Artificial Intelligence Models Reveal Sex-Specific Gene Expression in Aortic Valve Calcification

Visual Abstract


SUMMARY
Male and female aortic stenosis patients have distinct valvular phenotypes, increasing the complexities in the evaluation of valvular pathophysiology. In this study, we present cutting-edge artificial intelligence analyses of transcriptome-wide array data from stenotic aortic valves to highlight differences in gene expression patterns between the sexes, using both sex-differentiated transcripts and unbiased gene selections. This approach enabled the development of efficient models with high predictive ability and determining the most significant sex-dependent contributors to calcification. In addition, analyses of function-related gene groups revealed enriched fibrotic pathways among female patients. Ultimately, we demonstrate that artificial intelligence models can be used to accurately predict aortic valve calcification by carefully analyzing sex-specific gene A ortic stenosis (AS) is the most common valvular heart disease requiring intervention in both men and women (1)(2)(3). However, male and female patients have a distinct phenotype of AS, which may lead to diagnostic difficulties in terms of evaluating aortic valve pathology in women (4)(5)(6).
Likewise, the invasive either transcatheter aortic valve implantation or surgical aortic valve replacement (SAVR) rates remain substantially lower in women as compared with men (7).
The pathophysiology of AS is characterized by an inflammatory process due to endothelial damage caused by an interplay of lipid accumulation, mechanical stress, leaflet thickening, osteogenic differentiation, and ultimately calcification (8)(9)(10). A high-pressure gradient across the aortic valve arises as the aortic orifice becomes narrow and leaflet stiffness increases (11). When symptomatic AS is left untreated, the condition is rapidly fatal, leading to an adverse prognosis and high mortality (12). Recently, it has been shown that both moderate and severe AS carries a similar mortality rate (13). Although increasing attention has been paid to differences between the sexes in regard to cardiovascular diseases, there is a substantial lack of studies investigating sex differences in AS from a mechanistic point of view (14,15).
Gene expression studies have identified several target genes of interest when analyzing different pathophysiological processes in AS (16-18). However, none of these specifically addressed sex-specific transcriptomic patterns in AS. In addition, while traditional inferential statistics methods can offer great insights in datasets with a good observation (n) to predictor (p) ratio, there are multiple challenges in working with high dimensional datasets, such as gene expression data (19). Most regression models are known to underperform when there are ambiguities in how to select relevant predictors and when n < 10p, especially in situations where there are multiple nonlinear relationships (20). In addition, these datasets present with sparsity related challenges leading to an exponential increase in required data needed to detect significant findings. One solution to this problem is turning to artificial intelli-

RESULTS
The baseline characteristics of the cohort stratified by sex are summarized in Table 1. Lower standardized differences were achieved for the continuous variables that were included in the propensity matching algorithm than for the categorical variables. By matching for CABG an acceptable matching for     Table 2).

DISCUSSION
In this study, several different approaches were used  After the selection of transcripts whose levels were significantly different between calcified and nondiseased valve tissue, we identified 149 genes that were differentially expressed between males and females across the different categories of valvular disease. Subsequently, these differentially expressed genes between males and females were used to construct models for predicting valve calcification.
Three major conclusions can be drawn from these analyses. First, whereas a logistic regression provided only poor prediction of calcification, all supervised ML models reached a prediction accuracy of 100%.
Second, although the relative importance of the transcripts identified by the AI models differed somewhat between the models, there were some genes that were consistently included in the models. and has been associated with calcification mechanisms in different tissues (31)(32)(33). Furthermore, many genes repeatedly identified by different AI algorithms to be preferentially expressed in female valves have

CONCLUSIONS
In summary, the results from the multiple approaches in this study indicate that there are multiple combinations of sex-specific genes that can achieve good predictability of valvular calcification.
By using different AI models, we were able to highlight the potential of these gene sets that might otherwise not be possible if only traditional statistics analyses and single models were used. The conclusions of the present study are that sex-specific