Eighty consecutive patients (48 men and 32 women) who underwent Doppler echocardiography and cardiac CT within 3 months before aortic valve surgery were enrolled in Centro Cardiologico Monzino IRCCS (2018-2019). Exclusion criteria were AS due to rheumatic disease or radiotherapy, previous infective endocarditis, previous severe adverse reactions to an iodinated contrast agent, body mass index (BMI) ≥38 kg/m², and severely impaired renal function (estimated glomerular filtration rate < 30 ml/min/1.73 m²). In the final analysis, 56 age and sex-matched patients, with both bicuspid (BAV) and tricuspid (TAV) aortic valve morphology, were included. This study was approved by the Institutional Review Board of Centro Cardiologico Monzino IRCCS (n. R1348/20-CCM1418), and the participants gave written informed consent. The study protocol conforms to the ethical guidelines of the 1975 Declaration of Helsinki.

Cardiac CT (Revolution CT, GE Healthcare, WI) was performed using the following parameters: section configuration 256x0.625 mm, voxel size 0.625 mm, spatial resolution along the X-Y planes 0.23 mm, gantry rotation time 280 msec, and iterative reconstruction. A BMI-adapted scanning protocol (tube voltage and current) was used (11). All patients received a 50-mL bolus of contrast (Iomeprol 400 mg/mL, Bracco, Milan, Italy). Images were post-processed using dedicated software for AVC assessment (3Mensio Valves™, 3Mensio Medical Imaging, Maastricht, The Netherlands) (12) with manual correction. Contrast attenuation values (Hounsfield Units, HU) and contrast-to-noise ratio were measured at the level of the ascending aorta. Total AVF was assessed based on HU ranging between 30 and 350, adjusting the upper threshold by increments of 25 HU in either direction until blood pool was not highlighted, as previously published for non-calcific aortic valve tissue evaluation (2, 7). The calcium and fibrosis volumes were assessed and quantified using the ImageJ (Java-based image processing program developed at the National Institutes of Health) thresholding method ( Figures 1A–H ) (13). Indexed contrast-enhanced CT calcium volume and fibrosis volume were calculated by dividing the volume by the aortic annular area for each patient.

The workflow diagram of the deep learning analysis is reported in Supplementary Figure S1 . Each CT scan is composed of 256 or 224 slices (512 x 512 pixels) and stored in a DICOM object. In order to remove useless anatomical structures, such as aorta or ventricles, we selected the consecutive slices containing the aortic valve, which allows decreasing the computational burden and improving the overall network performances. Each selected slice was then used to perform a manual segmentation in order to generate the ‘ground truth’ images.

Thirty-eight out of 56 randomly selected samples were then used to train a U-Net model, with a default parameter, to segment the aortic valve. The remaining 20 samples served as the test set. The U-Net model was implemented with default hyperparameters, using the ‘Keras’ package; concerning the training step, we set the ‘Root Mean Squared (RMS) propagation’ optimizer to minimize the ‘binary cross-entropy (BCE) dice’ loss function, the number of epochs = 50 and the batch size = 10. The Dice Score was used to assess the segmentation accuracy, comparing the predicted mask and the ground truth.

Finally, the predicted masks were combined with the corresponding starting frames in order to extract the pixel intensity (Hounsfield Unit) of the aortic valve and, then, calculate the fibrosis and calcium volumes. The Pearson’s correlation index (r_p) allowed assessing the consistency between the automatic predictions and the calculations of fibrosis and calcium volumes, made by human operators.

Gene expression data were retrieved from Gene Expression Omnibus (GEO) public repository [GEO accession number: GSE102249 (14)]; this dataset collects the transcriptomic analysis, performed by microarray technology, of 240 valve tissues, coming from 120 males and 120 females. We re-annotated the Illumina probe ID, using the Ensembl human gene annotation (GRCh38, version 97); then we removed outliers and normalized data by log2 transformation and quantile normalization. The analysis was performed by R (v. 3.6.3).

Gene set enrichment analysis (GSEA) was conducted using GSEA pre-ranked tool (v 4.0.3) (15) and the biological functions were inferred by taking advantage of gene ontology biological processes (GO-BP), Reactome, WikiPathways, and BioCarta databases. To reduce redundancy and visually interpret GSEA results, the most significant GO-BP/Reactome/BioCarta pathway gene sets were identified by a false discovery rate (FDR) < 0.1. The network was drawn through the Enrichment Map software v.3.3.2 (16) implemented as a plug-in in Cytoscape v.3.8.2 (17).

Cellular composition of stenotic aortic valve tissue in women and men was inferred by the xCell R package (18), starting from the gene expression dataset of 240 samples. xCell raw scores were generated and transformed applying the spill-over compensation. Then, we assessed the association between the xCell scores and sex by Wilcoxon rank-sum test. The log₂ fold-change (logFC) of each cell type scores between men and women was computed and the pValues were adjusted with FDR correction. It is important to note that xCell results are expressed in terms of enrichment scores. Therefore, any inference on the cell counts or proportion of cell types is not provided.

The data were analyzed using IBM SPSS statistic 26 software. Continuous variables were expressed as mean ± standard deviation (SD) or, when variables had a skewed distribution, as median and interquartile range, while categorical ones were reported as frequency and percentage. Between-group differences were evaluated by Student t-test and by Pearson Chi-square (χ2) test. A pValue < 0.05 was considered to be statistically significant. Unpaired non-parametric Mann-Witney test was used to evaluate differences in indexed aortic calcification (iAVC) and fibrosis (iAVF).

Differential expression analysis between male and female samples was performed through the limma R/Bioconductor package (19). This statistical analysis was performed in the R environment v3.6.3. We deemed genes as significantly different at an FDR-adjusted pValue < 0.05. The robustness of the differential expression analysis results was assessed by exploring the histogram of the pValue distribution.

We studied 56 patients (28 men and 28 women) matched for age (71 ± 11 years) with severe AS (mean gradient 47.4 ± 13.6 mmHg, aortic valve area 0.82 ± 0.23 cm², and ejection fraction 64 ± 7%). Table 1 summarizes the patients’ characteristics. In brief, there was no difference between men and women in major cardiovascular risk factors, valve phenotype (17 BAV and 39 TAV in total), nor pharmacological treatment. As expected, men had higher body surface area and annulus diameter than women (1.89 ± 0.14 vs. 1.67 ± 0.17 m², respectively; p < 0.001) and (36.1 ± 3.0 vs. 31.9 ± 2.8 mm, respectively; p < 0.001), while women had lower aortic valve area than men (AVA; 0.74 ± 0.2 vs. 0.91 ± 0.2 cm², respectively; pValue = 0.007) but similar indexed AVA (0.44 ± 0.15 vs. 0.48 ± 0.12 cm²/m², respectively; pValue = 0.262).

Our results confirmed that, for a same hemodynamic severity in both sex, women had significantly lower iAVC compared to men (36 [95% Confidence Interval (CI): 27-49] vs. 72 [95% CI: 48-114] mm³/cm², respectively; pValue = 0.011). Furthermore, we showed that, in severe AS patients, the fibrotic tissue composition (i.e., iAVF) of the valve leaflets was significantly higher in women compared to men (83 [95% CI: 65-111] vs. 63 [95% CI: 45-82] mm³/cm², respectively; pValue = 0.026; Figures 1I, J ). Finally, the fibro-calcific ratio, which indicates the predominance of valve fibrosis if > 1.0, was significantly higher in women compared to men (2.57 [95% CI: 1.16-3.62] vs. 0.78 [95% CI: 0.43-1.95], respectively; pValue = 0.0018). Interestingly, the significant difference in iAVF volume between men in women were observed only in TAV morphology ( Supplementary Figure S2 ).

As the identification of valve in CT scans is a crucial step and could depend on the human operator experience, we validated our findings training a deep learning model (U-Net) to automatically recognize valve area in each frame without prior knowledge and calculate calcium and fibrosis volumes ( Figures 2A–C ). As shown in Figure 2D , we observed a high level of overlapping (Dice Score = 0.93+-0.09) between 400 predictions of valve area, performed by a human operator and the U-Net. Moreover, a Pearson’s correlation index r_p = 0.82 in 10 females and r_p = 0.93 in 10 males was gained comparing the calcium volume calculated by human (x-axis) and obtained by the deep learning automatic framework on the test set (i.e., the dataset containing images never seen by U-net during training), while a r_p = 0.96 in 10 females and r_p = 0.93 in 10 males was gained comparing the fibrosis volume calculated by human (x-axis) and obtained by the deep learning automatic framework on the test set ( Figures 2E–H ).

The public microarray-based gene expression dataset of 240 human stenotic aortic valves was analyzed to identify the genes and the biological pathways differentially expressed between women and men with severe AS.

The sex classification for each sample was performed based on the expression levels of XY-linked genes. In particular, Ribosomal Protein S4 Y-Linked 1 (RPS4Y1) and X Inactive Specific Transcript (XIST) genes were considered sex-specific markers, to split 120 male and 120 female samples. During the differential expression analysis, the XY-linked genes and pseudogenes were filtered out.

We observed that 3105 genes were significantly regulated between samples of women and men ( Figure 3 ). Setting a threshold of 0.2 log₂ fold-change (logFC), 80 genes were up-regulated in women, while 173 genes were up-regulated in men ( Figure 3A and Supplementary Figure S3 ). The heatmap ( Figure 3B ) shows that these 253 differentially expressed genes separate the specimens by sex. The top ten up-regulated genes in women and men are listed in Tables 2 and 3 , respectively.

A total of 4979 pathways characterizing women and men’s aortic valves were identified using GSEA ( Supplementary Tables S1 , S2 ). To enhance the interpretation of the results, we clustered the most significant pathways (FDR qValue < 0.1) into 33 macro-pathways (8 related to women and 25 related to men). The obtained enrichment network was drawn ( Figure 4 ), considering the computed NES and the gene counts ( Supplementary Figure S4 ).

The most representative pathways associated with women were related to the regulation of mRNA splicing, intraflagellar transport, post-translational modification (i.e., mannosylation), elastin fibers formation, insulin-like growth factor 1 pathway, and regulation of tyrosine kinase receptor.

In contrast, the most representative pathways associated with men belonged to the immune system response (innate and adaptive immunity), cell migration, and chemotaxis, constituting the wider groups of highly interconnected gene sets. Other molecular pathways enriched in men are inflammation (interleukin 1, 2, 4, 6, 8, 10, and 12), bone remodeling, cell proliferation, endothelial cell differentiation, cellular senescence, lipoprotein metabolism, reactive oxygen species, and reactive nitrogen species production, and regulation of platelet aggregation.

To reconstruct the cellular composition of the stenotic aortic valve from bulk transcriptome, we employed the xCell method to convert the gene expression profile to enrichment scores of several immune and stromal cell types across specimens (18).

Aortic valve leaflets harbor a heterogeneous cell population (20). Thus, to deconvolute cell types within the bulk transcriptional data, we selected 28 cell-type-specific gene expression profiles ( Supplementary Table S3 ).

The cell-type enrichment analysis showed that 14 cell types significantly differed between women and men (adjusted pValue < 0.05, Figure 5A ). Of note, as shown in Figure 5B , 4 cell types were identified as over-represented in women (logFC > 0) and 10 in men (logFC < 0).

The most representative cell types in women’s stenotic aortic valve leaflets were chondrocytes, fibroblasts, osteoblasts, and pericytes. Conversely, men’s stenotic aortic valve leaflets were characterized by the presence of i) innate immune system cells (monocytes, macrophages, and dendritic cells), ii) adaptive immune system cells (B and T lymphocytes), and iii) mesenchymal stem cells.