​

• Logistic Regression Machine Learning Model achieved 93% accuracy in predicting survival of patients with HF, while Automated Neural Network Deep Learning Model had 95% accuracy in predicting survival

​

• Key differentially-expressed hub genes that may play an important role in biological processes of HF: STAT3, CXCL10, TLR2, TIMP1, CCL2, CYBB, SPP1, & CD68.

​

• STAT3 had highest number of target miRNAs (148) & CCL2 had most target transcription factors (22)

​

​

�

Perisa Ashar*

Pratt School of Engineering, Duke University, Durham, North Carolina

Part 1: Role of AI (Machine Learning and Deep Learning) in Prediction of HF Deaths

A dataset containing various clinical parameters and demographic factors of 299 HF patients was analyzed (Chicco et al.; Ahmad et al.). These parameters and demographic factors included age, sex, presence of various conditions (anemia, diabetes, high blood pressure), presence of smoking, and follow-up time. Extreme outliers (>3X of IQR) were removed for data including creatinine phosphokinase levels, serum creatine, serum sodium, ejection fraction, & platelet count. SMOTE (Synthetic Minority Oversampling Technique) was applied to address imbalanced datasets. Various ML models (Logistic Regression, K-Nearest Neighbors Classifier (KNN), Support Vector Machine (SVM), Gaussian Naïve Bayes (GNB), Decision Tree Classifier, and Random Forest Classifier) with different train-test split ratios (10%-90%, 20%-80%, 30%-70%, 40%-60%, 50%-50%, 60%-40%, 70%-30%, 80%-20%, and 90%-10%) were used to predict survival. A sequential model automated neural network (ANN) was also developed and compared to the accuracy of machine learning models with a standard train-test split ratio of 80% : 20%. While training the ANN, there were a total of 300 epochs with a batch size of 40, and a validation split of 25%.

Part 2: Using Bioinformatics to Identify Key DEGS in HF Patients & miRNA/TF interactions

RNA-seq microarray dataset (GEO: GSE161472) was analyzed to investigate DEGs in LV free-wall tissue in patients with & without HF using the limma R package. Gene symbols corresponding to the top 200 entries of probe IDs with the most upregulated & downregulated genes (p <0.05) were used for enrichment analysis on StringDB & Cytoscape. Protein-protein interactions were used to identify overlapping hub genes from top 20 genes of 7 different topological analysis methods with CytoHubba: Maximal Clique Centrality (MCC), Maximum Neighborhood Component (MNC), Degree, and 6 centrality methods based on shortest paths (Radiality, Betweenness, Closeness, & Stress). Using miRNET, target miRNAs & transcriptions factors (TFs) were identified for each hub gene.

• Machine learning (ML) and deep learning can be employed to predict survival of patients with heart failure (HF) based on demographic factors and clinical parameters.
• Bioinformatics can be used to investigate differentially expressed genes (DEGs) in cardiac tissue of patients with and without HF.
• The combination of bioinformatics and artificial intelligence (AI) can be utilized to aid in prevention of deaths resulting from HF based on key parameters as well as cellular biomarkers for future treatment of HF.

​

• The logistic regression model (90:10 train-test split ratio) had the highest accuracy predicting survival of HF patients (F1-score = 93%)
• The next best models were KNN & Random Forest Classifier, which each had an accuracy F1-score = 89% at both the 80:20 & 90:10 train-test split ratios.

• Since machine learning models previously showed that the 90%:10% train-test split ratio with random seed of 0 was optimal for performance F1-accuracy score, the ANN was developed with a 90%:10% train-test split ratio with a random seed of 0. To optimize performance of ANN, other necessary parameters were as follows: batch size of 30 (with 300 total epochs) and a validation split of 25%
• ANN outperformed other ML models and had an overall accuracy rate of 95%.
• ANN had 98% accuracy on the training dataset and > 90% accuracy on validation

dataset during last (300^th) epoch.

​

​

Part 1: Role of AI (Machine Learning and Deep Learning) in Prediction of HF Deaths

​

Part 2: Using Bioinformatics to Identify Key DEGs in HF Patients & Investigate miRNA/TF Interactions Interactions

Fig. 7: String DB Network & Characteristics of corresponding to the top 200 entries of probe IDs with the most upregulated and downregulated genes from GEO dataset

Fig. 8: Top 20 HF Hub Genes Identified By 7 Different Topological Analysis Methods with CytoHubba

MCC

MNC

Radiality

Red = Hub Genes with Highest Scores (Most Gene Interactions)

Yellow = Hub Genes with Lower Scores (Fewer Interactions)

Degree

Betweenness

Closeness

Stress

• There were a total of 8 hub genes that overlapped with each topological analysis method: CCL2, TIMP1, TLR2, CXCL10, CYBB, SPP1, CD68, & STAT3

Fig. 9: Network of Hub Genes with Target TFs & miRNAs

Yellow = Hub Genes (Input)

Green = Transcription Factor Targets

Blue = miRNA Targets

• Artificial intelligence can accurately and effectively predict mortality in HF patients.
• Bioinformatics can identify potential candidate molecular biomarkers for characterization and help devise novel pharmaceutical interventions in patients with HF.

​