Abstract

The successful application of biosensor technologies in healthcare monitoring has created massive time-series data which are often distorted by noise, motion sources and environmental interference. Current machine learning and deep learning solutions tend to miss the ability to simultaneously feature localized signal properties, temporal dynamics on a long-range scale, and the ability to adapt to noise robustness. To solve these issues, the paper will suggest a multi-phase hybrid architecture to include Convolutional Neural Networks (CNNs), Transformer-based deep learning, and Reinforcement Learning (RL) in adaptive feature extraction and classification of noisy biosensor time-series data. The CNN module learns a discriminative local features, and the Transformer learns global temporal dependencies with self-attention mechanisms. A layer of RL-based optimization is presented that refines feature representations dynamically by adapting to changing noise levels by weighting and selecting, as well as by changing the noise levels. The suggested framework is tested on benchmark biosensor signals, both ECG and wearable sensor signals, in synthetic noise and real noise conditions. The experimental results show that the proposed model provides better performance: the accuracy of 95.7, F1-score of 95.2, and AUC of 0.97 are better than the traditional CNN, Transformer, and hybrid baselines. In addition, the RL component is highly robust to high noise. The proposed architecture offers a scalable and smart mechanism of real-time biosensor data analytics in the next-generation healthcare systems.