Cardiac Atlas based Phenotypes Extraction and Application in Biological Age Prediction (RF_TH_227)

The prediction of cardiac biological age has been a prominent research focus in the cardiovascular field, as it can be used to determine vascular age, a key indicator of cardiovascular disease risk. Traditional prediction methods primarily rely on conventional conventional biomarkers. In this study, we constructed a cardiac atlas based on end-diastolic short-axis CMR images from a healthy population and extracted high-throughput phenotypes.


Methods:

Atlas construction

This study received approval from the local institutional review board. All participants provided informed consent. A total of 736 individuals, representing a natural population without major diseases, were acquired using a 3T scanner (MAGNETOM Vida, Siemens Healthineers). The pipeline of cardiac atlas construction (Figure 1) include: 1) Cardiac segmentation: cine images were automatically segmented using a pre-trained biventricular segmentation network, followed by quality control and manual adjustments by an experienced radiologist. 2) Preprocessing and registration: Gaussian smoothing was applied, followed by rigid registration of the mesh images using the ICP algorithm. 3) Atlas construction: Using the LDDMM framework, rigidly registered images were refined through spatial registration and averaging. A template mesh representing the average cardiac shape was estimated, and atlas phenotypes, representing deformations from template to each subject, were derived using control point coordinates and momentum vectors in 3D space. Momentum represents the spatial displacement measures of variability within the population.

Cardiac age prediction

A random forest regression model was first trained using the 35 conventional biomarkers as input features to predict age. The model was then retrained by incorporating the momentum vectors as cardiac deformation features, comparing performance on the same training sets (656 samples) and test sets (72 samples). Initial screening was performed to exclude momentum vectors with low mean absolute displacement. Pearson correlation tests were used to evaluate the relationships between 35 biomarkers (atrial volume, ventricle volume, and wall thickness), combined biomarkers and atlas phenotypes (momenta) with age.

Results:

Correlations analysis

A total of 880 momenta were extracted from the 3D atlas, with 540 remaining after initial screening. Figure 2 shows that 29 out of 35 conventional biomarkers are significantly associated with age (p < 0.0001), and 185 out of 575 combined conventional biomarkers and momenta are significantly associated (p < 0.0001). The addition of atlas based momenta measurements slightly enhanced the significance of the correlation.

Prediction performance  

As shown in Figure 3, predicting age based solely on biomarkers yielded suboptimal results. The MAEs of the training and testing sets are 3.1 and 8.9, respectively.Incorporating momenta as additional input features improved prediction accuracy (The MAE of the test set decreased to 8.6). The regression scatter plot (Figure 3a) indicates better performance when momenta were included. Feature importance analysis (Figure 3b) revealed that 8 of the top 20 most significant features were derived from momenta, demonstrating their contribution to enhanced predictive performance.

Conclusion:

This study constructs a 3D cardiac morphological atlas, demonstrating that incorporating atlas-based phenotypes yields better biological age predictive performance than using biomarkers alone.