only specific genetic mutations can be modeled but also the effect of environmental, dietary modifiers, and potential cardio-toxic substances can be assessed. The platform provides a means to gain mechanistic insight into CM progression and establishes the foundation for future assessment of therapeutic and repurposing candidates. ZeCardioAI extracts clinically relevant cardiac parameters through DL-based chamber segmentation and ventricle border tracking, enabling automatic calculation of inotropic and chronotropic metrics, and ventricular strain. The developed ML classifiers (XGBoost for DCM, SVM for HCM) effectively distinguished WT from MUT larvae and SHAP analysis identified key features driving classification which align with known characteristics in human patients.

In disease biology research and drug discovery, animal model translatability to human is critical. Rather than preselecting features based on expected human phenotypes, we quantified a broad range of cardiac metrics to allow the ML algorithms to objectively identify which parameters are most discriminative for each specific genetic perturbation, acknowledging that while key disease markers are conserved, their relative contribution may differ between zebrafish and humans. We hypothesized that not all extracted features would be affected by the mutations, thus, the feature importance analysis served to detect biologically relevant phenotypes. The DCM-nexn model showed changes in ventricular size, ejection fraction, and strain that parallel clinical findings of systolic dysfunction and ventricular remodeling in DCM patients [46], including those with nexn mutations [47]. Notably, strain is emerging as an early diagnostic tool in patients with preserved ejection fraction but familial predisposition [20, 48] and, in our case, strain metrics were highly relevant, revealing subtle changes. AV delay, crucial for synchronized AV activation and optimal ventricular filling, was also among the top features in the DCM-nexn model identified by SHAP. Prolongation can cause suboptimal filling and diastolic mitral regurgitation and, while not a primary hallmark, is frequently observed in DCM patients due to chamber dilation disrupting conduction pathways, leading to bundle branch block, and higher heart failure risk [49]. Prolonged contraction time was also relevant, consistent with other DCM cellular models [50], and mice [51]. To further support the SHAP analysis, comparable classification performance was retained when retraining using only the top 10 SHAP-ranked features, indicating that SHAP successfully identified a set of biologically relevant parameters sufficient to characterize the nexn CM model.

The HCM-mybpc3 model recapitulated key aspects of human HCM, including altered conductivity, chamber dimensions, and contractility [52]. AV delay, while not a primary symptom, was the most influential feature, potentially due to myocardial disarray and fibrosis disrupting conduction [53]. Following the same approach as with the DCM-nexn model, comparable classification performance was observed when the model was retrained using only the top-10 SHAP-ranked features, further reinforcing the validity SHAP analysis. Prolonged relaxation time, though not among the top-10 features, was significantly increased, consistent with what is observed in HCM patients [54].

Overall, strain and chronotropic features such as AV delay and contraction/relaxation times emerge as sensitive, potentially early indicators. While abnormal strain has been shown to predict potential adverse outcomes [55], further research on AV delay as an early CM marker is needed.

The ability to detect clinically meaningful phenotypes underscores the value of zebrafish in cardiovascular research. Despite intrinsic differences, such as a single-chambered circulatory loop, absence of pulmonary circulation, and the capacity for heart regeneration, the zebrafish is well established for heart failure studies [56]. Its optical transparency, conserved signaling pathways, ease of genetic manipulation to mimic human mutations, similar heart rate, morphology, and function make it a powerful model for cardiac disease research [57].

However, cardiac imaging pipelines remain limited by imaging quality, fish line, and developmental stage. Moreover, to our knowledge, strain analysis has not been reported in larvae and has only recently been explored in adult fish [58, 59]. Our research expands on previous studies by leveraging zebrafish HTS with AI and xAI to improve phenotyping and model interpretability. While recent works also capitalize on the HTS potential of zebrafish [60, 61], ventricular geometry is simplified, and only few pipelines employ DL-based segmentation of both chambers, limiting their sensitivity to complex morphologies [17–19]. Although a direct quantitative comparison is not technically feasible due to substantial differences in input data, our review comparing recent similar works highlights that ZeCardioAI addresses a specific gap: the combination of high-throughput screening with advanced cardiac phenotyping, including strain analysis. Moreover, as emphasized across these studies, manual processing would be unfeasible at this scale. This further stresses the need for automation in cardiac analysis pipelines.

ZeCardioAI addresses these limitations with chamber-specific segmentation, regional strain analysis, and interpretable classification in a scalable, generalizable, and automated pipeline. Its throughput and interpretability make ZeCardioAI well suited for early-stage in vivo drug discovery, where sensitive phenotyping is essential for screening therapeutic candidates. We also note that this study focused on single-gene mutants, nexn and mybpc3, to isolate mutation-specific effects. While human DCM and HCM are genetically complex [4, 62], zebrafish enable future expansion to double mutants or inducible knockouts. The findings for the nexn and mybpc3 models, spanning morphology, strain, and conductivity abnormalities, align and contribute to the existing knowledge of CM pathophysiology, reinforcing the translational relevance of zebrafish models in advancing CV research.

In clinical settings, cardiac imaging often relies on 3D acquisition and established speckle-tracking techniques for strain estimation [20, 30]. In contrast, imaging in this study was limited to 2D, requiring proxy metrics in place of clinical standards. Furthermore, while speckle-tracking is applicable to adult zebrafish using high-frequency transducers to achieve the necessary axial and lateral resolution, the much smaller size of zebrafish larvae approaches the technical resolution limits of current ultrasound technology [63]. While we acknowledge that 2D imaging simplifies the heart’s 3D complexity and may overlook features along the depth axis, this approach represents a necessary trade-off to enable high-throughput screening. Implementing 3D or multi-angle acquisition would allow for even more detailed phenotyping but it would prolong imaging time, typically entail a loss of temporal resolution, and increase data processing load, rendering the platform unsuitable for large-scale cardiac screening. Accordingly, our bespoke boundary segmental strain serves as a 2D proxy for clinically established longitudinal strain, measuring myocardial wall shortening, while boundary radial strain approximates wall thickening and thinning typically assessed in parasternal short-axis echocardiographic views. Although these metrics lack full volumetric context, they successfully capture the contractility defects characteristic of CM, as reflected by our ability to accurately classify disease phenotypes and their high ranking in the explainability analysis. Still, 3D or multi-angle 2D imaging could be employed for lower-throughput additional validation. Moreover, inconsistent fluorescence patterns across the cardiac cycle precluded reliable speckle tracking. To address this, we implemented DL-based marker tracking to follow anatomical landmarks, enabling accurate strain calculation. The DLC network was trained on representative frames from 10 videos (totaling 150 annotated frames). This relatively small dataset size aligns with established DLC benchmarks, which show that transfer learning enables human-level accuracy with as few as 100–200 training frames [36]. Furthermore, unlike freely moving animal behaviors, our videos from zebrafish hearts present a constrained tracking challenge characterized by periodic, smooth motion and high-contrast fluorescent boundaries against a uniform background. These standardized imaging conditions significantly reduce the diversity of visual features the model must learn, allowing for robust generalization even with a small training set. Still, while DLC reports validation error on held-out annotated frames, frame-level validation is limited when the training data are small and temporally correlated. In practice, by qualitative visual inspection, we observed consistent tracking across independent datasets, validating the model’s ability to generalize to data unseen during training.

One key module of ZeCardioAI focused on data quality checks and standardization to obtain well-curated, high-quality datasets. We implemented a DL-based quality control model because manual video assessment is a major bottleneck in high-throughput screening, being time-consuming, repetitive, and prone to operator bias, thereby undermining end-to-end automation. Simpler automated metrics (e.g. intensity thresholds or blur measures) proved insufficient, as they failed to capture complex morphological artifacts such as bad orientation or cardiac deformation. Therefore, a CNN-based approach was chosen to robustly identify diverse quality failures. Nevertheless, some variability is unavoidable. For instance, WT and MUT samples showed partial overlap in UMAP projections. Some of the outliers could be traced to segmentation errors, while others reflected genotype-positive/phenotype-negative or genotype-negative/phenotype-positive cases, consistent with clinical CM observations [64, 65]. Additional sources of overlap likely arose from natural variation and dimensionality reduction constraints.

This variability is not necessarily detrimental. In fact, training on heterogeneous data may still support highly predictive models. In our study, XGBoost achieved strong performance in the DCM-nexn dataset, whereas its lower accuracy in the HCM-mybpc3 model may stem from limited data, as the algorithm typically requires larger datasets to generalize effectively. Expanding the mybpc3 dataset and refining curation will therefore be key to improving performance, particularly for downstream drug screening applications. Moreover, integration with transcriptomic datasets could further strengthen this framework by linking phenotypes to disease-driving pathways and prioritizing druggable targets, advancing mechanistic understanding [66]. Future efforts should also validate candidate compounds and gene targets through pharmacological treatments and CRISPR/Cas9-generated zebrafish mutants, while expanding datasets across additional genetic backgrounds.

Conclusion

ZeCardioAI’s combination of automated cardiac quantification and interpretable classification makes it well suited for HTS and drug discovery. Its modular design also supports adaptation to cardio-toxicity assays and broader cardiovascular research, positioning ZeCardioAI as a scalable tool for zebrafish-based translational studies. We anticipate that the integration of zebrafish and AI will be a transformative approach in the discovery of new therapies for complex diseases, such as CMs.

Finally, in alignment with the principles of open science and to ensure reproducibility, the complete ZeCardioAI framework is publicly available as a modular software suitable for local or cloud-based deployment (see Data availability section).

Key Points.

• Zebrafish genetic models for dilated and hypertrophic cardiomyopathy (CM) recapitulate clinical phenotypes.

• Fully automated, high-throughput fluorescent in-vivo imaging of zebrafish hearts allows for generation of large-scale and high-quality datasets for artificial intelligence applications.

• Specific deep learning models can successfully extract a plethora of metrics aiding the deep phenotypic description of CM disease models.

• Classifiers trained on heart metrics are able to distinguish healthy from diseased hearts with high fidelity, and explainable artificial intelligence enables the understanding of the parameter prioritization in the classification decision.

• The analytical capabilities provided in the ZeCardioAI platform can enable drug discovery in CMs and can be easily scaled up or expanded to other cardiovascular indications.

Author contributions

Ferran Arqué (Conceptualization, Methodology, Software, Validation, Writing, Review), Carole Jung (Conceptualization, Validation, Project Administration, Writing, Review), Beatriz Muriel-Moreno (Investigation, Validation), Laura López-Blanch (Investigation, Validation), Eduardo Saman (Methodology, Software, Validation), Davide DAmico (Funding Acquisition), Javier Terriente (Conceptualization, Funding Acquisition, Writing, Review), Paula Petrone (Conceptualization, Supervision, Writing, Review), and Sylvia Dyballa (Conceptualization, Methodology, Software, Validation, Project Administration, Writing, Review)

Conflicts of interest

F.A., C.J., B.M.-M., L.L.-B., E.S., D.D., J.T., and S.D. are affiliated with ZeCardio Therapeutics SL, a pharmaceutical company using zebrafish in cardiovascular disease modeling and drug discovery. F.A., J.T., and S.D. are affiliated with ZeClinics SL, a Contract Research Organization (CRO) using the zebrafish model for disease modeling, target validation, and drug screening.

Funding

This work was supported by the NextGenEU programme “Artificial Intelligence and Other Digital Technologies to Drive Their Integration into Value Chains” (NextGenEU-C005/21-ED 2021/C005/00149851); and the Ministerio de Ciencia e Innovación, Agencia Estatal de Investigación, “Ayudas para Doctorados Industriales” (DIN 2022-012817 to F.A.).

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author. The cardio-deeplab and cardio-deeplabcut Python packages are publicly available, along with the supporting code and classes used to run them and extract cardiac metrics from zebrafish heart videos. A Jupyter notebook demonstrating the extraction of cardiac metrics from an example video is provided for illustrative purposes. All resources are accessible at https://doi.org/10.5281/zenodo.17457010.