With the rapid development of the Internet of Things and human-computer interaction technologies, there is an urgent demand for the high-fidelity acquisition and accurate interpretation of human physiological information. However, traditional rigid sensors fail to address the mechanical mismatch with biological tissues. Although flexible electronic devices achieve conformal contact, their inherent signal drift, nonlinear response and hysteresis effects exceed the capability of traditional linear algorithms, which limits the application of wearable technologies in complex scenarios. To solve the limitations of hardware and software, innovations of flexible bio-MEMS (Bio-Micro-Electro-Mechanical Systems) in materials, device design, preparation technologies and integration, combined with artificial intelligence algorithms, provide effective solutions. This paper systematically reviews the latest research progress in this field: at the hardware level, it analyzes the innovative breakthroughs in materials such as transient electronics and nanocomposites, the structural innovations of bionic micro-nano structures (e.g., cilia arrays, synergistic cracks) in resolving the performance contradictions of sensors, and the engineering value of advanced processes such as micro-transfer printing and full 3D printing in realizing batch and high-precision manufacturing of devices. At the algorithm level, aiming at the pain points of physiological signals such as high noise and large individual differences, it discusses the applications of classical signal processing algorithms and neural network & deep learning algorithms (CNN, LSTM, etc.) in the processing of electromyography (EMG), electrocardiography (ECG) and electroencephalography (EEG) signals. It focuses on analyzing the effectiveness of Focal Loss optimization and attention mechanisms in dealing with class imbalance and cross-subject data, and elaborates on the transformation of artificial intelligence from a back-end "data analysis tool" to a front-end "booster" for sensor performance. Finally, this paper proposes the development direction toward "Hardware-Software Co-design" and "integration of sensing, storage and computing", and emphasizes the necessity of shifting algorithms from supervised classification to unsupervised "Anomaly Detection" under the background of sparse pathological data.