Artificial neural network models utilize chamber-specific predictors of cardiac fibrosis in ovariectomized and aortic-banded yucatan mini-swine
[EMBARGOED UNTIL 6/1/2023] Objective: Heart failure with preserved ejection fraction (HFpEF) is a disease associated with significant clinical pathophysiological heterogeneity in which maladaptive cardiac fibrosis, in both the right and left ventricles of the heart, plays a unique role in the manifestation of disease. Fibrotic remodeling quantified in this project occurs in a chamber-dependent manner on both sides of the heart. Extracellular matrix (ECM) remodeling is the core of this pathological process. The prevalence of HFpEF is greater in postmenopausal women with hypertension. Therefore, the goal of this study was to assess the role of female sex hormones on chamber-dependent differences i.e., left ventricle (LV) vs. right ventricle (RV), in ECM remodeling and regulation in a mini-swine model of pressure overload-induced heart failure (HF). To gain insight about the regulation of fibrosis in this model, biological inputs were measured in both the right and left ventricles and used as input variables in an artificial neural network model (ANN). This model will identify best predictors for experimental group status i.e., the combination of the loss of female sex hormone and/or pressure overload status, as an indication for the biological roles they play in the fibrotic remodeling process. Hypothesis: I hypothesized molecular markers involved in the bioregulation of the cardiac ECM can predict experimental group status in a chamber-specific manner. Methods: a) Animal model: An ovariectomy (OVX) model of surgically induced menopause was used to model the loss of female sex hormones. Separately, aortic banding (AB) was used to induce pressure-overload and mimic HFpEF. Animals that did not undergo ovariectomy were assigned to the intact (INT) groups and animals that did not undergo AB were assigned as control (CON). b) Data: 24 six month old female swine were categorized into 4 groups by ovariectomy and aortic-banded status: 1) Control, intact (CON-INT; n=6); 2) CON-OVX (n=5); 3) AB-INT (n=7) ;and 4) AB-OVX (n=6). c) Ninety-six biological measurements from both the LV and RV were considered including different mRNA, proteins, activity and/or abundance levels of various extracellular matrix components including structural proteins and regulatory pathways. d) Data preprocessing: Missing data were mean imputed and the min-max normalization method was used for all measures. One-way ANOVA models were used to identify mRNA or protein targets associated with group status for consideration in the ANN. Data were split into testing and training sets with one observation from each group (n=4 total) retained for later model testing i.e., 84 percent training and 16 percent testing e) Artificial neural network model: Measurements associated with group status were then used as input features in the ANN model. Multiple activation functions were considered. Different combinations of hidden layers and nodes within each layer were optimized. Cross-validation, confusion matrices, and F1 scores, percentage accuracy and balanced accuracy for each experimental group were used to describe the accuracy of the developed ANN model. Results: One-way ANOVA models indicated that in the LV, total collagen content, TIMP-1 mRNA, total JNK protein level, MMP-14 activity, MMP-2 activity and collagen I mRNA were associated with group status (p [less than] 0.1). In the RV, total collagen content and collagen I and III mRNA levels were associated with group status (p [less than] 0.1). These nine molecular markers were used to develop the ANN model. Cross-validation and confusion matrices indicate all nine targets formed a linear relationship predictive of group with an overall accuracy of 70.7 percent and F1 score of 0.81. Conclusion: Molecular mechanisms involved in the bioregulation of the ECM have analytical power to determine sex hormone and aortic-banding status in a pre-clinical model of pressure overload-induced HF. These findings indicate that nine biological measures could predict experimental group status in our pre-clinical swine model. Therefore, I identified these variables as potential biomarkers of fibrotic remodeling in a HFpEF phenotype with loss of female sex hormones and pressure overload. I also highlight the importance of these nine variables in the fibrotic remodeling process on both sides of the heart.