Wearable and implantable microfluidic technologies for future digital therapeutics

However, the accuracy and reproducibility of electrochemical sensors depend on the consistent delivery of the biofluid volume and flow to the sensor surface. Irregular biofluid input or fluctuating flow rates can be distorted, making it difficult to obtain precise concentration measurements and leading to significant errors. To address this issue, microfluidic systems are increasingly being integrated into electrochemical sensors, allowing for controlled collection, transport, and delivery of biofluids to the sensor interface.

Vinoth et al. proposed a fully screen-printed microfluidic wearable device that operates based on a pumpless capillary flow mechanism, allowing sweat to be passively guided into the sensing chambers without any external power.³³ The device is designed with four separate sensing chambers, enabling simultaneous analysis of multiple biomarkers such as lactate, Na⁺, K⁺, and pH using distinct electrochemical sensors. Fig. 2A illustrates the gradual influx of fluid into the microchannel over time, demonstrating the device's stable and reproducible fluid control architecture.

	Fig. 2 Representative microfluidic wearable devices. A) Sequential photographs demonstrating capillary-driven passive sweat intake in a fully screen-printed microfluidic patch. Reproduced with permission from ref. 33. Copyright 2021 American Chemical Society. B) Front and back views of the screen-printed device, highlighting distinct microfluidic pathways and electrode arrays. Reproduced with permission from ref. 37. Copyright 2024 American Chemical Society. C) Schematic of the wearable device showing flexible PET, conductive electrodes, sensing membranes, microfluidic channels, and skin adhesive. Reproduced with permission from ref. 12. Copyright 2021 Springer Nature. D) Exploded structural view of a wound-exudate monitoring device integrating battery, microfluidics, and electrochemical sensor arrays. Reproduced with permission from ref. 40. Copyright 2025 American Association for the Advancement of Science. E) Illustration of an exhaled breath condensate (EBC) monitoring mask featuring gradient micropillars, microchannels, and sensor reservoirs for passive fluid collection; scale bar: 4 mm. F) Real-time monitoring of various biomarkers from EBC under daily activity conditions including eating, resting, and exercising. Reproduced with permission from ref. 41. Copyright 2024 American Association for the Advancement of Science. G) An exploded view of a sweat-sensing microfluidic device with a detailed structure of the NFC electronics system. The inset image shows the structure and components of the NFC. H) Comparison of levels of selected analytes measured by a colorimetric sensing system and those measured by conventional lab analysis. Reproduced with permission from ref. 54. Copyright 2016 American Association for the Advancement of Science. I) Real-time measurements of sweat pH and sweat lactate levels in a rest-cycle-rest scenario. Reproduced with permission from ref. 55. Copyright 2024 American Association for the Advancement of Science. J) Photographs showing the structure of the device, before and after filling with a blue dye in water. K) A set of voltage decay curves in an in vitro experiment showing change of voltage resulting from different concentrations of phosphate buffer solution. The slow decay enables estimation of sweat arrival time. Reproduced with permission from ref. 56. Copyright 2019 John Wiley and Sons.

The internal architecture of the microfluidic channels plays a critical role in achieving quantitative control of fluid flow. In addition to capillary-driven designs, various studies have implemented microfluidic architectures such as spiral-shaped channels and capillary bursting valves (CBVs) to precisely regulate both the volume and flow rate of biofluids entering the sensors. CBVs are designed to hold back fluid until a threshold capillary pressure is exceeded, enabling on-demand flow release. For instance, spiral-shaped channels extend the fluid path, allowing sufficient residence time even with small volumes of biofluid, while their design inherently modulates flow resistance to support controlled fluid entry.^12,34 CBV structures, on the other hand, are engineered to permit fluid passage only after a specific pressure or volume threshold is reached, thereby enabling precise control over both the timing and quantity of fluid introduction.³⁵

Such precise regulation of fluid flow not only enhances the analytical accuracy on the sensor surface but also contributes to the signal stability, depending on the molecular recognition strategy employed. Various sensing mechanisms have been utilized for converting molecular interactions into electrochemical signals, including enzyme-based, aptamer-based, and molecularly imprinted polymer (MIP)-based approaches. Among these, enzyme-based electrochemical sensing is one of the most widely adopted strategies, wherein specific biochemical reactions with target molecules generate measurable electrochemical changes. For example, glucose oxidase (GOx)-based sensors have been extensively applied for monitoring glucose in sweat and have been effectively implemented across different microfluidic platforms, including three-dimensional paper-based microfluidic structures and position-resolved high-resolution sensing patches.^34,36 Similarly, other metabolic markers such as uric acid are also electrochemically detected using enzyme, MIP, or non-enzymatic strategies. Likewise, lactate oxidase (LOx) is another commonly used enzyme, particularly for monitoring lactate levels produced during exercise. It has been integrated into various wearable systems for real-time sweat analysis.³³ In another example, tyrosinase has been utilized in wearable devices for levodopa detection, where a spiral-shaped channel structure was employed to ensure stable fluid intake.¹² In addition, aptamer-based sensing strategies have emerged as highly specific molecular recognition tools that can replace traditional antibodies. These sensors use synthetic DNA or RNA sequences that bind selectively to the target analytes. Recent studies have successfully applied aptamer sensors for detecting trace levels of estradiol in sweat. The device used in this case was designed with surface hydrophilicity tuning and automated sampling structures, enabling fast and stable fluid acquisition and analysis.³⁵

Meanwhile, molecularly imprinted polymer (MIP)-based sensors utilize synthetic receptors tailored to the molecular structure of target analytes, enabling electrochemical detection without relying on antibodies or enzymes. Garg et al. developed a wearable electrochemical sensing platform for real-time cortisol detection using MIP technology.³⁷ The device integrates paper-based microfluidic channels with MIP electrodes (Fig. 2B). Sweat passively flows through the microfluidic channels and reaches the electrode surface, where the MIP receptors selectively bind to cortisol molecules. This binding induces a change in electrochemical impedance spectroscopy, which can be quantitatively measured in real-time. The system provides both high molecular recognition performance and improved operational lifespan and environmental stability compared to enzyme-based sensors, indicating promising potential for broader biomarker detection. This approach offers several advantages, such as durability, cost-effectiveness, and high stability under variations in temperature and pH. Additionally, the potential for repeated use makes it particularly suitable for wearable electrochemical sensing applications.

Another important consideration in wearable electrochemical sensors is the ability to reliably acquire sweat for measurement. Under resting conditions, natural sweat secretion occurs at a low rate and with irregular volume, which can limit the stability of sensor response. To overcome this challenge, sweat induction techniques using iontophoresis have been developed. This method involves applying a mild electrical current to the skin to enhance the transdermal absorption of compounds such as pilocarpine or carbachol, thereby promoting localized sweat production.³⁸ The ability to generate sweat on demand makes this approach well-suited for wearable devices requiring continuous and quantitative fluid analysis. Several studies have demonstrated the integration of iontophoresis with microfluidic channels and electrochemical sensors, achieving automated and quantifiable workflows from sweat induction to sensing.^34,35,39

Meanwhile, systems designed for passive operation under resting conditions, without external stimulation, are also being actively explored. These systems enable fluid analysis under more natural physiological conditions, minimizing concerns about skin irritation, electrode interference, or distortion of sweat composition. Nyein et al. developed a wearable patch incorporating a wheel-spoke electrode configuration and a spiral-shaped microfluidic channel (Fig. 2C).¹² The wheel-spoke structure not only supports analyte detection but also enables real-time calculation of sweat rate. Additionally, its design improves operational reliability under low-sweat flow conditions. The device employs a passive flow mechanism and hydrophilic filler to ensure reliable sweat collection even at low secretion rates. With radially arranged electrodes, the system allows simultaneous monitoring of pH, Cl⁻, levodopa, and sweat rate under resting conditions. Notably, the patch is designed for continuous wear over 24 hours, demonstrating its practical potential for long-term physiological monitoring in daily life.

Wearable electrochemical sensor research has traditionally focused on sweat as the primary biofluid for biomarker analysis. However, recent advances have expanded the range of target fluids to include other biological samples such as wound exudate and exhaled breath condensate (EBC). This expansion broadens the scope of wearable sensors beyond basic physiological monitoring toward clinically meaningful applications such as early disease detection, treatment response assessment, and prognosis prediction.

In this context, Wang et al. developed a wearable microfluidic electrochemical sensor system, termed iCares, designed for the analysis of wound exudate (Fig. 2D).⁴⁰ The device integrates three key microfluidic elements to enable efficient collection and transport of wound fluid: a Janus membrane that directs fluid intake based on surface tension, a wedge-shaped channel that supports unidirectional flow and filters impurities, and a 3D micropillar array that stabilizes fluid movement. These components operate together without the need for external pumps, ensuring stable delivery of wound exudate to the sensor array. The integrated electrochemical sensor array is capable of detecting wound-related biomarkers such as NO, H₂O₂, O₂, and pH in real-time, enabling precise monitoring of inflammation, infection progression, and antibiotic response. The device was evaluated in both a diabetic wound mouse model and actual chronic wound patients, demonstrating potential for integration with machine learning-based wound classification algorithms. Notably, in the diabetic mouse model, the sensor was able to detect early signs of infection before they became visually apparent, and machine learning analysis using data from 20 patients achieved 94% accuracy in predicting healing time and wound severity classification.

In addition to wound exudate-based analysis, wearable electrochemical sensor technologies targeting EBC are gaining increasing attention. Heng et al. proposed a mask-type wearable system that utilizes a bioinspired microfluidic architecture to collect condensed water vapor generated during respiration and analyze various metabolic molecules contained within (Fig. 2E).⁴¹ This device integrates a graded micropillar array, hydrophilic inner surface, and efflux channel to create a 3D microfluidic pathway that enables passive operation without external power—covering the entire process from collection to sensing and removal of condensate. The integrated sensor array is capable of real-time detection of NO₂⁻, NH₄⁺, alcohol, pH, and temperature, making it a promising non-invasive platform for respiratory biomarker monitoring. Fig. 2F presents in situ EBC analysis data collected from healthy participants wearing the device under various daily activity conditions, demonstrating its practical usability in terms of wearability, detection sensitivity, and resistance to motion artifacts. This holds significant potential for early diagnosis and continuous metabolic monitoring of respiratory diseases such as asthma and COPD. The EBCare device supports over 4 hours of continuous indoor and 8 hours of outdoor monitoring. Thanks to its tandem cooling architecture, it achieves up to five times higher condensate collection efficiency compared to conventional masks. Notably, stable condensate collection rates of over 4 μL min⁻¹ were maintained even under 1 sun solar exposure.

These advancements demonstrate that microfluidic technology is extending the reach of wearable sensors beyond sweat-based applications toward diverse fluidic diagnostics, positioning itself as a critical tool in the evolution of precision medicine and personalized health monitoring. The high-dimensional datasets acquired from these sensors move beyond single-analyte diagnostics and enable pattern-based disease classification. Particularly, the integration of such datasets with machine learning algorithms opens opportunities for disease prediction, treatment response assessment, and the development of tailored diagnostic solutions, paving the way for more sophisticated digital healthcare systems.^42,43

Overall, electrochemical microfluidic sensors offer high sensitivity and real-time monitoring capability but depend heavily on consistent biofluid acquisition and long-term electrode stability. While passive microfluidic architectures mitigate flow variability, they remain susceptible to environmental and user-dependent factors. Future work must focus on enhancing robustness under variable physiological conditions and improving long-term calibration-free operation.

Colorimetric sensors are highly practical not only for everyday use, but also for specialized applications such as analyte profiling for athletes as they can offer enhanced flexibility and deformability due to the inexistence of electrical components.^44–47 Notably, their passive, battery-free operation eliminates the need for charging and potential safety concerns often associated with battery-powered devices, while their non-invasive and easily applicable nature liberates the users from pain for analyzing specific substances that have previously required blood draws.^48,49 These advantages may allow people to manage health without visits to clinics or uses of large and expensive equipments.

Additionally, the use of external body fluids such as sweat is gaining increasing validity for biochemical monitoring as correlations between sweat and blood concentrations have been established for a variety of analytes.^50–52

The usage of colorimetric microfluidic devices can extend from simple sampling to integrated sampling and analysis. Choi et al. developed a microfluidic sampling device that incorporates capillary bursting valves to collect sweat with time information.³⁰ This device incorporates microchannels and microreservoirs with capillary bursting valves, enabling regulation of the flow of sweat. The CoCl₂ in the device reacts with sweat to make colour changes, which can visually inform the users of the presence of sweat in the device.

More quantitative studies advance such devices to analyse the colour for quantitative measurements can be found in numerous other studies. For example, Xiao et al. developed a microfluidic colorimetric sweat sensing device capable of measuring glucose levels from sweat.⁵³ The device measures glucose with glucose oxidase (GOx), HRP, and o-dianisidine. Notably, o-dianisidine was selected instead of the widely used KI solution because of its superior sensitivity compared to KI. Mechanically, the device incorporates check valves in the microfluidic chip to prevent backflow of the chemical reagents from the microchamber to the skin. Resultant colours can be imaged and analysed with a smartphone.

Microfluidic colorimetric devices can also support multiplexed analyte detection and quantification. In one study, a multi-sensing device was fabricated to sample sweat and measure pH, Cl⁻, glucose, lactate, and creatinine levels.⁵⁴ An exploded view of the device is shown in Fig. 2G. The device features distinct mechanical properties in that it has openings with a 3 mm diameter, allowing selective sweat extractions from up to 10 specific sweat glands. In the device, each analyte reacts with distinct solutions to make colour changes. The amount of sweat is measured by CoCl₂, which is deep blue in its natural state but changes to pale purple after reactions with sweat. Glucose reacts with GOx to produce H₂O₂, which reacts with HRP and KI to turn from yellow to brown. Lactate levels are measured by firstly reacting with a mixture of lactate dehydrogenase and nicotinamide adenine dinucleotide (NAD⁺) solutions and further interacting with the formazan dyes to shift from brown to yellow. Cl⁻ measurement facilitates Hg²⁺, Fe²⁺, and 2,4,6-tris(2-pyridyl)-s-triazine (TPTZ) to change from transparent to blue. pH levels are detected using the universal pH indicator compromising bromothymol blue, methyl red, and phenolphthalein. Creatinine is measured by a mixture of creatinase, creatininase, peroxidase, and 4-aminophenazone. Aside from analysing concentration measurements, temperature is also measured in the device and can be transferred to a smartphone wirelessly via near-field communication (NFC).

The resultant colours can be imaged by a smartphone and analysed with an image analysis software, which calculates the concentration of each analyte from the colours in the images. In the study, the analytes were also measured in conventional laboratory analysis, and the two results showed excellent agreement, as shown in Fig. 2H.

Because analyte contents in sweat can fluctuate with temporary physiological conditions of the body, some studies have focused on time-resolved profiling of sweat.⁵⁵ Cho et al. developed a band-type microfluidic device incorporating a sweat-activated colorimetric timer that can measure pH and lactate from sweat. The timer, based on a maltodextrin–iodine complex, is purple in its normal state but becomes lighter upon sweat exposure. pH is measured using bromothymol blue (BTB)-conjugated mesoporous nanoparticles. A human study was conducted on an unfit group and a fit group wearing the device to measure the lactate and pH profiles of sweat during exercise. The results of the unfit group are shown in Fig. 2I. The unfit group showed a large elevation in lactate levels and a decrease in pH levels, while the fit group showed small changes in lactate and pH levels while exercising.

Another microfluidic device that samples and analyzes sweat over time is introduced in a study by Bandodkar et al.⁵⁶ Primarily focused on dynamic sweat composition measurements, the introduced device utilizes flexible galvanic cells and passive valves that work together to serve the functions of a sweat-activated stopwatch. A brief structure of the device and its photographs before and during the blue dye filling in water are shown in Fig. 2J. The voltage decay of the galvanic cells can be tuned by external resistance. A large resistance slows the voltage decay, enabling accurate estimation of the sweat arrival to the device (Fig. 2K). From the sweat, pH and Cl⁻ can be measured, each with different mechanisms. pH reacts with the universal pH indicator dye, and Cl⁻ reacts with a selective colorimetric reagent present in the satellite zone of the device. In both cases, the resultant colours are imaged and analysed using a smartphone.

On-demand measurement can also be a useful function for a microfluidic device.⁵⁷ One study conducted by Mishra et al. introduces a soft microfluidic patch featuring a thin tab that activates a pump connected to the valve when pulled, which allows the user to activate the device at will. In the device, chloride, calcium, glucose, pH, urea, creatinine, and copper levels can be measured from sweat. Commercial colorimetric assays were used for measurements of all 7 analytes, which can be imaged by an external device and analysed using ImageJ software based on the colours of the images.

Though popular for its accessibility to wearable devices, sweat is not the only specimen used for microfluidic sampling. Tear fluid is also a promising candidate for microfluidic sampling. Wang et al. developed a microfluidic device for collecting tear and monitoring its biomarkers.⁵⁸ The device incorporates several pieces of filter paper loaded with colour dyes to measure vitamin C, pH, Ca²⁺, and proteins. Upon entering the device, the tear fluid visits four different chambers, each for different analyte measurements. In the chamber for vitamin C, the sweat undergoes a redox reaction with 2,6-dichloroindophenol to turn from blue to transparent. In the chamber for pH measurement, sweat reacts with the universal pH indicator to change between red and blue based on its acidity. In the chamber for Ca²⁺ measurement, the sweat reacts with a mixture of o-cresolphthalein complexone (o-CPC) and 8-hydroxyquinoline with AMP buffer to change the contrast of the purple complex. In the chamber for protein measurements, TBPB was used to facilitate the reaction with sweat to change the colour from yellow to green. The colours can be imaged with a smartphone and analysed using an artificial intelligence (AI)-assisted application. Major microfluidics for wearable devices are summarized in Table 1.

Overall, the use of readily accessible analytes such as those found in sweat, combined with the safety and comfort offered by colorimetric devices, highlights their potential as tools for healthcare support. Thanks to these aspects of colorimetric sensors, they can be employed in more extreme cases such as the point-of-care tests in pandemics like COVID-19.⁵⁹ Nevertheless, these devices largely rely on smartphone-based image analysis, which is susceptible to variations in ambient lighting conditions and camera angles, which can result in changes in color resolution, thereby limiting the accuracy of quantitative measurements.⁶⁰

Traditional drug delivery devices for neural applications are composed of a single rigid probe incorporating a microfluidic channel alongside recording and stimulating electrodes.^73,74 These conventional probes have evolved over the past few decades to become multimodal platforms capable of simultaneously analysing functional connectivity between different brain regions. Shin et al. have developed a multifunctional multi-shank microelectromechanical system (MEMS) neural probe, that enables both chemical modulation and neural signal recording across spatially distinct brain areas.⁷⁵ The device was specifically designed to target the CA3 and CA1 regions of the hippocampus, allowing simultaneous monitoring and modulation of long-range neural circuits. It consists of four silicon probes (Fig. 3A). One main multimodal shank incorporates all major functionalities with microchannels for drug delivery, an SU-8 optical waveguide for optical stimulation, and microelectrode arrays for neural recording. The remaining three shanks were each spaced 200 μm apart. The probes were specifically designed to be 7 mm long, to record the neural responses in both the CA3 and CA1 regions. Each shank is 128 μm wide and 40 μm thick and is integrated with eight 400 μm² iridium microelectrodes for high-resolution signal acquisition. To enable real-time chemical modulation, a microfluidic staggered herringbone mixer (μSHM) made of PDMS was integrated into the device. This mixer facilitates effective blending of two drug solutions in low Reynolds number conditions before delivery through 10 μm wide and 12 μm high microchannels. The optical waveguide, 40 μm wide and 15 μm thick, delivers blue light from an external source to the stimulation site with sufficient intensity to activate light-sensitive neurons. The authors investigated the functional connectivity of hippocampal CA3 and CA1 regions in Thy1-ChR2-YFP transgenic mice with the multifunctional multi-shank MEMS neural probe. Optical stimulation of the CA3 region was first performed to validate connectivity, as indicated by synchronized neuronal responses in the CA1 region. Following this, two synaptic receptor antagonists, cyanquixaline (CNQX) and (2R)-amino5-phosphonovaleric acid (AP5), were delivered through the microfluidic channels to block synaptic transmission. A significant reduction in CA1 activity confirmed the successful modulation of the circuit. Partial recovery of the CA3–CA1 response was observed 35 minutes after the injection, demonstrating the reversibility of the chemical intervention.

	Fig. 3 Implantable drug delivery microfluidic systems. A) A schematic of the multifunctional multi-shank MEMS neural probe. Reproduced with permission from ref. 75. Copyright 2019 Springer Nature. B) Illustration of the multifunctional hydrogel hybrid probes for neural applications. Reproduced with permission from ref. 78. Copyright 2021 Springer Nature. C) Exploded view of a wireless, battery-free, fully implantable optofluidic cuff system. Reproduced with permission from ref. 79. Copyright 2019 American Association for the Advancement of Science. D) Image of a long-acting subcutaneous ISL delivery nanofluidic implant. Reproduced with permission from ref. 80. Copyright 2023 American Association for the Advancement of Science. E) Illustration of a polymeric platform for the controlled dual-drug transscleral delivery of neuroprotective agents. Reproduced with permission from ref. 82. Copyright 2014 John Wiley and Sons. F) Schematic of an implantable microdevice for in vivo drug sensitivity testing within solid tumours. Reproduced with permission from ref. 83. Copyright 2015 American Association for the Advancement of Science. G) Real device image of IOP sensor. Reproduced with permission from ref. 93. Copyright 2014 Springer Nature. H) Reagent mixing part of glucose, lactate sensor based colorimetric assay. Reproduced with permission from ref. 84. Copyright 2019 Springer Nature. I) Electroosmotic flow for drug delivery and analysis of ISF. Reproduced with permission from ref. 95. Copyright 2021 Springer Nature. J) Protein sensor using active reset. Reproduced with permission from ref. 103. Copyright 2024 American Association for the Advancement of Science. K) Naloxone delivery via motor-driven actuator. Reproduced with permission from ref. 113. Copyright 2024 American Association for the Advancement of Science. L) Schematic illustrations of the closed-loop drug delivery based on electrophysiological signals. Reproduced with permission from ref. 114. Copyright 2023 Springer Nature. M) Integrating and folding process with multilayer component of chemical, optical, and electrical modalities. Reproduced with permission from ref. 117. Copyright 2025 Springer Nature.

Recent advances in implantable bioelectronics have focused on minimizing the foreign body response associated with long-term neural interfaces. A key strategy involves reducing the mechanical mismatch between implanted devices and brain tissue by employing material with high biocompatibility and elastic moduli closer to that of neural tissue.^76,77 Park et al. developed a hybrid multifunctional fibre-based probe, which integrates microscale polymer fibres into a soft poly(acrylamide)-alginate (PAAm-Alg) hydrogel matrix.⁷⁸ This hydrogel closely mimics brain tissue in softness (∼16.5 kPa) and mechanical resilience, providing a ‘stealthy’ interface that reduces tissue stress during brain micromotion. The device, thermally drawn from macroscale preforms, combines microfluidic drug delivery, optical stimulation, and long-term electrophysiological recording using tin microelectrodes (Fig. 3B). The microfluidic channels, made of poly(etherimide) (PEI), have inner and outer diameters of 54 μm and 115.4 μm, respectively. Each microelectrode array consists of seven electrodes, each 4.75 μm in diameter. The waveguide, built from a polycarbonate (PC) core with cyclic olefin copolymer (COC) cladding, measures 105.9 μm in diameter. These functional components are embedded in a hydrogel layer resulting in a total device diameter of approximately 334 μm. Finite element analysis (FEA) and in vitro experiments demonstrated that the bending stiffness of the device in the swollen state is significantly lower than that of traditional polymer or metal-based probes, reducing the mechanical strain on surrounding brain tissue. Importantly, the device is inserted in a dehydrated state, which temporarily increases stiffness for successful brain penetration without the need for additional coatings or shuttles. After implantation, the hydrogel rapidly rehydrates in vivo and reverts to its soft, compliant form. Functional validation was performed in transgenic mice. An adeno-associated virus expressing ChR2-eYFP was injected into the basolateral amygdala (BLA) and the hydrogel device was implanted into the ventral hippocampus (vHPC), where BLA axons terminate. The mice were put to a standard open field test. Upon optical stimulation of the vHPC, the mice displayed increased anxiety-like behaviour, evidenced by reduced time spent in the centre of the open field. To test if the glutamatergic inputs alter the anxiety-like behaviours in mice, a glutamate receptor antagonist cocktail with AP5 was delivered through the microchannels of the device. The cocktail suppressed the optically induced increase of anxiety, resulting in no significant difference in time spent in the centre of the open field compared to the control group.

Recent advances in implantable microfluidic systems have extended their application from the central to the peripheral nervous system (PNS). Zhang et al. present a wireless, battery-free implantable device for localized optical and chemical modulation of peripheral nerves.⁷⁹ This fully implantable platform consists of a soft elastomeric cuff (modulus ∼3 MPa), designed to match the mechanical properties of peripheral nerves (∼7 MPa for mouse sciatic nerve), thereby reducing mechanical strain and minimizing immune response (Fig. 3C). The cuff integrates four microfluidic channels (each 60 × 60 μm in cross-section) for drug delivery and a microscale inorganic LED (μ-ILED) (270 × 220 × 50 μm) for localized optical stimulation. A compact base station (∼5 mm radius, 4 mm thick), implanted subcutaneously at the thoracic spine, contains fluid reservoirs, electrochemical micropumps, and wireless power-harvesting electronics. The micropumps, actuated by water electrolysis in a KOH solution, deform a flexible SBS membrane to push drugs from the reservoir to the nerve surface, achieving flow rates of ∼1.5 μL min⁻¹ without generating significant heat. Each reservoir can be refilled through designated ports, enabling multiple cycles of drug delivery. The device was implanted in mice, with the cuff wrapped around the sciatic nerve and the base station anchored with sutures. Behavioural and histological comparisons between implanted mice, sham-operated controls, and animals implanted with rigid polyethylene (PE) cuffs revealed no nerve damage or functional deficits with the soft cuff system, while the PE cuff induced gait disturbances and immune cell infiltration due to its high stiffness (∼0.25 GPa). Transgenic TRPV1-ChR2 mice were used to evaluate the performance of the multimodal device. Upon blue light activation, these mice showed significant aversion behaviours, confirming functional engagement of nociceptive pathways. In a second experiment, the system delivered bupivacaine, a local anaesthetic, via the microfluidic channels. Thermal paw withdrawal tests showed a significant increase in latency after drug delivery, indicating effective neural blockage and decreased sensitivity to heat.

The application of implantable microfluidic devices has expanded beyond neural interfaces to target specific tissues for localized therapy. Pons-Faudoa et al. developed a long-acting, refillable nanofluidic implant (nISL) for sustained subcutaneous delivery of islatravir (ISL), a potent nucleoside reverse transcriptase translocation inhibitor under investigation for HIV pre-exposure prophylaxis (PrEP).⁸⁰ The implant, constructed from medical-grade titanium, contains a 278600 nanochannel silicon carbide membrane sealed inside a 570 μL drug reservoir (Fig. 3D). Two self-sealing silicone refill ports enable transcutaneous drug loading, eliminating the need for surgical retrieval or replacement. The nanochannels, protected from direct tissue contact by an outer microchannel interface, regulate ISL release through passive diffusion, determined by channel size, number, and drug solubilization kinetics. In vivo testing in rhesus macaques demonstrated stable plasma ISL levels over a 20-month period following a single transcutaneous loading event, with no refilling required. To assess preventive efficacy, male and female macaques were implanted with nISL and subjected to repeated low-dose rectal or vaginal SHIV_SF162P3 challenges, respectively. All male macaques (6/6) implanted with nISL remained uninfected after 10 weekly rectal challenges, while all control animals became infected. Similarly, in the vaginal challenge cohort, 5 of 6 treated female macaques remained uninfected, with the only infection occurring in an animal that had lost the implant. No systemic toxicity was observed during the study, with liver and kidney function remaining within normal ranges and no significant changes in total lymphocyte counts. Local tissue analysis revealed a moderate foreign body response, characterized by a fibrotic capsule surrounding the implant. However, drug release and tissue diffusion were not significantly impaired by the capsule thickness, collagen density, or vascularity.

Ocular applications also benefit from precise microfluidic control.⁸¹ Nagai et al. developed a polymeric platform for the controlled dual-drug transscleral delivery of neuroprotective agents to the retina in a rat model of light-induced retinal degeneration.⁸² The device consists of a microfabricated reservoir, and a controlled-release cover composed of varying ratios of poly(ethylene glycol) dimethacrylate (PEGDM) and tri(ethylene glycol) dimethacrylate (TEGDM) (Fig. 3E). By modulating the PEGDM/TEGDM ratio in the cover and drug matrix, the release kinetics of each drug could be independently tuned. PEGDM, with longer polymer chains, allows for increased swelling and enhanced drug permeability, whereas TEGDM, with shorter chains, forms a denser, less permeable network. Two therapeutic agents were selected: edaravone (EDV), a free radical scavenger, and unoprostone isopropyl (UNO), a large-conductance Ca²⁺-activated potassium channel activator. In vivo, the device was implanted onto the rat sclera, delivering the drugs directly through the outer eye tissues without the need for invasive intravitreal injection. Rats were exposed to intense light to induce photoreceptor degeneration, mimicking key features of retinal diseases. Electroretinography (ERG) revealed that co-delivery of EDV and UNO led to significantly greater preservation of retinal function than single-drug devices or phosphate-buffered saline (PBS) controls. Histological analysis confirmed that the outer nuclear layer (ONL) thickness was best preserved in the EDV/UNO group. TUNEL staining also showed reduced apoptotic cell death in this group.

Implantable microfluidic devices have also shown great promise in enabling localized delivery of therapeutics within tumours. Jonas et al. developed an implantable microdevice capable of performing high-throughput in vivo drug sensitivity testing directly within solid tumours.⁸³ The device is cylindrical in shape, measuring 820 μm in diameter and 4 mm in length, and is inserted into the tumour using a standard biopsy needle (Fig. 3F). It contains up to 16 spatially separated reservoirs, each measuring 150–350 μm in diameter and 150–250 μm in depth, spaced 400–750 μm apart to prevent drug diffusion overlap between compartments. Each reservoir is loaded with a single agent or drug combination, enabling parallel assessment of localized drug responses. Upon implantation, the drugs are passively released into spatially distinct regions of the tumour, allowing microdose scale interactions with the native tumour microenvironment. This platform was tested in vivo using mouse models xenografted with various human tumour types. The results demonstrated that the device could simultaneously evaluate the efficacy of multiple therapeutics. Importantly, the localized responses observed via histological and pharmacodynamic markers were shown to correlate with systemic treatment outcomes. This approach enables high-throughput, minimally invasive drug screening within living tumours, offering a powerful tool for predicting patient-specific drug sensitivity and optimizing treatment strategies prior to systemic administration. It holds strong potential for advancing personalized oncology and accelerating early-phase drug development.

Implantable drug delivery platforms enable precise, localized therapy but face challenges in long-term biocompatibility and refillability. While multifunctional integration with sensing modules advances closed-loop control, mechanical mismatch and chronic immune responses remain key barriers. Future research must optimize material softness, minimize surgical burden, and enhance autonomous control algorithms.

The pathophysiology of glaucoma is multifactorial, involving various contributing factors acting in a complex manner, but intraocular pressure (IOP) remains a major factor.^89–92 However, IOP can fluctuate significantly throughout the day due to factors related to a patient's daily activities, making continuous 24-hour monitoring more critical. In response to this need, Araci et al. developed a fully implantable device to continuously monitor IOP based on a microfluidic mechanism that leverages intraocular aqueous humor as the driving medium (Fig. 3G).⁹³ While one end of the sensor is exposed to the intraocular aqueous humor, the other end comprises a sealed microfluidic channel connected to a gas reservoir. As aqueous humor enters the microchannel via capillary action, the gas reservoir is compressed, causing an increase in internal pressure. The gas–liquid interface consequently shifts toward the gas reservoir until the internal pressure equilibrates with the IOP. This configuration allows real-time visualization of pressure changes through the displacement of the gas–liquid interface. The lens component of the sensor is fabricated using PDMS with a diameter of 8 mm and is coated with parylene-C to prevent air leakage.

Various biomarkers present in ISF can be quantitatively monitored using microneedle-based microfluidic sensors. ISF can be accessed in a minimally invasive and almost painless manner, as microneedles penetrate only the shallow subcutaneous layer by a few millimeters. Continuous glucose monitoring (CGM) device, which monitor glucose concentrations in the ISF using microneedle-based technology, are commercially available and widely adopted. To date, only glucose monitoring devices have reached commercialization, and existing CGMs still encounter challenges such as biofouling and signal drift.⁹⁴ In this regard, Nightingale et al. developed a sensor utilizing an optical flow cell-based colorimetric assay to detect alternative biomarkers such as lactate (Fig. 3H).⁸⁴ First, a 200 μm-diameter microdialysis probe with one end sealed by a semipermeable membrane (molecular weight cut-off of 13 kDa) is inserted subcutaneously. When sterile PBS is circulated through the tubing connected to the probe, components from the ISF diffuse into the PBS across the membrane. The resulting ISF–PBS mixture is drawn into a sensing unit that integrates a microfluidic chip, an optical flow cell, electronics, and a fluid reservoir cartridge, using a peristaltic pump. This setup constitutes a microfluidic analysis system with optimized temporal resolution enabled by a droplet-flow regime. Within the microfluidic chip, the ISF–PBS mixture is first combined with a reagent that reacts with the target molecule, and subsequently with a chromogenic reagent. The resulting mixture is then segmented into a stream of droplets by immiscible oil. The optical flow cell detects the color intensity of each droplet, which correlates with the concentration of the target analyte. This analytical approach is referred to as a colorimetric assay. The measured data are processed by an integrated microcontroller and transmitted wirelessly via Bluetooth. However, when this device was used for in vivo detection of glucose and lactate, the sensor successfully measured lactate concentrations up to 20 mM. In contrast, glucose detection was limited to 8 mM despite high sensitivity, which does not cover the full physiological range of glucose levels observed in diabetic patients.⁸⁴

Capillary action or diffusion can be used to extract ISF, but electroosmotic flow (EOF) offers more efficient and controllable fluid transport, which is crucial for reliable chemical analysis. Kusama et al. developed polymer-based ion-conductive porous microneedles (PMNs) not only for transdermal drug delivery but also for the extraction of ISF molecules and glucose sensing (Fig. 3I).⁹⁵ The 250 μm-long microneedles were able to penetrate the highly resistive stratum corneum, thereby reducing transdermal resistance and enabling effective molecular transport. Additionally, the formation of anode-to-cathode EOF was induced by sulfonic acid group-containing hydrogels, improving the iontophoretic extraction performance. In particular, PMNs filled with poly(2-acrylamido-2-methylpropane sulfonate) (PAMPS) significantly enhanced the glucose extraction rate.

Most previously reported sensors have primarily focused on the detection of small molecules, such as neurotransmitters, electrolytes, and drugs.^96–99 Small molecule sensing can often be performed electrochemically without the need for reagents.¹⁰⁰ Moreover, even when reagents are used, the high dissociation rates of small molecules make them well-suited for real-time detection.¹⁰¹ In contrast, protein detection inevitably requires the use of reagents, and current technologies remain insufficient for real-time monitoring of protein concentration changes due to the inherently slow dissociation kinetics of protein–receptor interactions.¹⁰² Therefore, Zargartalebi et al. developed a reagentless, continuous sensing platform by immobilizing a double-stranded DNA structure—comprising a ferrocene moiety and an aptamer—on the electrode surface (Fig. 3J).¹⁰³ ISF is drawn into the device through a microchannel composed of a hydrophilic material that generates capillary force. When the aptamer binds to a target protein within the ISF, the electron transfer rate between the ferrocene and the electrode is reduced due to hydrodynamic drag. This decrease in current enables reagent-free quantification of protein concentration. Furthermore, to overcome the limitations posed by slow dissociation kinetics, high-frequency voltage switching was employed to rapidly alter the sensor's potential, thereby utilizing drag and inertial forces to promote the dissociation of the aptamer–protein complex. These efforts can be further advanced by employing chemically guided structural design approaches to improve molecular specificity and enhance sensing precision.¹⁰⁴

Implantable microfluidic sensors offer unparalleled accuracy by accessing internal biofluids directly but at the cost of increased invasiveness and potential foreign body response. Microneedle-based approaches mitigate these issues, though long-term stability and biofouling still limit performance. Progress in anti-fouling coatings and minimally invasive insertion strategies will determine their clinical viability.

Opioids have been used for centuries to relieve acute pain, and as of 2017, more than 40 million people worldwide are reported to be dependent on opioids. However, in cases of opioid overdose, respiratory depression is induced, which can lead to cessation of breathing and result in brain damage or death within seconds to minutes. Ciatti et al. proposed a closed-loop system to enable real-time naloxone delivery by continuously monitoring physiological parameters such as oxygen saturation (SpO₂), respiration rate, and heart rate (Fig. 3K).¹¹³ The system incorporates a subcutaneous implant embedded with an optical sensor for continuous SpO₂ monitoring. Upon detection of a rapid decrease in oxygen saturation, naloxone is administered either intravenously or subcutaneously. In the intravenous configuration, the drug is delivered into the bloodstream by generating gas through electrolysis, which increases pressure inside the drug reservoir and drives injection. In the injection-type configuration, a miniature DC motor is used to actuate a naloxone-loaded needle. The system also allows real-time monitoring of the physiological response following drug administration.

Neuromodulation based on the brain's electrophysiological signals has shown efficacy in neurorehabilitation and the treatment of neurological disorders. In particular, Ouyang et al. developed a closed-loop neuromodulation implant that enables epilepsy management by detecting seizures using electroencephalography (EEG) signals and initiating drug release in response to electrophysiological activity (Fig. 3L).¹¹⁴ Biosignals such as EEG, electromyography (EMG), and body temperature can be recorded and used to guide appropriate implementations of optogenetic and pharmacological stimulation, controlled through an AI-driven system. In this system, a PtB-coated electrode interfaces with the dura mater of the skull via a small columnar structure composed of PDMS, allowing EEG signal acquisition. EEG, EMG, and body temperature data are wirelessly collected through a Bluetooth low energy (BLE) system-on-chip, which interfaces with an AI algorithm to determine the optimal timing for drug delivery. Drug release is actuated by a bipolar junction transistor (BJT). When a voltage is applied to a cross-type electrode structure, electrolysis in the electrolyte solution generates gas, which accumulates within the electrolyte chamber. The resulting increase in internal pressure deforms a flexible membrane, exerting force on the drug reservoir and thereby pushing the drug through a microfluidic channel. A deep learning-based AI system is incorporated to enable more precise drug control.

For a system to operate in a closed-loop manner, the sensing module and the drug delivery mechanism must be integrated into a single device platform. Since performance degradation can occur when multiple functional components are integrated into a single layer, these devices are typically designed with a multi-layer structure. Via-hole is commonly employed to interconnect the multi-layer structures.¹¹⁵ However, this approach is often unsuitable for thin and flexible materials due to mechanical limitations.¹¹⁶ To overcome this limitation, Lee et al. introduced a folded structural design that enables the fabrication of a skin-conformal device using thin and flexible materials (Fig. 3M).¹¹⁷ This approach involves fabricating the components in a single integrated layer, which is then folded to form the complete three-dimensional structure. The sensing and drug delivery modules are structurally decoupled, allowing them to function independently without mutual interference. The device measures blood flow changes via photoplethysmography (PPG) using optical sensing, and also records electrophysiological signals such as electrocardiogram (ECG) and EMG. Drug delivery is achieved through a micropump-based chemical delivery system. Additionally, mechanical flexibility is enhanced through a vertical cutting pattern, while durability is improved using serpentine interconnects. Furthermore, the insertion of a strain-isolating layer at the skin–contact interface allows the device to maintain stable functionality even under significant mechanical strain. Major microfluidics for implantable devices are summarized in Table 2.

Conventional drug delivery systems lack the ability to adapt to dynamic physiological changes, often resulting in suboptimal therapeutic outcomes. Closed-loop drug delivery systems offer a promising solution by integrating real-time biosensing, intelligent decision-making, and on-demand drug release into a single platform. By continuously monitoring physiological signals such as SpO₂, EEG, EMG, and body temperature, these systems can personalize drug administration and minimize risks associated with overdose or underdose. Recent advances, including implantable devices for opioid overdose intervention and AI-guided neuromodulation systems for epilepsy management, demonstrate the clinical potential of such platforms. Furthermore, innovative structural designs, such as folded architectures and strain-isolating interfaces, enable the seamless integration of sensing and delivery modules in thin, flexible, and durable formats suitable for long-term bio-interfacing. Together, these developments mark a significant step toward precision medicine through autonomous, adaptive therapeutic systems.

Closed-loop systems enable autonomous and adaptive therapy by linking real-time sensing with on-demand drug delivery. However, achieving seamless integration of sensors, actuators, and control algorithms in miniaturized, biocompatible platforms remains a key challenge. Future efforts should emphasize chronic stability, precise dosing algorithms, and regulatory validation to translate these intelligent systems into clinical practice.

One key class of materials are LMs such as eutectic gallium–indium (EGaIn), which remain in a liquid state at room temperature while exhibiting metal-like electrical conductivity.^122–126 EGaIn and other gallium-based LM alloys exhibit exceptional stretchability and possess intrinsic softness and biocompatibility, making them highly suitable for wearable and implantable devices.^127–129 Unlike solid metal films that crack under strain, LMs simply deform without fracturing, even under significant stretching.^130–132 For example, microfluidic channels filled with LMs have been used to create stretchable interconnects and sensors that maintain conductivity under repeated elongation.¹³³ As LMs are in a liquid state, they can be easily filled to the microfluidic channels using vacuum filling (Fig. 4A).¹³⁴ A practical consideration with LMs is the need for encapsulation (e.g. within elastomeric microchannels) to contain the liquid and prevent oxidation of the metal as gallium instantly forms a thin oxide skin (few nanometers thick). Despite the challenges associated with oxidation and containment, the unique combination of high electrical conductivity, intrinsic self-healing of ruptured conductive pathways, and exceptional deformability makes LMs particularly well-suited for use as interconnects in wearable and implantable microfluidic electronics, where conventional rigid metal wiring would be prone to fracture under mechanical strain.¹³⁵

	Fig. 4 System-level stretchability. A) Complex branched tree design completed with vacuum filling; scale bar: 5 mm. Reproduced with permission from ref. 134. Copyright 2017 The Royal Society of Chemistry. B) SEM images of the Au-TiO2 NWs with continuous gold coating; scale bar: 500 nm. Reproduced with permission from ref. 141. Copyright 2018 John Wiley and Sons. C) Schematic illustration of microstructures of polyacrylamide-based electrically conductive hydrogel. Reproduced with permission from ref. 143. Copyright 2023 Springer Nature. D) Optical images and an exploded view schematic illustration (lower left inset) of a skin-like composite that consists of a lithographically defined wavy filamentary network of polyimide, analogous to a collagen/elastin structure, embedded in a soft breathable elastomer, analogous to a biological ground substance. The image shows this material wrapped onto the tip of the thumb; scale bar: 1 cm. Reproduced with permission from ref. 145. Copyright 2015 Springer Nature. E) Photographs of hydrogel electronics with helical LM channels under 0% and 150% strain; scale bar (right): 5 mm. Reproduced with permission from ref. 148. Copyright 2018 John Wiley and Sons. F) An exploded-view drawing of a stiffness tunable probe with 3 layers of PDMS thin-films integrated with Pt electrodes for electrochemical sensing. Reproduced with permission from ref. 149. Copyright 2019 Elsevier. G) Capture and release of 20 μm particles with a flow rate of 200 μL min−1 at 0 mm and 4 mm stretching lengths; scale bars: 200 μm. Reproduced with permission from ref. 154. Copyright 2022 The Royal Society of Chemistry.

Another important class is stretchable conductive composites, typically composed of an elastomer matrix loaded with conductive fillers.^136,137 Examples include silicone or polyurethane embedded with silver nanowires, gold nanoparticles, carbon nanotubes (CNTs), graphene, or even LM droplets.^128,138,139 These composites leverage the inherent elasticity of polymers and the conductivity of metal or carbon fillers. For instance, silver nanowire networks in elastomers can form percolating pathways that remain intact under moderate strains (often up to 20–50% or more), enabling stretchable electrodes and wires. Such composites can achieve conductivities on the order of 103–105 S m⁻¹, which while lower than bulk metals, is often sufficient for bioelectronic applications.¹⁴⁰ Tybrandt et al. developed a soft, stretchable neural electrode grid using a gold-coated titanium dioxide nanowire (Au-TiO₂ NW) composite embedded in PDMS, achieving high conductivity (∼16000 S cm⁻¹), reliable performance up to 100% strain, and stable chronic neural recordings for over 3 months (Fig. 4B).¹⁴¹ Strategies like optimizing filler shape (e.g. 1D nanowires vs. 0D particles), surface functionalization for better adhesion, and creating microstructured “wavy” filler networks help improve the stability of conductivity during stretching. Recent studies have also developed composite inks for printing stretchable circuits, combining ease of fabrication with good mechanical compliance. A challenge with solid-particle composites is that large strains can still cause microcracks or filler network breakdown at very high deformations, but innovations like adding second-phase nanoparticles to bridge gaps or using soft conductive fillers (e.g. LM droplets) can mitigate this.

Conductive hydrogels represent a third class of stretchable materials, particularly attractive for bio-integrated devices.¹⁴² These hydrogels are polymer networks swollen with water or ionic liquid, often incorporating ionic salts or conductive polymers to facilitate charge transport. Their water-rich, tissue-like consistency yields excellent biocompatibility and matching of mechanical properties with human tissue. Intrinsically, purely ionic hydrogels have lower electrical conductivity than metals (typically only ionic conductance), but they can sustain very large strains (100–1000%) without damage. To boost their electrical performance, hydrogels are frequently enhanced with conductive components: for example, adding carbon nanomaterials, conjugated polymer chains (like PEDOT:PSS), or even dispersing nano/microparticles of gold or EGaIn. These hybrid conductive hydrogels maintain softness while achieving improved conductivity and are being explored for stretchable electrodes, sensors, and even circuits. A notable advantage of hydrogels is their biocompatibility and sometimes biodegradability, which can minimize foreign-body reactions and enable transient implantable devices. Chong et al. introduced a template-directed assembly approach to fabricate PEDOT:PSS-based hydrogels with record-high conductivity (∼247 S cm⁻¹), tissue-like modulus, and excellent stretchability (>600%), offering a promising platform for suture-free and adhesive-free bioelectronic interfaces in neuromodulation and electrophysiology applications (Fig. 4C).¹⁴³ However, hydrogels can dry out in vivo and purely ionic conduction is slower; thus ongoing research focuses on improving long-term stability (e.g. via hydrogel encapsulations or hydrogel–LM composites). In summary, by combining these materials (e.g., highly conductive LMs, robust elastomeric composites, and bio-friendly conductive hydrogels) engineers can achieve the required conductivity, stretchability, and biocompatibility for wearable and implantable microfluidic electronics.

Stretchable conductive materials such as liquid metals and composites bridge the mechanical mismatch between devices and soft tissue, enabling durable operation under strain. However, oxidation, encapsulation complexity, and long-term reliability under cyclic deformation remain unresolved. Future efforts should focus on self-healing materials and scalable fabrication methods for integrated stretchable systems.

Beyond serpentines, other architected structures provide stretchability. Kirigami patterns, inspired by paper cutting, introduce deliberate cuts or perforations in a material, which open up under strain to allow expansion. For instance, a thin film with an array of slit cuts can expand like a net when pulled, achieving large stretching with the intact segments deforming mainly by rotation and bending rather than tension.¹⁴⁷ Kirigami-based stretchable sensors and electrodes have shown strains of several hundred percent while maintaining function, by tailoring the shape, distribution, and orientation of cuts. This approach is powerful because it can turn even a normally rigid substrate (like paper or metal foil) into a stretchable mesh simply through patterning. Another design strategy uses fractal and biomimetic layouts. Fractal curves effectively elongate the path length within a given area, acting similarly to serpentines but with complex self-repeating motifs that improve strain distribution. A recent study demonstrated that a Peano-type fractal interconnect outperformed straight or simple wavy lines, showing lower hysteresis and higher stretchability in a microfluidic strain sensor. The third-order Peano fractal channels could be stretched nearly five-fold (∼490% strain) while maintaining reliable sensing performance. Such fractal designs mitigate stress concentrations and allow uniform extension, which is beneficial for sensitive devices. Similarly, helical coils and spring structures (in 3D) can enable stretchability – for example, a wire formed into a coil can extend and compress like a spring. Researchers have used helical microsprings for stretchable electrodes that accommodate motion in implants. For instance, Liu et al. developed hydrogel-based stretchable electronics with embedded 3D helical LM channels, achieving excellent mechanical compliance, stable electrical performance under 300% strain, and even full functional recovery after dehydration through simple rehydration (Fig. 4E).¹⁴⁸ These structural techniques ensure that as the device stretches, the strain energy is dissipated through geometrical changes rather than concentrated in the active materials, thus maintaining functionality under deformation. In summary, through clever designs like serpentine interconnects, kirigami cuts, and fractal/mesh layouts, system-level stretchability is achieved even if some constituent materials are not intrinsically stretchable.

Structural designs like serpentines and kirigami enable system-level deformability even with partially rigid components, offering a low-cost route to mechanical compliance. Nevertheless, optimizing these architectures for high-density functional integration and long-term fatigue resistance is still challenging. Combining structural engineering with intrinsically soft materials can further enhance durability in dynamic biological environments.

On the wearable side, skin-mounted microfluidic patches demonstrate system-level stretchability for diagnostics and therapy.^54,150 These devices typically consist of an array of microchannels and chambers in a stretchable polymer (e.g. PDMS), which can collect and analyze biofluids like sweat or ISF.¹⁵¹ A recent sweat-sensing patch integrates serpentine microfluidic channels that transport sweat to chemical assay wells on a flexible disk that sticks to the skin. The entire system stretches and bends with the skin, thanks to its thin silicone construction and wavy channel geometry. Such patches have been used to continuously measure electrolyte levels, metabolites, and even hormone indicators of stress, all while the wearer performs vigorous activity. By providing real-time biomarker data, these stretchable microfluidic sensors can inform digital therapeutic algorithms (for instance, detecting dehydration or stress and advising adjustments). Beyond diagnostics, transdermal drug delivery systems also benefit from stretchable designs.¹⁵² Flexible electronic bandages with microneedles or ultrasound emitters can adhere to dynamic skin surfaces and administer drugs in a controlled manner. For example, stretchable patches integrated with microneedle arrays have been developed for pain management and insulin delivery; they conform to joints or muscles and release drugs through the skin without hindering motion. Another approach is an osmotic-engined microfluidic patch that can achieve sustained drug release (over 24 hours) by using the body's own fluids as a driving force, all in a thin, elastic form factor that can be worn like a Band-Aid.¹⁵³ Stretchable microfluidic designs can also function as a trigger of releasing typical particles and cells. Fallahi et al. developed a stretchable microfluidic cell trapper composed of U-shaped microstructures that enables on-demand and size-selective release of particles and cells by simple mechanical elongation, offering a deterministic and label-free approach for single-cell studies and downstream analysis (Fig. 4G).¹⁵⁴

Importantly, fully stretchable therapeutic systems remain an active area of research, with challenges and opportunities ahead. Ensuring long-term reliability of stretchable components (wires, sensors, microfluidic pumps) is critical – repeated strain can cause fatigue, so materials like nanomesh conductors or self-healing composites are being explored. Biocompatibility and integration with the body's environment is another concern: devices must endure sweat, temperature changes, and immune responses. Stretchable encapsulation strategies (e.g. silicone overcoatings or permeable membranes) can protect electronics while allowing biofluid exchange. Energy and communication for these systems also demand innovative solutions, such as stretchable batteries or energy harvesting, and wireless modules that deform without damage. Despite these challenges, the prospects are exciting. Stretchable microfluidic electronics hold great promise for enabling closed-loop “digital therapeutic” capabilities – for instance, through skin patches that sense biomarker spikes and automatically release drug doses, or neural implants that monitor brain signals and delivers responsive neurostimulation. As materials and structural designs continue to improve, we anticipate stretchable therapeutic systems that are as soft and resilient as human tissue. This will enable truly seamless integration of medical devices with the body, advancing personalized medicine. Researchers envision a future of bioelectronics that can be wrapped around organs, woven into clothing, or applied to skin like stickers, all while providing real-time health interventions.^142,155 The convergence of stretchable materials, smart structural design, and microfluidic functionality thus forms a foundation for the next generation of digital therapeutics. Each innovation at the material or structural level directly translates to greater comfort, efficacy, and possibilities in the complete system, bringing medical technology closer to the patient in daily life.

While significant progress has been made in the development of stretchable microfluidic systems with soft, compliant materials, the practical deployment of such systems often necessitates integration with rigid components, including integrated circuits, batteries, antennas, and wireless communication modules.¹⁵⁶ These rigid–stretchable hybrid interfaces inherently introduce mechanical challenges that can compromise system reliability, especially under dynamic or long-term operation.^157,158

Mechanically, the large mismatch in elastic modulus between rigid (typically in the GPa range) and stretchable (typically in the kPa to low MPa range) materials leads to stress concentrations at their interface. These stress points can cause delamination, fracture, or performance degradation during repeated deformation such as bending, stretching, or twisting. To mitigate such failures, various strategies have been explored, including the use of strain-isolating geometries (e.g., serpentine interconnects), interfacial adhesive layers, and gradient modulus materials. However, these methods often add complexity to fabrication processes and may not be universally applicable across diverse device architectures.

From a microfluidic perspective, rigid–soft fluidic junctions also pose sealing and alignment issues, particularly when integrating soft microchannels with hard reservoirs or sensors. Long-term fluidic integrity and interface robustness are essential for applications involving chronic use or wearable deployment.

Addressing these integration challenges is critical for enabling next-generation soft electronics and digital therapeutics. Future research should focus on scalable, mechanically adaptive, and electrically reliable integration schemes to bridge the gap between rigid and stretchable domains.

In this context, the integration of microfluidic system, that is capable of analyzing biofluids—such as sweat, saliva, or ISF—with physical or electrophysiological sensors has gained considerable interest. These hybrid platforms allow real-time chemical sensing to be combined with vital sign monitoring, offering synergistic insights into both the internal biochemical environment and the body's dynamic physiological responses. For example, Song et al. presented an epifluidic electronic skin for multimodal health monitoring.¹⁶⁹ The authors integrated microfluidic channel-based electrochemical sensors to detect the chemicals in sweat with biophysical sensors to observe the body temperature and heart rate (Fig. 5A). Sweat is induced efficiently using hydrogel-coated iontophoresis electrodes, which actively stimulate localized sweat production. The hydrogel contains the muscarinic agent carbachol, enabling autonomous and prolonged sweat extraction without the need for vigorous physical activity or elevated ambient temperature. The collected sweat is transported through the microfluidic channels to a reservoir containing three biosensors, each targeting glucose, alcohol, and pH levels (Fig. 5B). Meanwhile, Ti₃C₂T_x (MXene)-based electrodes measure biophysical parameters such as body temperature and heart rate, enabling multiplexed physiochemical monitoring during daily activities, including beer intake (Fig. 5C). Owing to this multimodality of health monitoring dataset, the authors developed a machine learning model to predict the physiological impact of alcohol consumption, demonstrating the power of multimodal sensing for personalized health analysis.

	Fig. 5 Multi-modal bioelectronics including microfluidic system. A) Multimodal skin patch for health monitoring by measuring biochemical and biophysical signals. B) Microfluidic system integrated on the skin patch for sweat collection. C) Detection of various biosignals when the subject had a beer. Reproduced with permission from ref. 169. Copyright 2023 American Association for the Advancement of Science. D) Implantable neural probe with microfluidic-based drug delivery system. E) Behavioral change before (left) and after (right) the injection of bicuculline (BIC). F) Monitoring of changes in neural activities by BIC injection using multimodal neural probe. Reproduced with permission from ref. 64. Copyright 2022 Springer Nature.

Similarly, Xu et al. introduced a wearable skin patch capable of detecting multiple sweat metabolites (glucose, lactate, and uric acid) and electrolytes (Na⁺, K⁺, NH₄⁺), using a microfluidic collection system.¹⁷⁰ Sweat induction is achieved via iontophoresis, while stable transport through the microchannels ensures consistent amount of sweat during movement and physical exertion. In addition to chemical sensing, the patch includes physiological sensors for detecting pulse waveform, galvanic skin response, and temperature. This enables comprehensive monitoring of an individual's physiological response to various stressors. The authors collected multiplexed data by exposing participants to three different stressors and conducted questionnaires to quantify the stress level for each stressor. Using the resulting multimodal dataset, the authors successfully trained a machine learning model to classify responses to different stressors and predict individual anxiety levels, which cannot be feasible with a single biosignal alone.

Consequently, numerous studies underscore the growing potential of microfluidic-integrated, multimodal bioelectronics in advancing personalized healthcare. When combined with machine learning algorithms, which have been rigorously evolving in recent years, such systems enable accurate classification and prediction of health states based on complex, real-time physiological data. This integration opens new avenues for proactive health management by providing data-driven therapeutic strategies tailored to individual needs.

Multimodal sensing platforms provide richer physiological insights than single-modality systems by capturing both biochemical and biophysical signals. Yet, combining multiple sensors increases system complexity and may introduce cross-interference or calibration issues. Efforts should aim at achieving compact integration, robust signal decoupling, and data fusion algorithms that can handle heterogeneous inputs in real-time.

One prominent application is the integration of drug delivery microfluidic channels into shank-type neural probes, which are widely used for recording neural signals from deep brain regions. These neural signals reflect essential neurological activities, which are the basis of vital activities, as well as providing critical information about pathological conditions, including neurodegenerative diseases. By integrating microfluidic channels for drug delivery into such neural probes, the devices can advance the capability of conventional neural probes for theragnostic platforms that enable them to monitor and treat diseases or modulate neural activities simultaneously. In this regard, Yoon et al. proposed a shank-type neural probe capable of wireless drug delivery through embedded microfluidic channels (Fig. 5D).⁶⁴ A precise amount of drugs is released to the local brain region via an integrated electrolytic pump. The system is wirelessly controlled, allowing users to adjust the dosage and monitor neural signal changes simultaneously. For example, when a GABA_A receptor antagonist, i.e., bicuculline (BIC), was delivered into the substantia nigra of a mouse, the mouse halted exploratory behavior and began compulsive circling (Fig. 5E). Furthermore, neural activity was recorded through sixteen microelectrodes embedded on the same shank, revealing an increased firing rate in response to BIC (Fig. 5F). This system clearly shows how integrated drug delivery and real-time neural recording can enhance understanding of neuromodulation and provide a powerful tool for neurological research and therapy.

Furthermore, to determine the optimal timepoint of drug release, multimodal closed-loop systems combining microfluidic system for drug delivery with biosignal sensors have been studied. When the detected biosignals indicate the abnormal health state of the users, the loaded drug is released to the targeted regions to treat the symptoms. This feedback system can maximize drug efficacy while minimizing side effects. As an example, Lee et al. presented a skin patch capable of both monitoring hemodynamic and cardiovascular health states and simultaneously delivering drugs using a peristaltic operation.¹¹⁷ This multi-modal device optically detects PPG as well as record ECGs, body movement, and body temperature. Based on real-time physiological data, drug release is controlled via wireless system by employing electro-resistive microheaters, which thermally expand an actuation chamber to dispense the medication. The authors demonstrated the operation of closed-loop system under various conditions, highlighting its potential for intelligent, responsive theragnostic devices.

Together, these advancements illustrate the transformative potential of microfluidic systems when combined with real-time biosignal sensing. By enabling precise, on-demand drug delivery in response to physiological feedback, such integrated platforms pave the way for intelligent, closed-loop theragnostic devices that can revolutionize personalized healthcare.

Integrating sensing with localized drug delivery enables true theragnostic capability and closed-loop intervention. Despite this advantage, miniaturizing the combined modules without sacrificing precision and ensuring biocompatible long-term operation remain key barriers. Future systems should emphasize autonomous control, refillable reservoirs, and seamless coordination between sensing feedback and actuation.

For example, Park et al. introduced a soft LM-based neural interface placed on the cranium, capable of recording neural signals from multiple deep brain regions without hindering the animal's natural movement (Fig. 6A).¹³³ The neural probes are fabricated by printing LM, which has a Young's modulus comparable to brain tissue. This mechanical compatibility minimizes damage to neural circuits and immune responses, enabling reliable long-term signal acquisition (Fig. 6B). This approach facilitates circuit-level observation of brain activity, which is crucial for understanding neurological conditions, as brain functions are distributed across interconnected regions. However, electrophysiological signals alone do not completely reflect brain activity. Neurochemicals such as neurotransmitters and hormones also play key roles in modulating brain function. Thus, combining chemical sensing via microfluidic systems can complement electrophysiological data and enable more comprehensive neurological analysis.

	Fig. 6 Combining microfluidic systems with state-of-the art devices. A) Soft liquid metal-based neural interface developed on the cranium; scale bar: 1 mm. B) Neural signals detected from multiple deep brain regions; scale bars: 200 μV (vertical), 500 ms (horizontal). Reproduced with permission from ref. 133. Copyright 2024 Springer Nature. C) Schematic illustration presenting the measurement of intraocular pressure (IOP) using a pressure-sensitive transistor-based IOP sensor. D) Fluctuation of IOP following hyaluronic acid injection. Reproduced with permission from ref. 89. Copyright 2024 American Association for the Advancement of Science. E) Schematic illustration presenting the application of wireless, implantable device for managing cardiac activities. F) Multi-modality of the device capable of optically pacing and monitoring cardiac responses. G) Recorded electrocardiogram (ECG) in response to optical pacing using micro-LEDs. Reproduced with permission from ref. 183. Copyright 2022 American Association for the Advancement of Science.

Furthermore, the various types of bio-sensors designed to detect disease-specific biosignals have shown promising possibility for integration with microfluidic-based drug delivery system for efficient treatment.^11,175,176 For example, IOP is a critical biomarker for predicting glaucoma, an eye condition where increased pressure damages the optic nerve. Seo et al. developed a pressure-sensitive transistor-based IOP sensor to monitor IOP level in both the anterior and vitreous chamber of the eye (Fig. 6C).⁸⁹ This device reliably detected the IOP fluctuation following hyaluronic acid injection, indicating its feasibility to predict the progression of glaucoma (Fig. 6D). Integrating such biosensors with microfluidic drug delivery could enable closed-loop treatment systems that respond directly to pathological signals.

To further improve user comfort and enable untethered operation, wireless communication systems have been increasingly adopted in bioelectronics devices.^{123,177–179} This is particularly important for implantable electronics, where wired systems can generate tethering forces that lead to immune responses and limit long-term usability.^180–182 Magnetic resonant coupling-based wireless communication system is widely used into implantable bioelectronics since they can power devices without the need for bulky, additional batteries. In this regard, Ausra et al. designed the wireless circuit to supply sufficient power to micro-LED for optical stimulation and electrical recording (Fig. 6E and F).¹⁸³ The authors also demonstrated reliable cardiac pacing using six micro-LEDs and stable recording of cardiac responses, presenting the capability of this device to wirelessly modulate physiological activities of freely moving animals (Fig. 6G). When combined with these advanced technologies, microfluidic system can also be operated wirelessly and integrated into fully implantable platforms. This offers opportunities for developing intimate, multimodal bioelectronic systems capable of both monitoring and intervention.

As discussed throughout this section, there remain many untapped opportunities for integrating microfluidic systems into diverse types of bioelectronic platforms. Such integration will play a pivotal role in advancing the development of personalized, closed-loop therapeutic devices for precise and efficient health management. Continued research in this area will be key to unlocking the full potential of intelligent, patient-specific biomedical systems.

Integration with advanced bioelectronics greatly expands microfluidic functionality, enabling multimodal and wireless operation. However, combining multiple modalities increases fabrication complexity and potential interference between modules. Achieving seamless co-integration while maintaining device miniaturization and chronic stability remains a critical next step.

	Fig. 7 AI-enabled analysis and feedback. A) Categorization of ML algorithms. B) A meta learning framework based on TD-C learning. Reproduced with permission from ref. 204. Copyright 2023 Springer Nature. C) Image of a wearable, forehead-mounted patch; scale bars: 1 cm. D) Confusion matrix showing a high accuracy between two devices. Reproduced with permission from ref. 205. Copyright 2023 American Association of the Advancement of Science. E) SAAS based on piezoelectric micromachined ultrasonic transducers. F) Confusion matrix of sentence recognition task. G) Extracted features using t-SNE to visualize distinct clusters of different vocal states. Reproduced with permission from ref. 43. Copyright 2025 Springer Nature. H) Schematic of AI-enabled closed-loop system through BMI. I) Pain onset detection applying the SSM-based decoding strategy. Reproduced with permission from ref. 208. Copyright 2023 Springer Nature. J) Schematic illustration of implantable device (left) and electrical architecture including power management circuits (right). Reproduced with permission from ref. 114. Copyright 2023 Springer Nature.

Supervised learning is a machine learning approach that models the relationship between input and output data to make predictions, assuming that each input is paired with a known label. Depending on whether the target output is continuous (numerical) or discrete (categorical), supervised learning tasks are typically categorized into regression and classification problems.¹⁸⁹

Regression algorithms are used to predict continuous numerical values. A representative example is linear regression, which estimates outputs based on a linear relationship between input variables. While linear regression is computationally efficient and easy to interpret, it performs poorly with non-linear data and is susceptible to performance degradation in the presence of multicollinearity among input features.¹⁹⁰

In contrast, artificial neural networks (ANNs) are flexible models capable of capturing complex non-linear relationships.¹⁹¹ They consist of an input layer, one or more hidden layers, and an output layer. Deep neural networks (DNNs), which extend ANNs by adding more hidden layers, can learn high-dimensional and abstract features, leading to improved accuracy in tasks such as the prediction and classification of physiological signals.

Support vector machines (SVMs) are classification algorithms that identify the optimal hyperplane that maximizes the margin between different classes in a high-dimensional space.¹⁹² SVMs are particularly effective for binary classification and can be adapted for non-linear problems using kernel functions. However, they require careful parameter tuning and are less intuitive and harder to interpret in multi-class settings.

Decision trees are rule-based models that split data based on feature values to arrive at a prediction. They are easy to understand and visualize but are prone to overfitting and sensitive to noise. Random forests, an ensemble method that aggregates the predictions of multiple decision trees, mitigate these issues by improving model stability and generalization performance through averaging or majority voting.¹⁹³

Unsupervised learning is a machine learning approach that identifies hidden patterns or latent structures within unlabeled data. This approach mainly encompasses clustering and dimensionality reduction techniques, both of which are extensively applied in biosignal analysis tasks such as classification, visualization, and data preprocessing.¹⁹⁴

Clustering aims to uncover the underlying structure of data by grouping samples with similar features.¹⁹⁵ One of the most commonly used algorithms is K-means clustering, which partitions the data into a user-defined number of clusters (K), each represented by a centroid.¹⁹⁶ Although K-means is computationally efficient and straightforward to implement, it is sensitive to the initial placement of centroids.

Another widely used method, hierarchical clustering, builds clusters through iterative merging (agglomerative) or splitting (divisive) based on the similarity between data points. The resulting dendrogram visually represents the nested structure of clusters.¹⁹⁷ While this method offers insightful hierarchical relationships, it can be computationally intensive and less suitable for large-scale datasets.

Density-based spatial clustering of applications with noise (DBSCAN) is a density-based clustering algorithm that defines clusters as areas with a high density of data points.¹⁹⁸ It is particularly effective at detecting clusters of arbitrary shapes. However, performance of DBSCAN may deteriorate when clusters exhibit significant differences in density, and it is sensitive to hyperparameters such as the distance threshold and the minimum number of points required to form a cluster.

Dimensionality reduction techniques aim to project high-dimensional data into a lower-dimensional space while preserving its intrinsic structure, thereby facilitating visualization and analysis. Principal component analysis (PCA) is a linear method that identifies new orthogonal axes (principal components) aligned with the directions of greatest variance in the data.¹⁹⁹ PCA serves as a computationally efficient technique, well-suited for denoising and revealing inter-variable relationships, though it may not capture complex nonlinear patterns.

To address such limitations, t-distributed stochastic neighbor embedding (t-SNE) offers a nonlinear approach to dimensionality reduction by preserving local structures through modeling pairwise similarities in both high- and low-dimensional spaces.²⁰⁰ Especially useful for visualizing cluster structures in complex datasets, t-SNE has found broad applications in biosignal-based pattern recognition. Despite its strengths, the method demands significant computational resources, with results heavily influenced by initialization and hyperparameter settings such as perplexity.

Reinforcement learning (RL) is a machine learning paradigm in which an agent learns the optimal policy to maximize its cumulative reward by interacting with the environment.²⁰¹ Unlike supervised learning, RL learns through trial and error without the need for explicit labels, with feedback provided in the form of rewards. It is particularly well-suited for sequential decision-making problems in dynamic and uncertain environments, where each action can influence future states and rewards.

A key framework for formalizing RL problems is the Markov decision process (MDP).²⁰² MDPs model the environment using states, actions, transition probabilities, and reward functions, and rely on the Markov property, which states that the future state depends only on the current state and action. However, in real-world applications, it is often difficult to know the exact model, such as the transition probabilities or reward functions.

Q-learning is a value-based RL algorithm that learns an optimal Q-function, which estimates the expected cumulative reward of taking a specific action in a given state.²⁰³ By performing iterative updates, Q-learning converges toward the optimal policy. Simplicity and flexibility of Q-learning make it widely applicable. However, when the state space is large or continuous, tabular Q-functions have limitations, necessitating the use of function approximations or neural networks.

Monte Carlo methods are model-free techniques that calculate rewards over episodes of experience and use these to evaluate and improve policies. This approach is particularly useful for identifying long-term reward structures but can be inefficient for real-time learning since entire episodes must be completed before updates can occur. Machine learning algorithms for biosignal analysis are summarized in Table 3.

Machine learning enables powerful pattern recognition in complex biosignal datasets, but performance depends on data quality, labeling, and generalizability across individuals. Balancing model complexity with interpretability and ensuring robustness in real-world conditions are ongoing challenges. Future systems must adopt adaptive algorithms capable of learning with minimal calibration.

Kim et al. introduced a substrate-less nanomesh receptor engineered to detect proprioceptive signals through resistance changes induced by skin deformation.²⁰⁴ The acquired biosignals are analyzed through a meta-learning framework based on unsupervised learning, which utilizes time-dependent contrastive learning (TD-C learning) to extract meaningful features in a user-independent and data-efficient manner for recognizing various hand gestures (Fig. 7B). While TD-C learning is not discussed in section 2 due to its deviation from traditional classifiers such as kernel- or tree-based models, it represents a type of self-supervised learning that exploits temporal continuity to learn representations from unlabeled data. The system demonstrates rapid adaptability to new users and tasks, achieving over 80% accuracy after just 20 training epochs.

Kwon et al. developed a wearable, forehead-mounted patch designed to monitor electrophysiological signals, including EEG and EMG, during sleep (Fig. 7C).²⁰⁵ They proposed an integrated system capable of automatically classifying sleep stages and detecting sleep apnea events. The system employs a convolutional neural network (CNN), a subclass of DNNs composed of multiple hierarchical layers, to analyze EEG signals for the purpose of sleep state classification. As a supervised learning-based classification algorithm, the CNN was trained using data collected from both healthy participants and individuals with sleep apnea, enabling automated assessment of sleep quality and related disorders. The model architecture, incorporating batch normalization and max-pooling layers, demonstrated effective feature extraction from EEG data and achieved an accuracy of approximately 88.5% in real-time sleep stage classification (Fig. 7D). This performance is comparable to that of conventional polysomnography (PSG), underscoring the potential of the proposed patch as a reliable, at-home sleep monitoring and diagnostic tool.

In a more recent study, Liu et al. propose a skin-attachable acoustic sensor (SAAS) based on piezoelectric micromachined ultrasonic transducers (PMUTs) to capture subtle vibrations from vocal cords and surrounding skin for robust speech recognition (Fig. 7E).⁴³ Integrated with a CNN, specifically a residual network (ResNet) architecture, the system classifies ten vocal states with high accuracy, achieving 99.8% on both training and validation datasets (Fig. 7F). To assess the separability of extracted features, the authors applied t-SNE, which visualized distinct clustering of different vocal states in feature space after training (Fig. 7G). These results, supported by confusion matrix analysis, demonstrate the system's effectiveness for speech interaction, particularly benefiting individuals with speech impairments.

Similar to the previous advancements achieved through the application of AI in signal processing and data analysis, the integration of AI with microfluidic systems for analytical purposes is expected to overcome existing limitations and enable more precise digital therapeutics.

AI-driven processing enhances the sensitivity and autonomy of wearable and implantable platforms, yet requires efficient on-device computation and energy management. Maintaining privacy while handling high-dimensional physiological data is another key hurdle. Continued development of lightweight, secure, and adaptive AI models will be critical for clinical adoption.

Zhang et al. introduced an AI-enabled closed-loop system that decodes pain states in real-time through a brain–machine interface (BMI) and dynamically adjusts neuromodulatory interventions accordingly (Fig. 7H).²⁰⁸ Neural spike data recorded from the somatosensory cortex of mice were used to develop a Bayesian decoder based on a state-space model (SSM), trained using a supervised learning approach. The model was constructed by mapping neural activity to labeled pain states, defined by the presence or absence of nociceptive stimuli. Once trained, the decoder continuously interpreted neural spike patterns to estimate ongoing pain states and autonomously activated optogenetic stimulation when pain was detected (Fig. 7I). This closed-loop architecture enabled precise, condition-specific neuromodulation, initiating therapeutic intervention exclusively during nociceptive episodes. The study underscores the potential of BMI-integrated, AI-driven systems for targeted pain management and the broader treatment of neurological disorders.

While the wireless, battery-free implantable system developed by Ouyang et al. was previously discussed, the same system utilizes machine learning-based closed-loop feedback mechanism, which plays a central role in enabling autonomous neuromodulation (Fig. 7J).¹¹⁴ The system utilizes a supervised learning-based CNN model to analyze EEG data and detect seizures in real-time. Once a seizure is predicted, the system automatically delivers pharmacological treatment and has the capability to administer optogenetic stimulation, which is crucial for suppressing or treating seizures. The supervised learning-based algorithm identifies patterns in neural signals to predict seizures, with an accuracy rate of 0.92. The system also demonstrates reliable performance in long-term animal models and can continuously track sleep–wake cycles in freely moving animals. With AI-powered autonomous control, this technology enables real-time monitoring of acute, life-threatening conditions such as seizures and provides immediate intervention, offering significant promise for future neurorehabilitation and the treatment of neurological disorders.

As demonstrated by the aforementioned studies, AI plays a pivotal role not only in interpreting biosignals but also in enabling personalized therapeutic feedback systems that respond in real-time to a patient's condition. Future microfluidic systems are expected to adopt this feedback-driven architecture, which will facilitate early disease detection and prognosis. Building on this foundation, these systems may evolve into fully closed-loop digital therapeutics that autonomously optimize interventions based on an individual's physiological state. This approach will enhance treatment accuracy and responsiveness while minimizing the need for user intervention, ultimately contributing to more sustainable and accessible healthcare in daily life.

AI-integrated closed-loop systems demonstrate transformative potential in personalized therapy but raise concerns about safety, latency, and algorithm transparency in life-critical interventions. Achieving regulatory approval will require rigorous validation of both hardware and AI algorithms. Future designs must ensure fail-safe operation while maintaining adaptability to individual patient variability.