HIGH-PRECISION PHENOTYPIC ANALYSIS OF ECHOCARDIOGRAPHIC SIGNALS VIA ADAPTIVE STAGNATION-TRIGGERED WHALE OPTIMIZATION AND RANDOMIZED PCA

Abstract

Although echocardiographic video interpretation is one of the most important tasks to evaluate cardiac functional capacity, it is still resource-consuming, even with deep learning pipelines. This paper attempts to tackle the issue by applying a new hybrid optimization framework that applies Randomized Principal Component Analysis (RPCA) and Adaptive Stagnation -Triggered Whale Optimization Algorithm (AST -WOA) to dimensionality reduction and hyperparameter optimization, respectively. The first phase uses a SPP-ResNet50-Transformer backbone as the first one. It is based on this backbone that creates multi-scale spatiotemporal descriptors of echocardiographic sequences. RPCA then compresses the high-dimensional features, therefore retaining salient features and dropping redundant dimensions, thus decreasing the training overhead. AST-WOA then optimizes the features using an original meta-heuristic algorithm that transforms classical Whale Optimization Algorithm by adopting three new features: (1) OBL initialization to ensure diversity in populations, (2) non-linear cosine convergence factor, which is aimed at creating dynamically balanced trade, which controls the parameters of exploration and exploitation, and (3) stagnating convergence induced LEVy flight mechanism, through which reactivation of control parameters to encourage exploration is done selectively, particularly when convergence is stagnating. This self-aware optimization method is developed to avoid local minima and is especially useful in problems that have high computational complexity (as is the case of deep learning systems). The first illustrative example of the RPCA and AST-WOA frameworks was done using the EchoNet-Dynamic dataset as its experimental setting where it proved to be computationally efficient, provided better classification and ejection fraction predictions. The findings suggest that the suggested approach converges faster, is more robust to generalization and has a higher clinical interpretability than standard deep learning baselines and available meta-heuristic optimization algorithms.