Abstract
Real-time wearable sleep monitors process diverse biological signals while operating under tight energy and computation budgets. The existing algorithms are facing problems of high energy consumption due to separate hardware storage and computation units. In this work, textile-integrated in-memory neuromorphic computing electronics based on MoS(2) quantum dot fiber memristors was proposed for physical reservoir computing for the first time. Textile electronics convert raw electroencephalogram (EEG)and snoring audio directly into rich, high-dimensional state vectors based on intrinsic nonlinear dynamics. Leveraging 16 pulse-programmable conductance levels, the reservoir realizes an accuracy of 94.8%, 95.4%, and 93.5% in snoring events, sleep stages, and multimodal fusion, respectively. To enhance the robustness of feature extraction and improve classification performance under noisy conditions, the linear readout layer was replaced with a lightweight convolutional neural network. The hybrid neural network is 6 times faster than traditional deep-learning methods in 24-h segment EEG analysis. The memristors switch at ±1 V and sub-nanoampere currents, providing picowatt energy consumption suited to continuous on-body use. The results establish fiber memristor reservoir computing as an energy-efficient path to in-fabric, multimodal intelligence for next-generation home sleep analysis and wearable health care.