Smart closed-loop drug delivery systems

Abstract

The administration of therapeutics for long-term chronic disease management or treatment faces considerable challenges, such as the need for precise dosage control, timely delivery and adherence to medication regimens. Traditional drug delivery methods often result in suboptimal therapeutic outcomes owing to variable responses, fluctuating drug concentrations and lack of feedback from real-time monitoring. Smart closed-loop systems (CLSs) could address these limitations by integrating real-time biosensing with automated drug delivery, thereby personalizing treatments to individual needs. This Review explores the current landscape of CLSs, highlighting recent advancements in wearable and implantable technologies that facilitate continuous monitoring of biomarkers and offer responsive therapeutic interventions. We discuss the implications of device design and the trade-offs between wearable and implantable systems. In addition, we highlight the potential of artificial intelligence enhancement of CLS control algorithms by enabling systems to learn from and predict responses to achieve more effective and adaptive optimal therapies. Ultimately, this Review charts a path towards next-generation CLSs, emphasizing the integration of synthetic biology and engineered cells into implantable devices.

Graphical Abstract

Introduction

Traditional drug delivery, particularly for chronic diseases^1-3, faces considerable challenges, including poor adherence, variable absorption rates and side effects of fluctuating drug levels. Frequent dosing schedules further strain patients and health-care systems financially and logistically. Personalized medicine offers a promising alternative, tailoring drug delivery systems (DDSs) to individual physiological and metabolic profiles. Despite recent advances, managing chronic conditions still requires innovative strategies to address adherence issues in long-term therapies.

Innovations in DDSs are addressing precision and adherence challenges through closed-loop systems (CLSs), which integrate sensors, controllers and actuators. Sensors monitor physiological parameters in real time, transmitting data to the controller that determines the therapeutic response using adaptive algorithms. The actuator then adjusts the dosage of drug delivered based on this feedback, thus ‘closing the loop’. Together, these components dynamically adapt therapies based on patient-specific data in real time, mimicking the body’s feedback mechanisms⁴. Unlike traditional CLSs, modern ‘smart’ CLSs autonomously adapt to dynamic conditions and user needs without human intervention or fixed control strategies. Closed-loop glucose monitoring systems interfaced with insulin pumps exemplify successful personalized drug delivery. These systems adjust insulin delivery in real time based on glucose levels monitored continuously via subcutaneous sensors, a considerable improvement over traditional diabetes management methods. Similarly, the development of smart inhalers for asthma and implantable neurostimulators for pain management and epilepsy, which adjust their output based on real-time feedback from sensors, is expanding the scope of CLSs. Advances in sensors, remotely controlled drug delivery devices, power solutions and communication technologies are enhancing CLS capabilities and enabling integration with telemedicine platforms. These developments position smart CLSs as transformative tools for personalized medicine across diverse clinical applications.

This Review delves into these advancements, beginning with a detailed exploration of sensors, DDSs and control algorithms. It then examines the integration of these components into cohesive smart CLSs, providing insights into technological innovations, clinical applications and future perspectives. Finally, the Review addresses the challenges these systems must overcome to achieve widespread clinical implementation. By understanding these developments, the scientific and medical communities can better adapt to the evolving landscape of therapeutic management and enhance patient-specific treatments.

Sensors in closed-loop systems

Importance and working mechanism

The decision-making process of any CLS relies on input data from the sensor. The sensor recognizes and measures dynamic fluctuations in biomarker levels in the biofluid or the biophysical parameter; based on the readout, the delivery platform takes automated action to tune the dosage and duration of the counteracting drug. Accurate sensors can distinguish specific biomolecules in complex biological environments, ensuring precise dosing for personalized treatment⁵. However, long-term sensor performance is challenged by issues like surface fouling and bioreceptor degradation, which can compromise stability and sensitivity. Design innovations such as antifouling coatings and biocompatible materials are crucial for improving sensor longevity and performance. CLS sensors fall into two categories: biochemical and biophysical (Fig. 1), with signal generation mechanisms that might be continuous or sporadic, depending on the transduction method and application⁶.

Fig. 1 ∣. Sensors in closed-loop systems.

a, A typical closed-loop system uses sensor data to make an intelligent decision for the dosage and duration of the drug. Biochemical sensors used in closed-loop systems are mostly electrochemistry-based, using different bioreceptors to recognize and detect the target biomarker from varying biofluids, and they can be deployed in various wearable and implantable form factors. Biophysical sensors involving optical and electrical modalities can be combined with these biochemical sensors to develop hybrid monitoring platforms, which can deliver multi-source chemical–physical data to the drug delivery module. b, Various wearable and implantable sensing platforms hold promise for closed-loop system development and expansion. These range from continuous glucose monitors^202,203, multiplexed microneedle arrays²⁴ and wound-sensing and healing systems¹²⁸ to hybrid, multi-modal (biochemical and biophysical sensing) platforms⁵¹. BP, blood pressure; CGM, continuous glucose monitor; GI, gastrointestinal; HR, heart rate; I, current; ISF, interstitial fluid; MIP, molecular imprinted polymer; R, resistance; RE, reference electrode; SpO₂, oxygen saturation; V, electrical potential; WE, working electrode.

Biochemical sensing quantifies biomolecules on a functionalized transducer (for example, an electrode). Surface-confined biorecognition elements, such as enzymes, antibodies, nucleic acids, aptamers, molecular imprinted polymers and ionophores, sense metabolites, proteins, hormones, drugs or electrolytes^7,8. The interaction between the analyte and the immobilized receptor generates a signal change (chemical, mechanical, electrical, optical or acoustic) proportional to the analyte concentration (Fig. 1a). The reversible nature (reusability) of aptamer-based biosensors addresses the limitation of irreversible antibody immunoassays in the continuous monitoring of drugs and hormones in various diseases⁹. CLSs primarily rely on wearable or implantable electrochemical sensors for their high analytical performance, miniaturization capabilities, flexibility and low power needs, whereas wearable optical fibre sensors are still nascent¹⁰. Continuous sensors enable more precise, dynamic drug dosing with minimal human intervention, avoiding issues with single-use sensors, such as not capturing analyte fluctuations and the need for frequent sensor changing. Monitoring non-blood biofluids requires additional effort to establish the correlation between biomarker concentrations and gold-standard blood-based assays while accounting for individual variability^11,12. Miniaturized, low-power, flexible, skin-interfacing electrochemical sensors (that is, potentiostats) can measure these signals and transmit the data to wireless readers, such as mobile phones and tablets¹³.

Biophysical sensors measure vital signs (electrical signals, motion, temperature, skin properties, blood pressure and vascular dynamics) and hence do not require an immobilized receptor, avoiding surface fouling issues. The transducer uses electrical (potential, resistive or capacitive), optical or acoustic stimulation to measure the bio-signals¹⁴. Techniques such as electrocardiography, electromyography and electroencephalography, which measure electrical signals, aid in assessing disorders and overall health (Table 1). Optical sensors incorporate photoplethysmography (measuring blood volume changes in tissues) to calculate heart rate, oxygen saturation and blood pressure. Other optical techniques, such as fibre optics and surface plasmon resonance, are less desirable than photoplethysmography and electrode-based sensors for CLS due to miniaturization and mechanical resiliency restrictions, implementation challenges in wearables and high cost¹⁵. Temperature sensors monitor surface and core body temperature, and skin conductance values (galvanic skin response) track stress levels¹⁶. Biophysical sensors in commercial and clinical applications are integrated into a range of devices from smart watches to state-of-the-art clinical instruments (for example, X-trodes Smart Skin System, NeuroOne Evo EEG Electrode, Biopac Systems)¹⁷.

Table 1 ∣.

Sensor metrics applicable to closed-loop systems

Source	Biochemical
	Access techniques	Advantages	Challenges	Refs.
Blood	Venipuncture (tubes), capillary collection (microfluidic channels)	Main biological fluid for all biomarkers	Rapid coagulation, contamination Haemolysis	21
Interstitial fluid	Microneedles (hollow, solid, soft) — sensing and extraction, reverse iontophoresis, vacuum and microdialysis	Most biomarkers correlate well with blood Closed-loop system opportunities for ‘artificial pancreas’	Vacuum suction generates high flow rate (~1 μl min−1) with high power usage Low extraction rate via only swelling (flow rate ~nl min−1) Precise microneedle design needed to reach interstitial fluid (height ~700–800 μm) Sensing layers vulnerable to biofouling Reverse iontophoresis requires current (<1.5 mA cm−2), which may cause patient discomfort	20,186
Sweat	Active: exercise, iontophoresis and thermal stimulation (with microfluidic channels) Passive: diffusion and osmosis (with paper channel)	Active: fast generation (~1–4 μl min−1) with easy management on skin	Active: excessive sweat rate causes biomarker dilution Passive: fully exertion-free with slow extraction rate (~0.1–0.5 μl min−1) on skin Biomarker correlation highly dependent on blood-to-sweat partition mechanism and sampling technique	52,187,188
Tears and saliva	Tears: contact lens Saliva: mouthguard and pacifiers	Easy and frequent sampling possible Rich in biomarkers Applicable for a broad population (infants, geriatric, people with needle phobia)	Tears: highly sensitive organ for continuous monitoring Saliva: prone to biofouling Biomarker correlation in both biofluids hold mixed reviews	189,190
Gastric fluid	Capsules	Allows non-invasive internal monitoring Portable and convenient	Highly acidic environment (pH ~1–2) Transmitted signal loss from surrounding tissues	36
Wound fluid	Bandages and patches	Several closed-loop system opportunities	Sensors are prone to biofouling	126,133
Exhaled breath	Face masks	Contains biomarkers for respiratory disease and metabolic disorders	Breath composition is complex Requires proper moisture management Vulnerable to environmental contamination	40
Cerebrospinal fluid	Needles (20 or 22 gauge)	Contains biomarkers for neurodegenerative disorders	Invasive access from a sensitive location	42
Technique	Biophysical
	Category	Advantages	Challenges	Ref.
Electrocardiography	Biopotential	Reliable detection of cardiac diseases Atrial fibrillation monitoring leads to stroke prevention	Susceptible to motion artefacts	191
Electromyography	Biopotential	Assessment of muscle health and function Versatile applications — motor rehabilitation, prosthetics	Prone to participant-induced noise Clinical standards are invasive	192
Electroencephalography	Biopotential	Valuable neurological condition monitoring (epilepsy, sleep disorders)	Limited spatial resolution and signal-to-noise ratio	193
Photoplethysmography	Optical	Clinical standard measurements for heart rate, oxygen saturation Low power, non-invasive with small footprint	Susceptible motion artefacts Skin variability-dependence of results — tone, temperature, perfusion	194

Biofluid extraction and sensing

Biomolecules are assessed in various biofluids, such as blood, interstitial fluid (ISF), sweat, saliva, tears, gastrointestinal fluids and wound fluids¹⁸ (Table 1). Blood, ISF, sweat, saliva and tears can be sampled directly and transported via microfluidic channels to the sensing zone. During disease onset, elevated biomarker levels first appear in the blood and disperse into other biofluids and organs at different rates and amounts, depending on biochemical pathways and individual physiology^19,20. Blood has been used in wearable microfluidic channels for continuous drug monitoring in rats and humans²¹. However, the invasive nature of blood sampling techniques (via fingerpick or venipuncture) and high biofouling rates present major barriers to developing blood-based CLSs.

ISF is the next most targeted biofluid, surrounding the cells and blood capillaries. This proximity enhances the biomarker diffusion rate from blood to ISF, leading to a strong correlation between ISF and blood biomarker levels with minimal time lag²². ISF can be extracted via microdialysis²², reverse iontophoresis²³ or hollow microneedles²², or directly accessed via microneedles and implantable sensors^20,24,25. Microdialysis is invasive, and reverse iontophoresis requires an electric field for extraction. Microneedles, by contrast, are the least invasive tool for ISF sampling and sensing. They can be solid, hollow or soft, and their length usually ranges between 500 and 800 μm⁸. The tip of the needle arrays can be modified with a biorecognition element for direct ISF sensing after skin penetration. Alternatively, the microneedle array can be used to extract ISF (with vacuum or osmosis, at flow rates of roughly nl per min) to conduct sensing outside the skin. ISF glucose has been the main biomarker of interest for needle-based wearable continuous glucose monitors (CGMs), owing to its commercial relevance and strong correlation with glucose levels in blood¹⁸. Other biomarkers (metabolites, ions and amino acids) have also been monitored in ISF to quantify hydration, hormonal imbalance and oxidative stress²⁶.

Sweat is the next most accessible and used biofluid, generated on the skin via exercise or iontophoresis or extracted via osmosis²⁷. Exercise, which involves physical exertion, and iontophoresis, which requires electrically induced transdermal delivery of sweat-stimulating agents, produce transient sweat rates of around 1–4 μl min⁻¹. Such high rates can dilute the biomarkers, reducing sensing accuracy, but this can be amended by including a correction factor computed from the measured sweat rate and variations in blood levels under different physiological concentrations²⁸. For low sweat rates (nl min⁻¹) generated without external stimuli, osmotic extraction is possible using hydrogels and paper microfluidic channels^28-30, in which hydrogel properties and placement dictate sensing dynamics. Several sweat biomarkers correlate well (average correlation coefficient is in the 0.7–0.9 range) with blood and are widely used to quantify hydration, oxidative stress, cortisol, pH and hormonal imbalance, gout status, diabetic severity, and neurochemical balance^27,31.

Biochemical sensors have been embedded into contact lenses and mouthguards/pacifiers for continuous biomarker monitoring in tears and saliva, respectively, but with limited CLS opportunities^32-35. Tear glucose, lactate, pH and cholesterol have been studied to understand ocular disease progression but show limited correlation with blood levels^{2 ,31}. Biomarker detection in epidermal fluids may lag compared with blood, but personalized calibration and machine learning algorithms can correct this lag¹². Ingestible encapsulated sensors can track biochemical and biophysical signals in the gastrointestinal tract for diseases such as inflammatory bowel disease, diabetes and obesity^36-38, but they face limitations owing to their size, battery requirement and communication signal obstruction caused by the surrounding tissues. Exhaled breath, rich in volatile organic compounds, inorganic molecules, cytokines and pathogens, in the form of gases, aerosols or droplets, provides distinct signatures for respiratory diseases and metabolic processes, such as asthma, chronic obstructive pulmonary disease, lung cancer and tuberculosis³⁹. Breath sensors in respiratory masks track these markers^40,41 and could be integrated with a CLS drug-delivery module. For cerebrospinal fluid, which is important for tracking meningitis, multiple sclerosis and other neurodegenerative disorders⁴², and is typically accessed by lumbar puncture, implantable CLSs remain undeveloped.

Current and future sensing technologies for CLSs

Driven by the need for tight glycaemic control in managing diabetes, glucose–insulin CLSs, or ‘artificial pancreas’, exemplify the ‘sense-to-act’ concept foundational to future CLSs⁴³. Glucose biosensors in these systems enable precise, tailored insulin dosing. Commercial CGMs (Fig. 1b) use enzyme-functionalized needles for direct ISF sensing with an operational lifespan of up to 14 days⁴⁴ and are coupled to commercial insulin pumps.

Current CLS technologies for anaesthetic management rely on biophysical sensors to monitor physiological parameters, such as brain activity, heart rate and blood pressure. These measurements provide indirect data on the effects of anaesthetics like propofol and opioids, enabling clinicians to manage pain, prevent overdosing and control blood pressure fluctuations during surgery. Emerging technologies have shifted to direct measurement of propofol and opioid concentrations, offering real-time monitoring of anaesthetic levels⁴⁵, potentially enhancing the accuracy and responsiveness of CLSs.

Various flexible, skin-worn amperometric and potentiometric electrochemical sensors have been developed for the continuous monitoring of sweat metabolites and electrolytes^11,27. However, commercial sweat-sensing wearable patches have been successful only in collecting hydration (ions and sweat rate) information without CLS capabilities⁴⁶. Attempts by companies like Google, Novartis and Noviosense to develop smart contact lenses for CGM in tears also lack CLS capabilities.

Mimicking the function of a healthy pancreas for optimal glycaemic control requires monitoring analytes beyond glucose. Hence, sensor development is headed towards multiplexed biomarker detection. Multi-analyte sensors in microneedle array patches that enable simultaneous monitoring of diverse molecular markers²⁴ will reinforce intelligent decision-making in CLSs⁴⁷. Parallel innovations include diabetic retinopathy therapy (glucose monitoring and genistein delivery) with contact lenses in rabbits⁴⁸ and ingestible capsules for tracking core body temperature, heart rate, pH, haemorrhage and medication status³⁶. Hybrid multi-modal platforms for simultaneous monitoring of biophysical and biochemical markers promise an even brighter future for CLSs and therapeutic interventions^49-51, capturing comprehensive multi-source data to improve CLS decision-making capabilities. For example, combining blood pressure measurements with biophysical sensing of motor and non-motor symptoms and biochemical sensing of levodopa (l-DOPA) in ISF and sweat guide Parkinson’s disease management^13,52. Similarly, hybrid sensor systems combining blood pressure and electrochemical sensors coupled to drug delivery modules are useful for real-time blood pressure regulation in hypertension⁵¹ (Fig. 1b). Progress in hybrid sensing modules has been driven by coupling different transducers onto single, miniaturized skin-interfacing form factors⁵¹, and machine learning algorithms integrating and mining these rich, multi-source data are expected to optimize therapeutic interventions and drug dosing in CLSs^53,54.

Drug delivery in closed-loop systems

Importance of controlled drug delivery

The actuator of a CLS is the drug delivery unit, which is designed to modulate drug dosages in real time, driven by input from biosensors that continuously monitor physiological changes. Advancements in DDSs are moving towards technologies that enable controlled and more precise drug release, offering superior management of plasma concentration compared with traditional methods such as injections or pills (Fig. 2a). Sustained release systems, the most common in implantable drug delivery technologies, maintain therapeutic levels^55-58 and reduce the need for frequent dosing. However, they lack the adaptability required for CLSs, in which drug release needs to respond dynamically to physiological changes. By contrast, tuneable release systems adjust drug release rates in real time based on sensor feedback, making them ideal for CLSs. These systems harness stimuli-responsive materials to respond to external stimuli such as heat⁵⁹, light⁶⁰, electromagnetic field⁶¹ or ultrasound⁶².

Fig. 2 ∣. Drug delivery profiles and strategies for closed-loop systems.

a, The affinity of a drug-delivery system (DDS) to a closed-loop architecture depends on the drug-release profiles that it can provide. The conventional drug release profile shows peaks and troughs in drug concentration, leading to suboptimal therapeutic windows. The sustained drug release profile maintains a constant drug concentration over an extended period, ensuring prolonged therapeutic action. The controlled release profile demonstrates dynamic adjustments in drug concentration based on real-time feedback, which is ideal for closed-loop systems (CLSs). b, Various implantable and wearable controlled drug delivery strategies and tools enable temporally tuned release of therapeutics. c, Various approaches of DDSs for closed-loop applications include a commercial wearable insulin pump from the Minimed 780 G system, a hybrid solution featuring an implantable DDS powered by an external source and coupled with a wearable biosensor⁹³, a hollow microneedle-based minipatch equipped with a biosensor and an electroosmotic pump⁷⁴ facilitating drug delivery through skin-mounted technology, and another fully wearable device comprising two microneedle modules, using reverse-iontophoresis for biosensing and iontophoresis for drug delivery²³. DC, direct current; EEG, electroencephalography; ICP, ionic concentration polarization; MEMS, microelectromechanical systems. Part c, implantable DDS and wearable sensor, adapted with permission from ref. 93, Science. Part c, hollow microneedle-based minipatch, reprinted (adapted) with permission from ref. 74, American Chemical Society. Part c, microneedle sensor and DDS, adapted from ref. 23, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

Wearable DDSs

Wearable DDSs empower patients with chronic conditions to manage their treatments discreetly, improving quality of life (Fig. 2b). Insulin pumps are the most notable example, having been on the market for decades and integrated with CGMs in ‘hybrid’ CLSs, such as the MiniMed (Medtronics), Omnipod (Insulet Corporation), t:Slim X2 (Tandem Diabetes Care) and Accu-Check Insight (Roche). These devices administer insulin through a cannula inserted into the skin, using mechanical solutions to regulate drug dosage. Although superior to oral or injectable insulin methods in dynamic glycaemic control⁶³, and with fewer gastrointestinal side effects⁶⁴, mechanical pumps cannot fully replicate pancreatic islet function, particularly in maintaining stable euglycaemia. Clinical reports often show wider fluctuations in blood glucose levels for patients using these systems compared with healthy individuals⁶⁵. Moreover, device bulkiness, aesthetic concerns and infection risk of transcutaneous cannulas or catheters, which might reduce patient compliance, have spurred research into more compact, user-friendly DDSs. For example, iontophoresis patches use low voltage (<10 V) and constant current (<1.0 mA cm⁻²) to drive charged small molecules⁶⁶ through the stratum corneum⁶⁷ via electrophoresis and electroosmosis, the latter being the dominant transport mechanism for neutral drugs⁶⁸. The drug release rate (μg h⁻¹) scales with the electric charge transfers at the electrode surface, enabling precise dose control by integrating Coulomb counters into the circuit⁶⁹. Limitations include poor skin tolerance for some drugs, variability of iontophoresis depending on individual skin characteristics⁶⁹ and slow transport across the stratum corneum (minutes to hours). Slow transport can be addressed by combining iontophoresis with other methods that increase skin permeability. Electroporation, for example, applies high-voltage pulses (>100 V) to create temporary pores in the skin’s lipid bilayers and cell membranes⁷⁰, allowing large drug molecules to migrate by electrophoresis and diffuse across the skin⁷¹. However, electroporation requires high-voltage DC–DC converters with microcontrollers for pulse generation⁶⁹, presenting technical challenges.

Microneedles provide a minimally invasive alternative, creating microchannels in the skin to deliver drugs. They are classified into dissolving and hollow types. Dissolving microneedles, made of hydrogels, permit controlled release through stimuli-responsive materials, such as polypyrrole (electrically controlled)⁷² and 1-tetradecanol (thermally controlled)⁶⁴. Other designs include microneedles that perforate and retract from the skin, using iontophoresis to enhance drug penetration⁷³, and hollow microneedles integrated with electroosmotic pumps or ultrasonic atomizers for controlled release of drugs in liquid formulation^22,74,75 (Fig. 2c). Smart contact lens-based drug delivery systems have progressed considerably in treating ocular diseases and visual impairment conditions, driven by innovations in lens responsive materials and fabrication strategies^12,76,77. Such systems can be developed with hydrogels, polymers and biocompatible materials (like chitosan or agarose) using soft lithography, electrospinning and printing (direct and inkjet). Drugs can be loaded directly into the contact lens material through mixing and molecular imprinting and can be released using varying stimuli (like pressure, heat, electric field, magnetic field or pH change).

Wearable DDSs face challenges such as variability in skin permeability, low delivery rates and limited drug payloads (Table 2). Intra-patient and inter-patient variability, regional blood flow and skin condition and pH further influence iontophoretic systems⁷⁸. Measuring skin permeability before device application and adjusting stimulation parameters (such as current and voltage) could improve dosing reliability by accounting for individual differences. Drug payload limitations in dissolving microneedles^79,80 could be mitigated by using higher-potency drugs to maximize therapeutic efficacy.

Table 2 ∣.

Metrics related to drug delivery strategies applicable to closed-loop systems

Approach	FDA-approved, commercially available
	Drug delivery strategy	Advantages	Challenges	Reference
Wearable	Mechanical insulin pump	Rechargeable and refillable Reliable tuneable drug delivery (basal ~0–35 U h−1, bolus ~0.05–35 U), easy to integrate with control systems Proven in closed-loop systems	Painful cannula insertion and frequent cannula replacement Risk of mechanical failure or infusion set blockage Bulky, unattractive to wear	MiniMed 780G System User Guide t:slim X2 Insulin Pump with Control-IQ Technology User Guide
Wearable	Insulin patch	Same as mechanical insulin pump but tubeless and more discrete to wear	Limited drug loading capacity	Omnipod 5 Automated Insulin Delivery System User Guide Medtrum TouchCare Nano Pump
Implantable	Intrathecal pump	Same as mechanical insulin pump but tubeless and more discrete to wear Refillable Proven for chronic pain management and spasticity control	Not rechargeable Not proven in closed-loop systems Pump and catheter mechanical issues	SynchroMed III Intrathecal Pump for Chronic and Cancer Pain PROMETRA II PROGRAMMABLE PUMP For use with Intrathecal Catheter SynchroMed III Intrathecal Pump for Severe Spasticity
Approach	In development
	Drug delivery strategy	Advantages	Challenges	Refs.
Implantable	Ionic concentration polarization Electrophoresis Electrostatic gating	Operating voltage (<3 V) Power consumption (μW) No moving parts	Molecule size-dependent and charge-dependent Long-term degradation of electrodes	95-98
Implantable	Electrochemical	Operating voltage (<10 V) Minimal heat generation (<0.2 °C)	Power consumption (mW–W) Long-term degradation of electrodes	90,92,93
Implantable	Electromagnetic	Large pumping force	Power consumption (mW–W) Potential interference with external devices Requires magnetic field application Heat generation	82,94
Implantable	Piezoelectric	Large pumping force No moving parts	Requires high voltage (tens to hundreds of volts)	83,84,86,89
Wearable	Microneedles Dissolving Hollow	Minimally invasive and painless Release of small and large molecules	Basal delivery due to passive diffusion (dissolving) Limited drug loading capacity (dissolving) Need for an integrated pumping system (hollow) Variability in skin permeability	22,59,64,73,195-197
Wearable	Iontophoresis	Operating voltage (<10 V)	Slow release due to the stratum corneum Variability in skin permeability Release of small molecules (up to a few kDa) Limited current/voltage window for safe operation	59,64,69,73,128,198
Wearable	Electroporation	Release of small and large molecules	Variability in skin permeability Operating voltage (>100 V)	69,71

Implantable DDSs

Implantable devices enable tuneable, precise drug delivery through advanced technologies like microelectromechanical systems actuators (Fig. 2b). These devices enable fine control of drug release rates and device miniaturization⁸¹. In most pump-microelectromechanical systems actuators, a rapidly oscillating diaphragm connected to a drug reservoir generates pressure to eject the drug through unidirectional exit valves^81-83. The pump’s flow rate can be adjusted by modifying the applied electrical signal⁸⁴, offering greater precision than plunger pumps used in insulin delivery⁸⁵. Piezoelectric microelectromechanical systems actuators, driven by alternating high-voltage currents, use piezoelectric ceramics to oscillate the diaphragm⁸⁶, providing strong pumping forces and consistent flow rates⁸⁷. However, the high voltage required (tens to hundreds of volts)^88,89 complicates power unit design. Electrochemical actuators are an attractive alternative for their low operating voltage (<10 V) and power consumption, minimal heat generation and pumping force^90,91. They use an applied potential to electrolyse water in the pumping chamber, producing oxygen and hydrogen gases that increase chamber pressure to push the drug out⁹². These actuators are particularly suited for rapid drug elution (such as in seizure-triggered delivery^91,93) and have been tested for local drug delivery in the brain⁹⁰ and peripheral nerves⁹² within optofluidic systems for wireless neuromodulation. Electromagnetic pumps share many strengths of electrochemical pumps but generate more heat⁶⁹ and require high electrical currents and magnetic field strength, potentially influencing drug efficacy⁸⁸. These pumps use a current-generated electromagnetic field to oscillate a magnet attached to the diaphragm within the pump⁸¹. For example, a radiofrequency electromagnetic field has been used to induce a ‘flapping’ motion in a nitinol cantilever actuator, squeezing the pump chamber to eject a drug solution⁸². An alternative design includes two permanent magnets affixed to the pump membrane, repeatedly reversing polarity to induce drug flow, providing reversal capabilities to prevent drug over-release^93,94.

Other cutting-edge approaches harness the interactions between ions of electrolyte solutions, drug molecules and nanochannel interfaces within nanofluidic platforms, enabling control over molecular transport by leveraging electrokinetic and electrostatic phenomena. Ionic concentration polarization is used to create zones of ion accumulation and depletion at the inlets and outlets of nanochannels upon the application of an electric field. This effect is most pronounced in narrow channels with overlapping electrical double layers, where ion accumulation and depletion are enhanced. This process reduces or interrupts the transport of drugs across nanochannels, regardless of the polarity of the applied potential^95,96. By contrast, electrophoresis directly drives charged drug molecules through the nanochannels under the influence of the electric field. Ionic concentration polarization modifies the local ionic environment whereas electrophoresis provides direct control over the direction and rate of molecular movement^95-97. However, it typically requires higher-intensity electric fields to sustain transport, which increases power consumption. Another approach is electrostatic gating, which involves modulating the Debye length — the distance from the surface over which ions shield electric fields — by applying a voltage to embedded electrodes. In nanochannels with dimensions comparable with the Debye length, electrostatic gating influences the ionic distribution on the nanochannels surface, enabling controlled drug diffusion^98-101. However, in high ionic-strength solutions, such as ISF, the Debye length is shortened, complicating precise modulation with low potentials in in vivo settings. These electrokinetic and electrostatic approaches offer several advantages, including low power consumption and the elimination of mechanical parts, reducing the risk of failure and making them ideal for implantable devices. Nonetheless, their effectiveness is size-dependent and charge-dependent, which might limit their applicability, especially because many drugs are neutral in charge.

Control algorithm in closed-loop systems

The control algorithm processes the sensor inputs to calculate, in a feedback loop, the optimal drug dosages needed to maintain physiological parameters within desired ranges. Proportional-integral-derivative controllers, a class of feedback control systems that continuously adjust output based on past, present and predicted errors are widely used for their reliability and ease of implementation^102,103. However, they may struggle with addressing time delays associated with most sensors (related to the transport time of the target analyte from blood to other biofluids for electrochemical sensing) and DDSs (owing to the route of administration and pharmacokinetics of the drug). Model predictive control has emerged as a preferred alternative¹⁰⁴, leveraging personalized models to adapt to individual pharmacokinetics in real time, thus minimizing latency issues¹⁰⁵. Additional control strategies, such as sliding mode control, fuzzy logic control¹⁰⁶, artificial neural networks¹⁰⁷ and reinforcement learning¹⁰⁸ offer diverse approaches suitable for applications ranging from anaesthesia^109,110 to neurostimulation^111,112. For the management of type 1 diabetes (T1D), artificial neural network models have been developed using clinical datasets to identify patient dynamics and regulate blood glucose levels, establishing personalized insulin–glucose relationships¹⁰⁷.

Control algorithms must undergo rigorous testing and validation under settings that mirror daily patient use. In silico simulations allow developers to test control algorithms under various conditions, such as sensor errors, pump occlusions, physical activity and meal responses, to assess system robustness and optimize performance. For example, FDA-approved simulators based on glucose data during meals in individuals with T1D offer a platform for modelling the CLS behaviour across different age groups, aiding in the refinement and validation of these systems prior to clinical deployment¹¹³. This simulation-driven development accelerates the pathway to clinical adoption and enhances the reliability and efficacy of CLSs in real-world scenarios.

The role of AI

Artificial intelligence (AI) has revolutionized the field of biosensors and drug-delivery CLSs by enabling intelligent, adaptive and personalized health-care solutions. In biosensors, machine learning algorithms enhance the accuracy and sensitivity of analyte detection by filtering noise, correcting signal artefacts and identifying complex patterns within biochemical data¹¹⁴. For example, gradient boosting and regression models analyse feature importance through metrics like gain or coefficients, indicating the influence of each variable on the outcome. Reinforcement learning models derive feature importance indirectly by learning which states or actions maximize cumulative rewards, whereas support vector classifiers use support vectors to identify boundary-defining features, which divides the data into different classes. Fuzzy logic systems prioritize features based on their membership functions and rules. In general, regression methods rely on mathematical equations to capture these relationships, gradient boosting models iteratively refine predictions using residual patterns, and support vector classifiers construct hyperplanes to separate data into meaningful classes (Table 3).

Table 3 ∣.

Different machine learning algorithms applicable to closed-loop systems

Machine learning approach	Machine learning algorithm	Application in closed-loop systems	Ref.
Gradient boosting	XGBoost	Predicts patient-specific dosages accurately Can detect anomalies in trends	16
Regression	Artificial neural networks	Linear and non-linear therapy adaptation by predicting physiological responses to maintain homeostasis Useful in complex biological systems with non-linear dynamics	199
Fuzzy logic	N/A	Operates on degrees of truth, enabling more nuanced decision-making in complex systems Enhances adaptability by accommodating variations in patient responses and environmental conditions Enables incorporation of human reasoning to improve system robustness Enhances drug delivery precision and patient safety	164
Reinforcement learning	Control Lyapunov function	Adaptive control strategies to ensure stability while optimizing patient outcomes Reinforcement learning algorithms learn optimal treatment policies from patient data, whereas Lyapunov functions ensure stability Real-time therapeutic adjustments by accounting for real-time individual patient responses	200
Support vector classifier	N/A	Accurately classifies complex patterns in medical data Predicts patient’s dynamic response to treatments and interventions in real time	201

Various machine learning models have been developed to determine the best approach for fusing physiochemical data with psychological phenomena such as stress and anxiety. Extreme gradient boosting (XGBoost), linear and radial basis function support vector machines, logistic and ridge regression, and conventional decision trees have been evaluated for their performance. The XGBoost machine learning model yields higher accuracy (versus classification and regression models) in assessing stress and anxiety using an abundance of physiochemical data as the input sequence¹⁶. Moreover, XGBoost is a scalable machine learning algorithm that uses gradient boosting to build decision trees sequentially, correcting errors from previous iterations. Its high speed, accuracy and ability to handle large datasets and missing values make it highly suitable for real-world applications¹¹⁵. Medtronic’s MD-Logic Artificial Pancreas System (utilized by the MiniMed 780 System) uses fuzzy logic control to automate diabetes management, reflecting the vision of integrating AI in closed-loop devices¹⁰². Additionally, FDA-approved neurostimulation devices for epilepsy and chronic pain use machine learning models such as artificial neural networks to adjust therapy in real time based on the patient’s movements and neural activity patterns. Recent advancements have enabled systems to automatically deliver naloxone (an opioid antagonist) by analysing biophysical data, such as electrocardiography, heart rate, respiratory rate and oxygen saturation, to detect overdose conditions^116,117. Advanced machine learning algorithms, such as deep learning and reinforcement learning, further refine these systems, learning from vast datasets of patient responses to better tailor drug dosages in real time. Predictive models can also be used to analyse individual risk factors and environmental variables, enhancing prevention. Machine learning can even account for the interpersonal differences (like existing health conditions, genetic factors and lifestyle) by generating personalized parameters to enhance the accuracy of sensor reading and drug delivery rate. Therefore, the convergence of machine learning with biosensing and drug delivery technologies represents a paradigm shift towards personalized and responsive health care, offering new approaches for managing complex diseases.

Smart closed-loop systems

The advancement of glucose sensors and insulin pumps has paved the way for the creation of modern ‘artificial pancreas’ (Fig. 3a). These CLSs and their ‘hybrid’ variants that integrate manual control — like meal announcements for insulin boluses — have demonstrated substantial clinical benefits. For instance, ‘hybrid’ closed-loop insulin delivery systems are effective in enhancing glucose control by increasing time-in-range and reducing the risk of hypoglycaemia in individuals with suboptimally managed T1D. Such ‘hybrid’ systems are superior to conventional treatment such as multiple daily insulin injections⁶³, continuous insulin pump infusion¹¹⁸ or even traditional CLSs^119-121. In anaesthesia, CLSs have provided better control over anaesthetic depth, facilitating shorter recovery times and reducing the incidence of inadvertent awareness^122,123.

Fig. 3 ∣. Advances in closed-loop systems, working principle and design consideration.

a, Timeline of milestones in the integration of sensors and drug delivery systems in closed-loop systems (CLSs) for diabetes. In 1962, the pioneering work of Clark and Lyons²⁰⁴ introduced the concept of enzyme electrodes. This technology was later commercialized by YSI, leading to the launch of the YSI Model 23 Analyzer in 1975, which marked an important step in blood-glucose analysis²⁰⁵. Advancements in insulin delivery systems in 1978 saw the introduction of commercial insulin pumps²⁰⁶. Subsequent developments in 1987 introduced glucose test strips, enabling self-testing through mediator-based enzyme electrodes²⁰⁷. By 1991, subcutaneous glucose monitoring technologies emerged, enabling real-time monitoring of glucose levels²⁰⁸. The integration of continuous glucose monitor (CGM) systems began in 1999 with the introduction of a CGM system for physician use²⁰⁹. Remotely programmable insulin pumps emerged in the same year. In 2003, the first wireless insulin pump system was approved. This device could automatically receive glucose readings wirelessly from a blood glucose-meter and use an integrated calculator to suggest appropriate insulin doses. Early prototypes of ‘artificial pancreas’ systems and ‘hybrid’ CLS clinical trials started in 2006. In 2016, these efforts culminated in the FDA approval of the Medtronic MiniMed 670 G, the first commercial hybrid closed-loop insulin delivery system that combines CGM technology with automated insulin delivery (Medtronic Innovation Milestones). b, Schematic representation of a smart CLS working principle. Unlike traditional drug delivery systems, these systems leverage machine learning and artificial intelligence to predict patient needs before symptoms manifest, and make decisions based on complex, continuously updated data patterns. Rather than simply responding to immediate data, these systems analyse trends and adjust treatment protocols dynamically, enabling a high degree of personalization. Collected data are transmitted to physicians for remote monitoring and personalized treatment adjustments. c, Design considerations for each component of a CLS. The successful implementation of effective and reliable CLSs is driven by the performance and user-friendliness of sensors and drug delivery systems, the reliability of control algorithms and communication systems, and the longevity and efficiency of the power unit.

Clinically approved CLSs

Clinically approved CLSs for T1D, such as the MiniMed 780 G (Medtronic), t:slim X2 (Tandem Diabetes Care), Omnipod 5 (Insulet Corporation) and TouchCare Nano (Medtrum), exemplify the advancements in this field. The MiniMed 780 G and t:slim X2 integrate insulin pumps with CGMs using proprietary closed-loop algorithms (SmartGuard and Control-IQ+) to adjust insulin delivery based on CGM readings. However, both systems have cumbersome tubing systems that can be unattractive to users. By contrast, the Omnipod 5 and TouchCare Nano patch pumps offer tubeless insulin pumps that are smaller and more discreet, providing more freedom and flexibility. Current state-of-the-art CLSs for diabetes use CGMs that have a mean absolute relative difference (error) of around 8% versus blood-glucose levels measured using gold-standard techniques (commercial blood glucose meters, YSI glucose analysers (YSI Life Sciences - The Gold Standard)). The drug infusion pumps in these systems have precise low flow rates (>μl min⁻¹) and adaptive and model predictive control algorithms. They are powered by rechargeable batteries and support wireless communication via near-field communication or Bluetooth for data transfer and management (Fig. 3b). Although these systems enhance diabetes management and quality of life, they come with several drawbacks: the inconvenience of fingerstick tests for calibration, uncomfortable tubing, painful cannula insertions and the need for cannula replacements every 2–3 days to prevent occlusion. Users also face pod failures, adhesive issues and high costs, with annual expenses estimated at US$6,000¹²⁴.

Applications in development and new opportunities

CLSs integrating sensing and drug delivery in miniaturized, discrete, stable and cost-effective structures could enhance acceptability. An example of such a system is a CLS that features two microneedle devices designed to deliver insulin via iontophoresis and sense glucose from ISF using reverse iontophoresis²³. Despite its promise, this system lacks a mechanism to refresh the sensor solution after each measurement, limiting its autonomy towards a fully closed-loop architecture. Alternatively, a patch using an electrochemical device with sweat-based sensors and thermoresponsive microneedles for drug delivery has been investigated for diabetes management^59,64, with challenges related to environmental variability when it comes to thermal control. In this system, the microneedles continuously release low basal levels of drugs as they dissolve, which could be undesirable for certain therapeutic indications. Hollow microneedles have been functionalized for glucose sensing and coupled with an electroosmotic pump for insulin delivery within a flexible device, showing promise for integration⁷⁴. However, challenges are anticipated in translating this technology for clinical use due to the substantial differences in skin characteristics between humans and rats (such as in epidermal thickness, hair density, hydration levels, stretchability and blood flow). Another innovative approach uses heaters and a thermally expandable polymeric actuator combined with a hollow microneedle as a drug delivery unit of a CLS, demonstrated using only humidity and temperature sensors as proof of concept¹²⁵.

For wound healing, sensors that monitor critical parameters at the injury site, integrated with systems that deliver antibiotics, anti-inflammatory agents or growth factors, could enable a real-time, responsive treatment to promote faster recovery and prevent complications^126-129. Some recent developments in this area are advancing towards CLSs^{126,128,130-133} (Fig. 1b); these include devices that simultaneously monitor biophysical markers, metabolites and inflammatory biomarkers, and conduct wound healing through antimicrobial delivery or electric stimulation. However, the bulkiness of the sensor system (components and battery) and the need for retrieval after healing, pose major challenges. This motivates a shift to bioresorbable and battery-free systems¹³⁴ and heat-induced adhesion for selective patch attachment–detachment at the wound site¹²⁶.

Despite this progress, integrating continuous biochemical sensors for analytes beyond glucose remains a challenge for expanding CLSs into other therapeutic areas. Therefore, efforts have also focused on improving CLS decision-making based on electrophysiological monitoring rather than just continuous biochemical sensing⁴. Notably, an implantable closed-loop solution (MiNDS) has been explored in neural applications for local intracerebral drug delivery using refillable pumps connected to a miniaturized neural DDS through fluidic channels¹³⁵. MiNDS integrates a neural electroencephalogram activity-recording electrode, enabling real-time analysis to detect changes in local field potential activity due to drug delivery. This application holds promise for treating neurological disorders requiring chronic monitoring and real-time intervention, such as epilepsy and Parkinson’s disease, in which it is beneficial to monitor epileptic discharges and beta-band local field potential oscillations, respectively.

Areas of opportunities for CLS development include disease conditions that: (1) exhibit distinct signals for prompt detection and timely drug delivery, (2) are highly dosage-sensitive, (3) have medications that are stable within implantable or wearable delivery devices, and (4) can benefit from extensive patient data to enhance CLS decision-making capabilities using AI. For example, neurological disorders like Parkinson’s disease would benefit from closed-loop delivery of L-DOPA, as overdosing can cause dyskinesias (involuntary movements), whereas underdosing results in worsening motor symptoms. Fingertip sweat⁵² and blood¹³⁶ can be used for decentralized l-DOPA monitoring, whereas motor and non-motor symptoms are tracked using physical sensors (accelerometers, gyroscopes and magnetometers), blood pressure and machine learning processing¹³. Similarly, hypertension represents another opportunity for CLS development; wearable blood-pressure monitors with epidermal or implantable controlled drug delivery of calcium-channel blockers¹⁰¹ could help to prevent complications like hypotension or uncontrolled hypertension, which can lead to stroke or heart attack. CLSs have already shown promise in the detection and emergency treatment of opioid overdose^116,117, in which sensors capable of detecting respiratory depression (such as photoplethysmography sensors, electrocardiography sensors, and inertial measurement units) can trigger an actuator (such as an electrolytic pump or a DC motor-driven mechanism) for the immediate release of naloxone to reverse the effects.

Emerging biohybrid CLS concepts

Bacteria and cells can be engineered to detect disease markers and autonomously synthesize and secrete therapeutics on demand^137-144 in response to chemical¹³⁷, optical¹⁴⁵ or electrical¹⁴⁶ stimuli. This type of biohybrid CLS reduces the need for sensor replacements and drug refills. For instance, cells were bioengineered to secrete interleukin-1 receptor antagonist (an anti-inflammatory molecule) in response to inflammation in the murine K/BxN model of inflammatory arthritis¹³⁷. Similarly, electrosensitive β-cells have shown real-time glucose control by releasing insulin upon electrical stimulation in diabetic mice¹⁴⁶. In another study, cells were genetically engineered with an optogenetic circuit, enabling precise control of insulin secretion in response to far-red light, enabling real-time regulation through a smartphone interface¹⁴⁵. However, maintaining cell viability and function long-term in vivo remains a challenge¹⁴⁴.

Challenges

Researchers and manufacturers of CLSs encounter numerous design considerations due to the complexities associated with different components interplaying in these systems (Fig. 3c), which inevitably give rise to various challenges.

Power

Power generation, storage and consumption are critical to CLS functionality and influence patient acceptance. Key power demands in a CLS come from the biosensor, DDS, control algorithm and wireless communication. Although miniaturized lithium-ion and lithium-polymer batteries are commonly used for their high energy density, safety and form factor limitations call for sustainable, autonomous energy sources. Integrating energy-harvesting modules with flexible batteries or stretchable supercapacitors presents a promising approach. Energy harvesters, the system’s most challenging component, can convert chemical, solar, thermal or kinetic energy into electrical power. For example, biofuel cells using biochemical reactions in sweat with electrolyte-based supercapacitors^147-149 and sweat-activated batteries^150,151 are garnering interest in wearable technology. However, biofuel cells have a limited lifespan due to enzymatic degradation. More durable options include triboelectric nanogenerators and piezoelectric generators, which convert mechanical energy from body motions^152-154, and photovoltaics, which harness light and offer the highest energy density¹⁵⁵. Physical motion and light are not always available, and thermoelectric generators, which leverage localized skin-based thermal gradients¹⁵⁶, can provide continuous power under a wider range of conditions.

Biofuel cells, thermoelectric generators and photovoltaics produce DC, whereas triboelectric nanogenerators and piezoelectric generators produce alternating current, requiring rectifiers to convert alternating current to DC for energy storage in batteries or supercapacitors. Near-field communication enables wireless power transmission to the CLS when an external transmitter and the system’s receiver are in proximity¹⁵⁷. These technologies can collectively provide sustainable power in the microwatt to milliwatt range¹⁵⁸, which is sufficient for the closed-loop architecture. Therefore, integrating multiple energy-harvesting strategies can optimize overall power yield. Despite strides in this field¹⁵⁹, capturing biochemical or mechanical energy of the human body and designing efficient circuit management modules remains a challenge. Consequently, the field relies primarily on wireless rechargeable batteries for powering wearable and implantable devices¹⁶⁰.

Communication

CLSs typically have separate platforms for sensing, drug delivery and control systems. Technological challenges prevent the integration of these components into a single unit, yet interconnectivity remains essential. Even so, a fully autonomous integrated platform would still require communication on two levels: (1) with users to manage emergency medication discontinuation or other anomalies, and (2) with central medical databases for data transfer and analysis. Wireless communication is ideal for CLS connectivity, offering usability and biocompatibility. Key considerations for communication protocols include power consumption, data volume, bandwidth and device circuit compatibility. Main strategies for wireless communication include standardized protocols such as Bluetooth Low Energy, passive radio-frequency identification and near-field communication. Bluetooth Low Energy and radio-frequency identification have power consumption in the order of microwatts to milliwatts, suitable for connecting the separate components of CLSs. Further optimization can adjust sensor sampling, actuation frequency and data transmission intervals based on patient needs¹¹.

Wireless connectivity in CLSs can be disrupted by factors like distance or interference. Mitigation strategies include real-time alerts, automatic reconnections and predictive algorithms to maintain drug delivery based on recent physiological data trends during sensor disconnections. However, predictive algorithms can introduce errors, necessitating failsafe measures such as pre-set basal rates and manual input of physiological data to the drug delivery module, measured using external sensors. These approaches could temporally ensure safe operation. Additionally, local buffering of sensor data could facilitate a seamless return to normal operation once reconnection is restored. Another critical risk is unauthorized access to sensitive data or malicious manipulation of therapies. Cybersecurity vulnerabilities in current designs (such as firmware modification and data breaches) highlight the need for robust mitigation strategies, including robust encryption, authentication, access controls and firmware modification prevention schemes¹⁶¹.

Future CLS generations could integrate all components into a single unit with shared circuitry to eliminate wireless connectivity issues and reduce power consumption. Near-field communication could enable battery-free operation, allowing an external device to control the system, retrieve data and interface with medical databases.

Foreign-body response

The natural foreign-body response to an implanted device involving inflammation and progressive fibrotic encapsulation can hinder drug diffusion from the implants^162,163 and analyte transport from the biofluid to the sensor’s transducer. It can also reduce the sensor’s mechanical response and shield electrical potentials, compromising sensitivity. Altogether, these factors could impact the function and longevity of implant-based or microneedle-based CLSs.

Foreign-body response can be mitigated by coating implants with biologically inert materials¹⁶⁴. For example, minimally invasive microneedle sensors often use anti-biofouling coatings¹⁶⁵. In the Eversense CGM implant, all electrical and optical components are enclosed within a biocompatible capsule⁴. However, most implants cannot be fully encapsulated with inert materials without compromising their function. Alternative strategies include dynamic implant actuation for local immunomodulation of the foreign-body response¹⁶⁶ and limiting surface roughness (1–4 μm) to minimize cell adhesion¹⁶⁷. Neutral surface charges could prevent protein adsorption^168-170, and hydrophilic properties with chemically stable functional groups could accelerate tissue repair by encouraging anti-inflammatory macrophage polarization (M2 phenotype)^171,172. Moreover, if implant stiffness cannot match the mechanical properties of surrounding tissues, the implant’s outer layer can be modified for mechanical compatibility. For microneedles, using extremely soft materials with elastic modulus similar to biological tissues shows promise^173,174. Implant design should also avoid sharp discontinuities and angles and minimize thickness and curvature. The implantation site also influences foreign-body response and power/data transmission efficiency. Subcutaneous placement, with milder foreign-body response, is ideal for drug diffusion and transport¹⁷².

Regulatory

Regulatory compliance has become a considerable challenge for medical device manufacturers, particularly due to the European Union’s Medical Device Regulation^175,176, which imposes rigorous standards, especially for software used in disease diagnosis, prevention, monitoring, prediction and treatment. These rules place greater emphasis on AI-driven software and expand the definition of a ‘medical device’ beyond physical instruments and implants.

In the USA, regulatory pathways include the FDA’s 510(k) clearance, de novo classification and premarket approval. Although the 510(k) process is less burdensome, demonstrating substantial equivalence to market CLSs is challenging due to the field’s novelty, with most predicates limited to diabetes applications. Without a suitable predicate, manufacturers must pursue the more demanding de novo classification or premarket approval process. Market entry is further complicated by the need to secure reimbursement from health-care systems or insurers, which is often contingent on demonstrating cost-effectiveness and clinical benefits through post-market studies. To address these challenges, manufacturers can incorporate clinically approved technologies into CLSs, leveraging established safety and efficacy data to streamline regulatory approval. Early collaboration with regulatory bodies is another key strategy, helping developers anticipate hurdles and facilitate translation of CLS innovations into clinical practice.

Other challenges

Manufacturing techniques, such as 3D printing, inkjet printing and lithography, have improved the performance and scalability of sensors and DDSs. However, practical considerations of compatibility, effectiveness and comfort remain crucial for the widespread adoption of CLSs. For implantable DDSs, reservoir-based designs are ideal for drug storage, but their capacity is constrained by physical size. Therefore, safe and efficient transcutaneous drug refilling is a practical necessity to minimize surgical replacement. Drug stability is another consideration, as hydrolysis, decomposition and unintended reactions with ISF can degrade drug efficacy over time. Refilling with solid rather than liquid drug formulations offers up to 1,000-fold higher drug loading efficiency and superior stability⁵⁵, because hydrolysis and decomposition occur primarily at the surface in contact with ISF¹⁷⁷. Infection risks are more prominent in systems with transcutaneous components, such as drug refilling ports, catheters or cannulas. Mitigation strategies include using sterile techniques and prophylactic antibiotics to reduce bacterial inoculum, repulsive polymer coatings to prevent bacterial adhesion, and antimicrobial-eluting or contact-killing surfaces to inhibit bacterial growth^178,179. Furthermore, as bioactive materials can degrade or lose function when subjected to standard sterilization processes^180,181, selecting materials that withstand these processes without compromising their bioactivity is essential for ensuring device reliability.

Deploying CLSs in low-resource environments presents unique challenges that require thoughtful design. Reducing device complexity and incorporating intuitive, user-friendly interfaces can empower patients to manage treatments with minimal training. Using low-cost sensors with simplified calibration, along with basic telehealth interfaces, could be advantageous. For instance, mobile health (mHealth) platforms using basic connectivity (short message service (SMS)-based data transmission) could enable health-care providers to remotely monitor patients and adjust treatments in real time, even in areas with minimal digital infrastructure. By focusing on engineering strategies that prioritize frugal innovation, CLSs can bridge health-care gaps and bring high-quality, patient-specific treatments to underserved populations worldwide, ensuring that the benefits of precision medicine reach even the most remote and resource-limited areas.

Outlook

The future of CLSs promises a range of developments that will enhance medical practice. Ideally, CLSs should be self-powered, all-in-one systems, with efficient communication and data processing, capable of delivering multiple drugs based on the multi-source sensor data (Fig. 4a). Patient resistance to new technologies could present a barrier to the adoption of CLSs. Educating patients on the benefits of CLSs coupled with clinical trial evidence of their safety and efficacy is essential. User-friendly designs emphasizing comfort and simplicity could further facilitate adoption.

Fig. 4 ∣. Next-generation closed-loop systems.

a, The next phase of closed-loop system (CLS) evolution will focus on integrating self-powered single platforms with optimized drug reservoir efficiency to enhance patient acceptance. Bi-hormonal therapies will further improve treatment efficacy for certain diseases. b, Wearable CLSs of the future will be capable of real-time monitoring of multiple health parameters using electrochemical and biophysical sensors powered by a combination of solar cells, biofuel cells and nanogenerators. A low-power microcontroller unit (MCU) will control the operation, while a digital display will provide information to patients. c, Advanced implantable CLSs will integrate engineered cells that sense and release therapeutics. Biosensors will ensure the viability of these cells, and embedded electronics will enable wireless communication and precise, on-demand cell stimulation for controlled therapeutic release. BFC, biofuel cell; BP, blood pressure; CGM, continuous glucose monitor; ECG, electrocardiography/electrocardiogram; EMG, electromyogram; HR, heart rate; MN, microneedle; PENG, piezoelectric nanogenerator; SpO₂, oxygen saturation; TEG, thermoelectric nanogenerator; TENG, triboelectric nanogenerator.

The choice between wearable and implantable CLSs depends on clinical needs, treatment invasiveness and patient engagement. Implantable CLSs are ideal for extended therapies requiring minimal patient interaction, such as chronic pain management or hormone replacement. Although they offer discretion, their surgical placement presents challenges for maintenance and timely troubleshooting. By contrast, wearable CLSs are generally non-invasive, suitable for short-term to long-term treatments and require active patient involvement. This imparts patient autonomy, enabling them to more closely monitor their health and immediately respond to device alerts.

Wearable CLSs, deployed as headwear, wristbands, textiles or patches, could monitor a range of biochemical health indicators (such as nucleic acids, proteins, ions and metabolites) in sweat and ISF in real time. Microneedles could be functionalized to extract ISF through a microfluidic channel interfaced with electrochemical sensors, and sweat could be extracted using a hydrogel. A wearable CLS should also integrate biophysical and biochemical sensors connected to a low-power microcontroller unit, which could be powered by a battery harvesting energy from solar cells, biofuel cell arrays or nanogenerators, and execute machine learning-based real-time analysis of sensor input data to control the DDS. An advanced DDS would feature an easy-to-refill reservoir and a digital display and audio for sharing information, such as refill time, volume infused and residual drug. The microcontroller unit could transfer data via Bluetooth Low Energy to smart gadgets (phones, watches) and the cloud. Remote monitoring and feedback from health professionals could be facilitated via Wi-Fi, ensuring seamless communication and user support (Fig. 4b).

Biohydrid platforms represent the next-generation implantable CLSs. Materials science, synthetic biology and immunology expertise will be key for creating an immune-protected microenvironment conducive to cell survival and function while preventing overgrowth¹⁴⁴. Safety protocols will be necessary to enable complete device retrieval in cases of malfunction or adverse reactions (Fig. 4c). Furthermore, integrating ‘smart’ molecules that can reversibly toggle their activity on and off in response to specific analyte levels offer potential for adoption in CLS and reduce toxicity risks of excessive drug release^182-184.

The future of CLSs for drug delivery holds great promise for advancing personalized medicine and telemedicine. Key challenges include sensor stability and accuracy, DDS miniaturization, adaptable control algorithms and issues with power, communication and regulatory compliance. Addressing these problems is essential for the widespread adoption of CLSs. As CLSs evolve, their use is likely to expand into remote and challenging environments, such as space missions¹⁸⁵, military operations, industrial settings and underserved areas. Integrating AI could drive precision and full automation. Growing market demand and competition will spur innovations, positioning CLSs as a standard health-care tool globally.

Key points.

• Closed-loop systems (CLSs) face barriers such as sensor stability, miniaturization of drug delivery systems and regulatory concerns, requiring extensive clinical validation for broad patient acceptance and health-care integration.

• The trade-offs between wearable and implantable CLSs involve balancing clinical needs with treatment invasiveness and end-user engagement. Implantables offer long-term solutions whereas wearables are more suitable for short-to-mid-term treatments.

• Artificial intelligence-driven control algorithms improve CLS performance by learning from patient data to predict disease progression and abnormalities, and optimize drug delivery, moving towards full automation with minimal human intervention.

• The integration of engineered cells in implantable CLSs holds promise for autonomous drug delivery, although overcoming immune rejection and maintaining long-term cell viability and function in vivo remain unresolved.

Citation diversity statement

We acknowledge that papers authored by scholars from historically excluded groups are systematically under-cited. Here, we have made every attempt to reference relevant papers in a manner that is equitable in terms of racial, ethnic, gender and geographical representation.