A novel hybrid liquid neural network with time dependent neuromorphic and transformer encoder based reservoir dynamics for respiratory sound classification

Abstract

Chronic obstructive pulmonary disease is often misdiagnosed due to its similarity to other respiratory disorders, which delays early intervention. In liquid neural network reservoirs, leaky and spiking neurons enhance sensitivity to low-frequency variations while preserving the continuity of respiratory sounds. Time-dependent activation coefficients and adaptive leakage rates improve computational efficiency, stability, and memory retention. To further enhance the discrimination between normal and chronic obstructive pulmonary disease respiratory signals by capturing long-range dependencies and temporal dynamics, a transformer encoder module is integrated. In this study, respiratory sound recordings obtained from individuals with chronic obstructive pulmonary disease and healthy controls are subjected to Z-score normalization. The liquid neural network reservoir is configured to include both leaky and spiking neurons. To improve performance metrics, a timedependent activation coefficient is assigned to the leaky neurons, while a time-dependent leakage coefficient is added to the spiking neurons. To enhance the modeling capabilities, a transformer encoder block is integrated into the reservoir architecture. The accuracy values of the liquid neural network reservoir with leaky and spiking neurons are 97.70 % and 91.63 %, respectively. When a time-dependent activation coefficient is applied to the leaky neurons and a time-dependent leakage coefficient to the spiking neurons, the accuracy values increase to 99.84 % and 94.25 %, respectively. When the transformer encoder is integrated into the liquid neural network reservoir, the prediction accuracy values for the leaky and spiking neuron architectures are 98.19 % and 93.43 %, respectively. Furthermore, the robustness analysis revealed that the differences in accuracy values ranged from -0.94 to -4.83.