Occam’s Razor–Guided Neural Architecture Search for Biological Signal Analysis in Cardiovascular Disorders

Abstract

In the context of the classification of heart sounds using deep learning, challenges related to computational efficiency and implementation persist in settings with limited resources. Objective: The study proposes a comparative approach for the systematic evaluation of hybrid neural architectures. This approach employs the Neural Architecture Search (NAS) as an analytical tool guided by the principle of Occam's Razor. The purpose of this employment is to identify optimal models in terms of accuracy, complexity, and resource consumption. Methods: Contrary to conventional NAS methodologies that are characterized by the construction of architectures from the ground up, the present research utilizes pre-built models—namely, CNN-GRU, CNN-LSTM, SENet, EfficientNet, and MobileNet—to conduct a comparative analysis under specific hardware constraints. These constraints include, but are not limited to, maximum RAM occupancy, Flash storage capacity, number of MAC operations, and computational latency. Results: The analysis revealed that CNN-GRU models exhibit an exceptional balance between performance and structural simplicity. Using cochleographic representation, CNN-GRU achieved an accuracy of 0.9790, while maintaining an optimal Occam Score of 0.9623 among all evaluated architectures. Conclusions: The integration of the principle of simplicity within the NAS process enables the selection of efficient neural architectures applicable to TinyML devices. This, in turn, contributes to the development of lightweight, comparable, and sustainable cardiovascular diagnostic solutions.