image: The self-powered intelligence concept brings together flexible electronics, on-device AI, and sustainable energy harvesting, aiming to enable long-term, stable, and intelligent personalized healthcare.
Credit: ©Science China Press
Imagine a future in which a flexible patch worn on the chest continuously monitors your heartbeat and performs the analysis directly on the device. When it detects an arrhythmia, it sends an alert before you feel anything wrong, all without ever needing a battery replacement: the heat you naturally dissipate and the small motions of daily life are enough to keep it running. Behind this vision are several technology streams that are now converging. A new review published in National Science Review by a team led by the Beijing Institute of Technology, together with collaborators from the Harbin Institute of Technology, the University of Glasgow, Imperial College London, and the National University of Singapore, systematically maps out this research landscape.
The review begins with the most fundamental challenge: power. Wearable and implantable medical devices currently rely on batteries that require periodic replacement or recharging, which is inconvenient for wearables and poses surgical risks for implants. The paper surveys five energy harvesting approaches that draw power from the human body and the surrounding environment: photovoltaic generation converts indoor and outdoor light and offers some of the highest power densities in wearable contexts; thermoelectric generation, which exploits the temperature gradient between skin and ambient air and benefits from the body is stable, round-the-clock thermal output; piezoelectric and triboelectric nanogenerators, which capture mechanical energy from high- and low-frequency body motion respectively; and radio-frequency energy harvesting, which collects ambient electromagnetic signals. The review notes that while a resting adult dissipates considerable metabolic power, body-worn harvesters typically capture only microwatts to milliwatts. Selecting a harvesting strategy is therefore not about maximizing a single efficiency metric, but about matching the harvester to the anatomical site, the expected activity pattern, and the power demands of the target application. The paper further discusses multi-source energy harvesting and AI-driven power management strategies that help maintain stable system operation under fluctuating energy availability.
With the power foundation established, the review turns to the physical platform: flexible electronics. Human skin stretches, sweats, and moves; internal organs undergo peristalsis and immune surveillance. Conventional rigid circuit boards are poorly suited to these dynamic environments. The review surveys key advances in this area: researchers are developing electrodes with thicknesses in the micrometer range and serpentine interconnects that maintain electrical integrity under repeated mechanical deformation; additive manufacturing approaches such as screen printing and laser sintering offer pathways toward scalable fabrication; biocompatible polymer substrates, including polydimethylsiloxane and polyimide, provide material platforms for long-term skin contact or implantation. The paper also covers hardware-level signal decoupling strategies that separate responses to different stimuli, such as temperature and pressure, at the physical sensor layer, supplying cleaner data for downstream AI processing.
The third pillar is on-device intelligence. Most current wearable devices operate by transmitting data to a smartphone or the cloud for analysis. This offloads computational complexity from the device, but the transmission itself consumes considerable energy, and the latency from acquisition to result can be a meaningful factor for time-sensitive health monitoring. The core concept articulated in the review, termed self-powered intelligence, proposes a different approach: AI inference runs locally on a low-power chip embedded in the device, so that data are processed as they are acquired rather than being shipped elsewhere. Running AI models under microwatt-to-milliwatt power budgets requires extreme algorithmic efficiency. The review surveys multiple research directions that address this challenge, including compressing neural networks to fit within microcontroller-class hardware, designing cascaded inference strategies that balance continuous monitoring accuracy against power consumption, and automatically searching for network architectures suited to specific hardware constraints. The contribution of the paper lies in bringing these technical threads together within a unified analytical framework tailored to wearable medical devices, and discussing their applicability and limitations in concrete clinical contexts. The review also examines non-Von Neumann computing architectures, such as memristor-based designs that fuse memory and computation in a single physical unit, thereby reducing the energy cost of shuttling data between separate processor and memory elements. This direction remains at an early stage of research, but it offers a notable pathway toward sustained on-chip AI inference under extreme energy constraints.
For the chest, flexible electrode patches can continuously acquire electrocardiogram and respiration signals, and researchers are exploring their potential for out-of-hospital chronic heart failure monitoring and early sleep apnea screening. For the wrist and limbs, textile-embedded sensors combined with lightweight classifiers may assist in distinguishing Parkinsonian tremor from benign age-related hand tremor. For implantable use, researchers are developing millimeter-scale sensors to monitor gastric motility and deep-tissue electrophysiological signals ……. The review emphasizes that most of these applications remain at the laboratory proof-of-concept stage, with substantial distance yet to cover before large-scale clinical deployment. It identifies three principal barriers to translation: manufacturing scalability, as the vast majority of flexible electronic devices are still fabricated by hand in academic labs; long-term reliability under the corrosive and mechanically demanding conditions inside the human body; and clinical validation and interpretability, since compressed AI models must retain sufficient accuracy and offer explainable outputs to earn trust from clinicians and clearance from regulators.
The review arrives at an opportune moment. Global wearable device shipments have surpassed 2 billion units annually, and the implantable medical device market is projected to exceed 150 billion dollars by 2030. The question of how to run intelligence on the body, continuously and autonomously, is moving from a niche research topic toward a central concern of the medical device industry. This review provides a systematic and grounded technical reference for addressing that question.