Transforming Healthcare: Intelligent Wearable Sensors Empowered by Smart Materials and Artificial Intelligence

Abstract

Intelligent wearable sensors, empowered by machine learning and innovative smart materials, enable rapid, accurate disease diagnosis, personalized therapy, and continuous health monitoring without disrupting daily life. This integration facilitates a shift from traditional, hospital‐centered healthcare to a more decentralized, patient‐centric model, where wearable sensors can collect real‐time physiological data, provide deep analysis of these data streams, and generate actionable insights for point‐of‐care precise diagnostics and personalized therapy. Despite rapid advancements in smart materials, machine learning, and wearable sensing technologies, there is a lack of comprehensive reviews that systematically examine the intersection of these fields. This review addresses this gap, providing a critical analysis of wearable sensing technologies empowered by smart advanced materials and artificial Intelligence. The state‐of‐the‐art smart materials—including self‐healing, metamaterials, and responsive materials—that enhance sensor functionality are first examined. Advanced machine learning methodologies integrated into wearable devices are discussed, and their role in biomedical applications is highlighted. The combined impact of wearable sensors, empowered by smart materials and machine learning, and their applications in intelligent diagnostics and therapeutics are also examined. Finally, existing challenges, including technical and compliance issues, information security concerns, and regulatory considerations are addressed, and future directions for advancing intelligent healthcare are proposed.

Wearable sensors, empowered by AI and smart materials, revolutionize healthcare by enabling intelligent disease diagnosis, personalized therapy, and seamless health monitoring without disrupting daily life. This review explores cutting‐edge advancements in smart materials and AI‐driven technologies that empower wearable sensors for diagnostics and therapeutics. Current challenges, limitations, and future opportunities in transforming intelligent healthcare are also examined.

1. Introduction

Intelligent healthcare^{[ 1 ]} is a smart healthcare service system that can dynamically access information, connect all elements related to healthcare, and actively, effectively, and intelligently manage and respond to medical needs.^{[ 1} , ^{2 ]} It represents a shift in the medical paradigm, efficiently linking people, institutions, medical devices, information, and supplies.^{[ 2 ]} The development of wearable sensing systems empowered by smart materials and machine learning (ML) enhanced technologies have spurred the advancements of intelligent healthcare. They can cooperate with other technologies such as communication technologies of 5G, Wi‐Fi, and the Internet of Medical Things to provide more effective and accessible intelligent healthcare delivery for individuals, families, and communities,^{[ 3 ]} thus holding great potential to transform traditionally slow and tedious hospital‐centered approaches into faster, more efficient, intelligent, personalized, and patient‐centric ones.^{[ 4 ]} This evolution will usher healthcare into a revolutionary intelligent era, where patients are no longer confined to clinical settings but can rely on wearables and smartphones. Doctors can gain insights into health problems even before symptoms arise, allowing for timely, personalized, and highly effective medical interventions without the need for in‐person visits.^{[ 5 ]}

In intelligent healthcare systems, wearable sensors can continuously collect real‐time physiological data without disrupting our daily lives.^{[ 4} , ^{6 ]} Due to their continuity, non/minimal invasiveness, and wide range of sensing targets from vital signs^{[ 7 ]} to biomarkers,^{[ 4} , ^{8 ]} wearable sensors can be used for various biomedical applications, such as monitoring lifestyles for preventive healthcare, tracking drug metabolism for smart medicine, and assessing physiological parameters for disease diagnosis and treatment. Smart materials such as self‐healing materials, metamaterials, and responsive materials can endow wearable sensors with high sensitivity and unique features. Although wearable sensors excel in collecting extensive data, they sometimes fail to generate meaningful actionable clinical information for direct diagnosis and decision‐making. ML facilitates deep data analysis, providing critical information for decision‐making,^{[ 9 ]} enabling more accurate, efficient, and effective disease diagnosis and treatment,^{[ 10 ]} and making the voluminous data collected from wearable sensors more instrumental in the management of diseases.^{[ 11 ]} Thus, the fusion of smart materials, ML with wearable sensors enables more intelligent and personalized diagnosis and therapy (Figure 1 ). Despite significant advancements in wearable sensors,^{[ 4} , ^{12 ]} artificial intelligence (AI),^{[ 13 ]} and AI‐based medical sensors,^{[ 11 ]} a comprehensive overview of smart materials and their synergy with ML in wearable sensing technologies is still lacking. This article aims to fill this gap by reviewing recent developments in wearable sensors and biomedical applications empowered by smart materials, such as self‐healing and metamaterials, alongside ML techniques.

Figure 1.

Overview of smart materials and ML‐enhanced wearable sensors for intelligent healthcare. Reproduced with permission.^{[ 14 ]} Copyright 2016, American Chemical Society. Other images of the figure are assets from Freepik.com.

In this review, we emphasize the fusion of wearable sensors with smart materials and ML technologies for intelligent healthcare and discuss their applications in diagnosis and therapy. The advances in smart materials and ML used in wearable sensors are first introduced. Wearable biophysical and biochemical sensors powered by ML and smart materials for intelligent healthcare use are then discussed, highlighting their applications in smart preventive healthcare, intelligent disease monitoring, smart medicine, and intelligent rehabilitation. We then conclude with current challenges and limitations of intelligent healthcare in transforming healthcare, along with the exciting prospects and opportunities for the future.

2. Smart Materials in Wearable Sensors

Although the configurations vary depending on applications, materials used in wearable sensors typically consist of sensing, conductive and substrate materials.^{[ 4} , ^{15 ]} Utilizing smart materials to construct these layers can provide wearable sensors with unique functionalities and enhanced performance, such as self‐healing after damage, degradation, or deformation when needed, and selective target sensing. Commonly used smart materials for flexible sensors include self‐healing materials, metamaterials, responsive sensing materials, biodegradable materials, and shape‐morphing materials. Biodegradable materials are primarily designed for implantable sensors, where biocompatibility and controlled degradation are essential for temporary medical monitoring.^{[ 16 ]} Their role in wearable sensors is limited due to challenges in durability, mechanical robustness, and long‐term stability required for continuous external use.^{[ 16 ]} Shape‐morphing materials, such as shape‐memory polymers and hydrogels, are more frequently applied in soft robotics, biomedical implants, and reconfigurable devices rather than wearable sensors.^{[ 17 ]} Their integration into wearable electronics remains relatively uncommon due to challenges in material responsiveness, integration complexity, and long‐term mechanical stability. Since this review focuses on wearable sensors, biodegradable materials and shape‐morphing materials are discussed only briefly, due to their limited use in wearable applications.^{[ 18 ]} Therefore, this review highlights self‐healing materials, metamaterials, and responsive sensing materials, which play a more significant role in enhancing the performance, adaptability, and durability of wearable sensors (Figure 2 ).

Figure 2.

Smart materials including self‐healing materials, metamaterials, and responsive materials. Self‐healing can be realized based on physical, chemical, or physiochemical interactions. Metamaterials mainly include mechanical, acoustic, and electromagnetic metamaterials. Mechanical metamaterials. Reproduced with permission.^{[ 19 ]} Copyright 2018, Wiley‐VCH. Acoustic metamaterials. Reproduced with permission.^{[ 20 ]} Copyright 2022, AAAS. Electromagnetic metamaterials. Reproduced with permission.^{[ 21 ]} Copyright 2024, AAAS. Mechanical responsive materials: polyvinylidene fluoride and lead zirconate titanate. Reproduced with permission.^{[ 22 ]} Copyright 2013, Elsevier. Optical responsive materials. Reproduced with permission.^{[ 23 ]} Copyright 2018, Springer Nature. Thermoresponsive materials: Poly(N‑isopropylacrylamide) (PNIPAM). Reproduced with permission.^{[ 24 ]} Copyright 2018, American Chemical Society. Electrical responsive materials: Prussian blue, ZnS:Cu. Reproduced with permission.^{[ 25 ]} Copyright 2021, Elsevier. Chemical responsive materials: polyaniline. Reproduced with permission.^{[ 26 ]} Copyright 2002, Elsevier.

2.1. Self‐Healing Materials

2.1.1. Mechanisms

Self‐healing materials are designed to automatically repair themselves after damage, thereby extending their lifespan and enhancing their reliability.^{[ 27 ]} Among them, highly conductive, non‐conductive, and target‐sensitive self‐healing materials can be used for wearable electrodes, substrates, and sensing elements, respectively. Self‐healing materials can be broadly classified into intrinsic and extrinsic types. Intrinsic self‐healing materials (Figure 3a) can heal autonomously without the addition of healing agents. They can repeatedly heal through the reorganization of the polymer matrix based on the diffusion of polymer chains and the reformation of broken dynamic bonds. Intrinsic self‐healing materials usually have low glass transition temperatures with enhanced polymer chain movement and abundant sites for covalent or non‐covalent interaction at the damaged interfaces.^{[ 28 ]} In contrast, extrinsic self‐healing materials (Figure 3b) require healing agents that are encapsulated in microcapsules or embedded within microvascular networks. These healing agents are usually reactive precursors and catalysts.^{[ 28b ]} When damage occurs, the microcapsules or microchannels in the microvascular networks rupture and release the healing agents, filling the damaged areas and repairing the cracks. The healing process can be triggered by light, temperature, or solvent exposure. To summarize, intrinsic materials offer repeatable healing but may lack mechanical strength and require suitable external conditions, while extrinsic materials provide strong, efficient healing but are often limited to one‐time use and have complicated compositions.

Figure 3.

Self‐healing materials. a,b) Mechanisms of intrinsic and extrinsic self‐healing materials. c) Self‐healing insulative materials. Reproduced with permission.^{[ 31 ]} Copyright 2021, Elsevier. d) Self‐healing conductive materials. Reproduced with permission.^{[ 35a ]} Copyright 2024, Springer Nature. e) Self‐healing thermo‐sensitive materials. Reproduced with permission.^{[ 38 ]} Copyright 2020, Wiley‐VCH. f) Self‐healing force‐sensitive materials. Reproduced with permission.^{[ 34 ]} Copyright 2022, Wiley‐VCH.

The self‐healing process that involves physical, chemical, or physiochemical interactions at the molecular level is often used.^{[ 27} , ^{29 ]} Chemical healing process involves the formation of covalent bonds, such as disulfide bonds, boronate ester bonds, imine bonds, and metal coordination. In contrast, the physical healing process is based on non‐covalent interactions such as hydrogen bonds, ionic bonds, electrostatic interactions, and host‐guest interactions, as well as supramolecular chemistry.^{[ 27b ]}

2.1.2. Self‐Healing Insulative Materials

Insulative self‐healing materials, usually made of elastomers such as Polydimethylsiloxane (PDMS) and Polyurethane (PU), can work as substrates in wearable sensors. Their self‐healing capability is typically achieved through dynamic covalent bonding or reversible non‐covalent interactions. For instance, self‐healing PDMS has been reported to utilize hydrogen bonding^{[ 30 ]} or metal‐ligand interaction^{[ 14 ]} to realize repair damage. Based on the self‐healing PDMS elastomer, a novel strain sensor was developed by immersing the treated self‐healing PDMS substrate in a Polypyrrole precursor solution and reacting in situ to form a strain‐sensitive layer (Figure 3c).^{[ 31 ]}

2.1.3. Self‐Healing Conductive Materials

Self‐healing conductive materials require both high conductivity and self‐healing capability. They are usually used for electrodes and sensing elements. These materials can be classified into intrinsically^{[ 32 ]} and composite^{[ 30} , ^{33 ]} conductive self‐healing materials. The former are usually conductive polymers and ions‐based hydrogels such as self‐healing poly(3,4‐ethylenedioxythiophene):poly(styrenesulfonate) (PEDOT:PSS)^{[ 32a ]} and poly(ionic liquids) hydrogels,^{[ 34 ]} while the latter are made through incorporating self‐healing materials with conductive materials.^{[ 30} , ^{35 ]} For instance, a self‐healing conductive heterogeneous multiphase conductor with continuous morphology and macroscale phase separation was reported with an insulative self‐healing phase composing hydroxyl‐terminated PDMS and boron oxide nanoparticles and a conductive phase of PEDOT:PSS, showing ultrafast universally autonomous self‐healing with full recovery of pristine tensile and electrical properties in less than 120 and 900 s (Figure 3d).^{[ 35a ]} Similarly, a conductive room‐temperature self‐healing composite consisting of a supramolecular organic polymer and metallic nickel nanostructured microparticles exhibits high conductivity of 40 S/cm and 15s healing time with 90% healing efficiency of the initial conductivity.^{[ 36 ]} Embedding a carbon nanotube network in self‐healing PDMS‐MPU_0.4‐IU_0.6 can also result in a self‐healing conductor, which also shows its capability for electrocardiogram (ECG) sensors and electrodes of strain sensors.^{[ 30a ]} Liquid metals, such as eutectic gallium‐indium (EGaIn) or gallium‐indium‐tin (Galinstan), exhibit self‐healing properties due to their fluidic nature and high surface tension. When damaged or fractured, the liquid metal can flow and coalesce back into a continuous structure, restoring electrical and mechanical functionality almost instantly. This makes such material ideal for applications in flexible electronics, wearable sensors, and soft robotics, where mechanical stress and deformation are commonly encountered. For instance, a bilayer liquid‐solid conductor (b‐LSC) with liquid metal^{[ 37 ]} as a conductive layer and VHB as an encapsulation layer is developed to achieve stretchable and self‐healable conductors.

2.1.4. Self‐Healing Sensing Materials

Self‐healing sensing materials are materials sensitive to stimuli such as temperature, pressure, and strain, which can lead to changes in electrical signals or color. Some of them are made of self‐healing conductive materials, which can be used as sensing materials for bioelectric sensors such as ECG sensors,^{[ 30} , ^{33 ]} temperature sensors,^{[ 38 ]} and force sensors such as tactile^{[ 30} , ^{39 ]} and strain sensors.^{[ 35b ]} For instance, a self‐healing conductive material is reported to act as a thermo‐sensitive element and achieved by free polymerization of monomers (3‐sulfopropyl methacrylate potassium salt and methyl methacrylate), which shows no signal degradation in responsivity under 200% strain or bending (Figure 3e).^{[ 38 ]} Besides, a self‐healable conductive hydrogel made of poly(acrylic acid), amorphous calcium carbonate, and MXene nanocomposite was also reported for strain sensors. The hydrogel can detect human motions and heal at room temperature without any external assistance.^{[ 35b ]} Similarly, an adhesive ionic conductive hydrogel made of cellulose nanofibrils and phenylboronic acid‐ionic liquid/acrylamide cross‐linked network was also reported to respond to strain as well, showing remarkable stretchability (1810 ± 38%), toughness (2.65 ± 0.03 MJ m⁻³), and self‐healing property (92 ± 2% efficiency) under the dynamic boronic ester bonds and physical interactions (hydrogen bonds and electrostatic interactions) (Figure 3f).^{[ 34 ]}

2.2. Metamaterials

Metamaterials are artificially designed and manufactured materials, exhibiting unprecedented physical properties that are not found in conventional materials.^{[ 40 ]} These unique characteristics include zero or negative Poisson's ratio, negative refractive index, negative magnetic permeability, and negative dielectric constant.^{[ 41 ]} Metamaterials can be broadly categorized into various types such as mechanical, acoustic, electromagnetic, and thermal materials, etc.^{[ 42 ]} Due to their unique characteristics, metamaterials play a crucial role in wearable sensor technology. They can enhance electromagnetic sensors by improving signal transmission in wearable antennas^{[ 43 ]} and wireless communication.^{[ 44 ]} In mechanical strain sensors, they can provide greater flexibility and high sensitivity for strain and pressure detection.^{[ 41} , ^{45 ]} Additionally, metamaterials are utilized in thermal and infrared sensors for wearable temperature regulation and thermal imaging,^{[ 46 ]} as well as in acoustic sensors,^{[ 47 ]} improving sound detection in wearable hearing devices.

2.2.1. Mechanical Metamaterials

Mechanical metamaterials are characterized by unique mechanical properties, such as negative elastic modulus, negative Poisson's ratio, and zero shear modulus.^{[ 40} , ⁴¹ , ^{48 ]} These materials can be fabricated based on topology optimization and ML‐based design methods.^{[ 48} , ^{49 ]} They are mainly categorized into origami, chiral, and lattice types based on the structural phases of their assembly units.^{[ 41} , ^{50 ]} These materials have found extensive applications in wearable sensing, particularly in tactile and strain sensors.^{[ 41} , ^{45 ]} For example, auxetic metamaterials have been reported to enhance the sensitivity of stretchable strain sensors by 24‐fold (Figure 4a).^{[ 19 ]} This improvement arises from the synergistic effect of the reduced structural Poisson's ratio on PDMS thin films combined with auxetic frames and strain concentration in the active single‐wall carbon nanotube layer. Unlike conventional strain sensors, which often face sensitivity limitations due to conflicting effects of bidirectional expansion and compression, auxetic mechanical metamaterials with a negative Poisson's ratio exhibit stretchability in both directions, thereby overcoming these limitations. Similarly, auxetic PU foam, obtained through hot‐pressing commercial PU foam and subsequent chemical modification, has been used to construct strain sensors capable of detecting compressive strains of up to ≈40% and tensile strains of up to 80%, offering a higher gauge factor, better linearity, and excellent stability.^{[ 51 ]} Additionally, 3D‐printed metamaterial capacitive sensing arrays with curved segments and structurally stretchable patterns have been reported to detect normal forces on curved deformable surfaces without interference from stretching stimuli.^{[ 52 ]}

Figure 4.

Metamaterials. a) Mechanical metamaterials. Reproduced with permission.^{[ 19 ]} Copyright 2018, Wiley‐VCH. b) Electromagnetic metamaterials with structured conductive fabrics. Reproduced with permission.^{[ 21 ]} Copyright 2024, AAAS. c) Electromagnetic metamaterials featuring a coaxially shielded internal structure. Reproduced with permission.^{[ 53 ]} Copyright 2024, Springer Nature. d) Acoustic metamaterials. Reproduced with permission.^{[ 54 ]} Copyright 2024, Springer Nature.

2.2.2. Electromagnetic Metamaterials

Electromagnetic metamaterials are synthetic structures engineered by arranging uniform metallic components to manipulate electromagnetic waves. They can be used to enhance electromagnetic wave transmission,^{[ 43 ]} mitigate interference from extraneous loading,^{[ 21} , ^{53 ]} and create secure near‐field multibody area networks.^{[ 44 ]} For example, Tian et al. reported an energy‐efficient and secure wireless body sensor network interconnected through radio surface plasmons propagating on metamaterial textiles made of conductive fabrics.^{[ 43 ]} The metamaterials can support surface‐plasmon‐like modes at radio communication frequencies, enhance transmission efficiencies by three orders of magnitude, and confine wireless communication to within 10 cm of the body, addressing issues related to energy inefficiency and vulnerability to eavesdropping. This group also demonstrated an electromagnetic metamaterial comprising a structured conductive fabric supporting spoof surface plasmonic modes, showing the capability to mediate near‐field interactions between body tissues and wireless signals for highly sensitive and interference‐immune sensing of vital signs such as respiration rate, heart rate, and pulse rate (Figure 4b).^{[ 21 ]} Additionally, a textile metamaterial featuring a coaxially‐shielded internal structure was reported to mitigate interference from extraneous loadings and enable communication between near‐field reading devices and battery‐free sensing nodes placed within the network (Figure 4c).^{[ 53 ]} Metamaterials built from arrays of discrete, anisotropic magneto‐inductive elements were also reported to create a mechanically flexible system capable of battery‐free communication among near‐field communication (NFC)‐enabled devices placed anywhere close to the network.^{[ 44 ]}

2.2.3. Acoustic Metamaterials

Acoustic metamaterials can manipulate sound waves in ways unattainable by conventional materials, exhibiting properties such as zero or negative sound refractive index, high anisotropic sound transmission, and acoustic invisibility for cloaking and camouflage.^{[ 47 ]} These characteristics make them highly effective for developing advanced sensors that manipulate sound waves, offering enhanced capabilities in health monitoring and medical imaging. For example, Tang et al. reported an injectable ultrasonic sensor made from absorbable metagel, comprising a hydrogel matrix embedded with periodic air columns.^{[ 54 ]} This metagel, with stimulus‐responsive functionalities (e.g., pressure, temperature, and pH responsiveness), can generate tunable acoustic reflection spectra, enabling the independent detection of intracranial pressure, temperature, pH, and flow rate (Figure 4d).^{[ 54 ]} Additionally, active acoustic metamaterials integrated with piezoelectric transducers can drive transdermal drug delivery through localized acoustic fields and acoustic streaming, allowing rapid and on‐demand management of acute diseases.^{[ 55 ]}

2.2.4. Thermal Metamaterials

Thermal metamaterials, engineered to manipulate heat flow in unprecedented ways, can be broadly categorized into conduction‐based, convection‐based, and radiation‐based metamaterials.^{[ 56 ]} Conduction‐based thermal metamaterials, such as thermal cloaks^{[ 57 ]} and concentrators^{[ 58 ]} guide heat flux around or toward specific regions, enabling applications in thermal management for electronics^{[ 46 ]} and energy harvesting. Convection‐based metamaterials optimize heat transfer in fluids, enhancing cooling efficiency in industrial systems. Radiation‐based thermal metamaterials, like photonic structures and emissivity‐tunable surfaces, control thermal emission and absorption through either far‐field or near‐field thermal radiation, proving useful in infrared camouflage^{[ 59 ]} and thermophotovoltaics.^{[ 60 ]} These categories collectively offer potential innovative solutions in the fields of smart materials.

2.3. Responsive Sensing Materials

Responsive materials are a group of materials whose physical or chemical properties can change in response to an external stimulus. These materials are utilized across various fields, including sensing, actuation, drug delivery, and imaging. When integrated with wearable healthcare sensors, responsive materials can serve as signal‐transducers, real‐time monitors, or response‐based controllers. Generally, responsive materials can be categorized based on the type of stimulus: physical stimuli, including mechanical, light, heat, electrical, and magnetic fields; and chemical stimuli, such as the presence of certain chemical species or pH changes.

2.3.1. Mechanical Responsive Materials

Mechanical stimulation refers to the application of physical forces, such as vibration or pressure, to a material or target, triggering a response based on the mechanical interaction. One of the most well‐known and widely used mechanical responsive materials is piezoelectric materials, such as polyvinylidene fluoride (PVDF) and lead zirconate titanate (PZT)‐based ceramics. Such materials generate electrical voltage over their surface in response to the external forces due to the lack of inversion symmetry of their lattice electrical state. It should also be noted that piezoelectricity is reversible; that is, the materials can also generate a mechanical response when applying electrical voltages. Applications of such materials range from vibrational wave sensing,^{[ 61 ]} ultrasound transducing,^{[ 62 ]} actuators,^{[ 63 ]} and even energy harvesting.^{[ 64 ]} Another group of mechanically responsive materials is the hydrogels that are incorporated with activable elements. These materials, depending on the variation of activatable elements, can achieve mechano‐chromatic change if integrated with mechanophores such as spiropyran and rhodamine (Figure 5a),^{[ 65 ]} strain‐stiffening if having helical topologies such as polyisocyanopeptides,^{[ 66 ]} and mechanical‐induced crosslinking if containing piezoelectric based initiators (zinc oxide) (Figure 5b).^{[ 67 ]}

Figure 5.

Responsive materials. a) Force‐induced chromic materials. Reproduced with permission.^{[ 65 ]} Copyright 2024, Wiley‐VCH. b) Mechanically induced crosslinking of hydrogel. Reproduce with permission.^{[ 67 ]} Copyright 2021, Springer Nature. c) Up‐conversion nanoparticles based in vivo 3d printing of scaffold. Reproduced with permission^{[ 72 ]} Copyright 2024, Wiley‐VCH. d) Photodiode‐based wireless UV sensor. Reproduced with permission.^{[ 77a ]} Copyright 2018, American Association for the Advancement of Science. e) Electroluminescent pressure sensor based on ZnS:Cu. Reproduced with permission.^{[ 78 ]} Copyright 2019, Wiley‐VCH. f) Magnetic field‐controlled microrobots based on ferrimagnetic materials. Reproduce with permission.^{[ 84 ]} Copyright 2023, American Association for the Advancement of Science. g) Thermoresponsive valve based on PNIPAM hydrogels. Reproduce with permission.^{[ 91 ]} Copyright 2020, American Association for the Advancement of Science. h) Chemical species responsive conformation changes of aptamers and corresponding electrochemical sensors. Reproduce with permission.^{[ 111 ]} Copyright 2005, Wiley‐VCH. i) Bacterial DNase‐induced degradation of DNA gels and its sensor. Reproduce with permission.^{[ 110 ]} Copyright 2021, American Association for the Advancement of Science.

2.3.2. Light‐Responsive Materials

Light‐responsive materials are widely utilized to build sensors, actuators, imaging systems, drug delivery devices, and even for therapeutic applications. Light‐induced color‐changing (photochromic) materials are widely used in smart windows,^{[ 68 ]} UV sensors,^{[ 69 ]} and transduction devices.^{[ 70 ]} Traditional photochromic materials, including silver halide compounds, azobenzene derivatives, and diarylethene, usually undergo a reversible alteration in chemical structure upon light exposure. Other than this, fluorescence molecules, quantum dots, and up‐conversion nanoparticles can also be categorized as photochromic materials since they respond to light stimulations by offering light with different wavelengths. For instance, nanocomposites of TiO_2‐x/rGO/MB‐based materials can achieve multiple color changes depending on the light intensities.^{[ 71 ]} Another example of utilizing photochromic materials is to utilize a core‐shell NaYF₄‐based up‐conversion nanoparticles to achieve non‐invasive 3D printing of bone scaffold (Figure 5c).^{[ 72 ]}

Another category of light‐responsive materials is liquid crystalline polymeric materials,^{[ 73 ]} which are polymer materials that consist of self‐assembled molecules that can form orderly structures. The inclusion of the azobenzene group in the molecular design offers photo‐responsive features as the azobenzene group undergoes reversible cis/trans isomerization when shining with UV light. Utilizing such properties, light‐driven soft actuators^{[ 74 ]} and robots can be made. For instance, a soft materials‐based robot with four stretchable legs has been developed using copolymers of liquid crystal monoacrylates and diacrylates as mesogen with azobenzene as the chromophore.^{[ 75 ]} Photodiode, which is a p‐n junction semiconductor device that can convert photons to currents, has been widely utilized in optical sensors and portable spectrometers.^{[ 76 ]} Recently, a wearable and broad‐band UV dosimeter with only fingernail size has been developed with a photodiode array for the control of solar exposure and precise phototherapy (Figure 5d).^{[ 77 ]}

2.3.3. Electrical and Magnetic Responsive Materials

Electrical responsive materials encompass a wide range of materials that can respond to external applied electric field. Electroluminescent and electrochromic materials generate optical responses when an electric field is applied. Examples of such materials include zinc sulfite (ZnS) and ZnS:Cu serving as electroluminescent materials under alternating current and if combined with the flexible dielectric materials, applications on wearable sensors of strain, temperature, and humidity (Figure 5e).^{[ 78 ]} Another category is electrochromic materials that undergo color change after applying voltages. Most of them, including Prussian blue and tungsten (IV) oxide, are electrochemically sensitive and alter their oxidizing states upon electrical stimulations. If combined with sensors, electrochromic materials can serve as optical displays for low‐power visualization of sensing signals.^{[ 79 ]} Other electrical responsive materials include ionic hydrogels that can have mechanical deformation due to osmotic pressure difference after applying electric fields,^{[ 80 ]} dielectric elastomers, and liquids that are actuated laterally by voltages.^{[ 81 ]}

Magnetic field responsive materials that are utilized in wearable sensors are typically soft magnetic composite materials that consist of a soft polymer‐based matrix, including hydrogels and elastomers, and dispersed magnetic elements such as hard magnetic materials (i.e., hexagonal ferrites)^{[ 82 ]} and soft magnetic materials (iron oxide nanoparticles).^{[ 83 ]} These magnetic elements align their magnetic dipoles according to the external magnetic stimulus. Torques, forces, and field‐induced stiffening are generated based on the alteration and strength of the applied magnetic field. Magnetic actuators and robots, including implantable soft robots, can easily be fabricated. For instance, soft magnetic microfiberbots are fabricated using styrene‐ethylene‐butylene‐styrene and ferromagnetic particles (NdFeB) to achieve robotic embolization (Figure 5f).^{[ 84 ]} Tactile sensors can also be designed utilizing a combination of magnetorheological elastomer and a magnetic to transduce force signals to local magnetic flux change.^{[ 85 ]}

2.3.4. Thermoresponsive Materials

Thermoresponsive materials mainly consist of polymers that either have a lower critical solution temperature (LCST) close to room temperature or possess shape memory capabilities. For the first category, when reaching the LCST, polymers undergo a reversible phase transition that changes their physical properties in terms of size and shape. A typical thermoresponsive polymer material is Poly(N‑isopropylacrylamide) (PNIPAM), which has an LCST of 32 °C and is therefore suitable for body temperature range sensing^{[ 86 ]} or actuation applications.^{[ 87 ]} Other polymers, including poly(N‐vinylcaprolactam),^{[ 88 ]} poly(ethylene glycol) (PEG) based copolymers,^{[ 89 ]} and polypeptide^{[ 90 ]} can also achieve responsive near body temperature. The application of thermoresponsive materials on sensors can be vast. For instance, PNIPAM has been utilized as a fluid‐controlling valve for the sweat sensor (Figure 5g)^{[ 91 ]} and a thermal switch of single molecule detection if coupled with Au nanoparticles. As for the shape memory polymers (SMPs), they can return to their original shape after being deformed when exposed to a specific temperature owing to the existence of at least two separated phases (e.g., combined crystalline and amorphous phase). Common shape memory polymers materials include polyurethanes,^{[ 92 ]} PEG‐based materials,^{[ 93 ]} and polylactic Acid (PLA).^{[ 94 ]} When combined with biosensors and electronics, such materials can serve as substrates with switchable rigidity and configurable shapes, which can reduce the size of surgical incisions.^{[ 95 ]} Other applications, such as nerve probes and stimulators,^{[ 96 ]} bone‐tissue engineering,^{[ 97 ]} and drug delivery,^{[ 98 ]} have also been developed using SMPs as substrates.

2.3.5. Chemical‐Responsive Material

Chemical‐responsive materials can generally be divided into two categories depending on their types of response. The first category involves changes in physical properties induced by the concentration or presence of certain species, including polymer swelling, self‐assembly, sol‐gel transitions, and host‐guest interactions. Examples include hydrogel swelling in aqueous conditions, deoxyribonucleic acid (DNA) oligonucleotide‐triggered assembly of gold nanoparticles, calcium ions‐induced gelation of alginate, and cyclodextrin/ferrocene interactio((ns. One example of this type of chemical‐responsive material is the pH‐sensitive hydrogels that have different swelling ratios owing to having either acid or basic functional groups on the polymer backbones. Typical applications of such material include stimuli‐responsive drug delivery/release^{[ 99 ]} and pH‐dependent sensors.^{[ 100 ]} Other than this, one of the applications of wearable sensors is electrochemical aptamer‐based biosensors. DNA/ribonucleic acid (RNA) oligonucleotides (aptamers) are specially designed as host molecules to the target analytes. Upon binding, the aptamer undergoes conformational changes (Figure 5h), which are then reported by electrochemical active species based on their redox activity. Metabolites,^{[ 101 ]} hormones,^{[ 102 ]} proteins,^{[ 103 ]} and even bacteria^{[ 104 ]} can be quantified through this approach.

The other kind of chemical‐responsive material involves changes triggered by chemical reactions. Specific reactions, including doping, redox, degradation, enzyme catalysis, and crosslinking can alter molecular structures and doping states, inducing macroscopic responses such as deformation,^{[ 105 ]} collapse,^{[ 106 ]} stiffness or color changing,^{[ 107 ]} and electrical properties alteration.^{[ 108 ]} One example of chemical‐responsive materials that have been applied in the wearable sensor field is the polyaniline‐based pH sensors.^{[ 109 ]} When doping with anions (such as chloride ions, and sulfonic ions) and exposed at low pH, its emeraldine forms can be reversibly changed to its emeraldine salt forms with an increase in electrical conductivity, increasing redox potential, and change of color. Another instance of chemical responsive materials that render healthcare sensing applications is the degradation of DNA‐gel under the presence of bacterial DNase to determine S. Aureus (Figure 5i).^{[ 110 ]} The rate of degradation is converted to electrical permittivity with the help of an interdigitated electrode and a half‐wave rectified LC module.

3. ML in Wearable Sensors

Wearable sensors can acquire continuous high‐resolution multi‐parametric data related to a patient's health condition. However, these complex data can consist of multivariable, non‐linear physiological patterns that are difficult to identify using conventional processing methods. To address this challenge, ML—defined as a computer program capable of acquiring knowledge by extracting distinct or desired features from raw data—has been integrated into wearable sensor technology. This enables accurate interpretation of data collected by wearable sensors, allowing both users and clinicians to conveniently and efficiently adjust treatments or take appropriate actions. ML has already shown great promise in advancing wearable technology for various applications, including diagnosing and predicting diseases,^{[ 112 ]} as well as monitoring^{[ 112} , ^{113 ]} and evaluating athletic performance (Table 1 ).^{[ 114 ]} ML can be broadly categorized into three types: unsupervised, supervised, and reinforcement learning (Figure 6a,b), each playing a crucial role in enhancing the functionality of intelligent wearable medical sensors. Unsupervised learning is widely used for pattern recognition, feature extraction, and clustering, enabling wearable sensors to identify hidden health trends without requiring labeled data. For example, clustering algorithms can be used to group patients based on their physiological signals (e.g., heart rate variability or sleep patterns), while dimensionality reduction techniques help refine raw sensor data, improving signal quality and interpretation. Supervised learning is essential for disease classification, anomaly detection, and predictive modeling, where labeled sensor data such as ECG, electroencephalogram (EEG), or photoplethysmography (PPG) signals are used to train models for diagnosing conditions such as arrhythmia or sleep apnea. Finally, reinforcement learning (RL) plays a critical role in closed‐loop control systems for wearable sensors, enabling real‐time decision‐making and personalized healthcare interventions. By continuously adapting based on user feedback and sensor data, it optimizes adaptive control mechanisms in applications such as closed‐loop glucose monitoring for diabetes management,^{[ 115 ]} and intelligent electrophysiological systems.^{[ 116 ]} Compared to supervised and unsupervised learning, RL is less frequently applied in wearable sensors due to several challenges. These include high computational demands, the need for extensive interaction data for training, real‐time processing constraints, and safety concerns in medical applications.^{[ 117 ]} For example, supervised learning leverages historical labeled data, and unsupervised learning discovers patterns in pre‐collected datasets, allowing both to function effectively with limited real‐time data. However, RL requires ongoing interactions with the environment to refine decision‐making policies. In the context of wearable medical sensors, continuously experimenting with a user's physiological responses poses ethical and safety concerns, making it difficult to acquire sufficient real‐world interaction data for RL training. Given these limitations and the fact that supervised and unsupervised learning currently play a more dominant role in wearable sensor data analysis, this review will primarily focus on their applications. In this section, we summarize several key processing techniques reported in the field of wearable sensors.

Table 1.

Applications of ML in Wearable Devices.

Sensor system	ML	Model type	Applications
Soft magnetoelasticity sensing system	PCA, SVM	Classification	Wearable system for assisted speech[ 120a ]
Triboelectric‐based finger‐bending sensors	PCA, SVM	Classification	Object shape recognition[ 120b ]
Motion capture system	RF	Classification	Gesture recognition[ 121 ]
Electronic tongue	DT	Classification	Detection of bovine mastitis in milk samples[ 112b ]
Piezoelectric sensor	DTW‐KNN	Classification	Facial movement recognition[ 112a ]
Graphene vibrotactile‐sensitive sensors	KNN	Classification	Material texture recognition[ 113a ]
Textile triboelectric sensor	FNN	Regression	Ambulatory cardiovascular monitoring[ 113b ]
Quadruple tactile sensors	MLP	Classification	Object recognition[ 122 ]
Triboelectric sensor	MLP	Classification	Tactile recognition and gesture cognition[ 123 ]
Soft sternal patch with multifunctional sensors	RNN	Classification	Sleep apnea and sleep stages recognition[ 126 ]
Smartphone sensor‐based	CNN‐LSTM	Classification	Human activity recognition[ 127 ]
Strain sensor array	CNN	Regression	Monitoring blood pressure and cardiac function[ 128 ]
Flexible fibre sensor	CNN	Classification	Human activities recognition[ 129 ]
Hydrogel Sensors	CNN, FNN	Classification	Soft actuator posture identification[ 112c ]
Triboelectric smart glove	PCA, CNN	Classification	Sign language recognition[ 130 ]

Note: ML, Machine Learning; PCA, Principal Component Analysis; SVM, Support Vector Machine; RF, Random Forest; DT, Decision Tree; DTW‐KNN, Dynamic‐Time Warping K‐Nearest Neighbors; K‐NN, K‐Nearest Neighbors; FNN, Feedforward neural network; MLP, Multilayer Perceptron; RNN, Recurrent Neural Networks; CNN‐LSTM, Convolutional Neural Networks and Long Short‐Term Memory; CNN, Convolutional Neural Networks.

Figure 6.

Schematics of ML algorithms. a) Supervised learning. b) Unsupervised learning. c) Support vector machine (SVM). d) Tree‐based algorithms including decision tree (DT) and random forest (RF). e) K‐nearest neighbors (K‐NN). f) Neural network‐based algorithms including FNN, RNN, and CNN.

3.1. Unsupervised ML Algorithms

Unsupervised ML algorithms are designed to infer patterns from datasets without explicit labels or predefined outputs. The primary objective of these algorithms is to discover inherent structures, relationships, or clusters in the data, enabling the identification of underlying patterns that may not be immediately apparent. Therefore, in the field of wearable sensors, unsupervised ML algorithms are often employed as data preprocessing tools to complement supervised learning algorithms, enabling more accurate and efficient model performance. Here, we briefly introduce some commonly used unsupervised algorithms for dimensionality reduction and clustering analysis in wearable sensor fields.

3.1.1. Data Dimensionality Reduction

Dimensionality reduction is a key feature of many unsupervised algorithms, as it enables the simplification of complex, high‐dimensional datasets by transforming them into a lower‐dimensional space while preserving essential patterns and relationships. Techniques such as Principal Component Analysis (PCA), t‐SNE (t‐distributed Stochastic Neighbor Embedding), and non‐negative matrix factorization are commonly used for this purpose. Among these, PCA remains one of the most prominent techniques due to its computational efficiency and ability to retain key data variance. For example, PCA is characterized by its ability to reduce the dimensionality of high‐dimensional datasets while preserving the most significant variance within the data. By transforming correlated variables into a smaller set of orthogonal (uncorrelated) principal components, PCA facilitates efficient data processing and interpretation. It effectively minimizes redundancy and noise, isolates the most relevant features, and enhances computational performance. Additionally, PCA is particularly advantageous for simplifying complex data structures from 3D to 2D, enabling better visualization, and improving the accuracy of subsequent ML tasks such as classification and clustering. For example, PCA is adopted to reduce the dimensionality of waveforms and statistically evaluate the triboelectric outputs collected from elastoplastic silk fibroin electrode bioelectronics during the workout.^{[ 114a ]} Similarly, PCA enables optimization of the signal‐to‐noise ratio (SNR) from various single‐unit action potentials collected through the nanostructured PEDOT functionalization on the liquid metal sensing platform, ensuring clearer electrophysiological insights.^{[ 114b ]}

3.1.2. Clustering Analysis

Clustering analysis is the core aspect of unsupervised algorithms, aimed at discovering inherent patterns and grouping data points based on their similarities, all without requiring predefined labels. Algorithms such as K‐Means, Hierarchical Clustering, and Density‐Based Spatial Clustering of Applications with Noise are commonly used to divide data into distinct clusters, ensuring that points within the same cluster are more similar to each other than to those in different clusters. Among these, Hierarchical Clustering Analysis (HCA) stands out for its flexibility and adaptability to diverse data distributions. For example, HCA is characterized by its ability to build a nested hierarchy of clusters without the need to predefine the number of clusters. The method can be agglomerative, where each data point starts as its cluster and is successively merged based on similarity, or divisive, where all data points begin in one cluster and are recursively split. HCA's flexibility in allowing the user to define clusters at any level of the hierarchy, along with its capability to adapt to non‐spherical and unevenly sized clusters, is a key advantage. Additionally, the dendrogram—a tree‐like representation of the clustering process—provides a visual and intuitive way to understand relationships within the data, offering a robust framework for segmenting populations or activities based on sensor‐derived features and delivering actionable insights into behavior and health metrics for wearable sensor applications. For example, a hierarchical clustering approach using Ward's linkage algorithm with Euclidean distances is applied to wearable‐derived hourly average accelerometer activity data as a novel digital biomarker to identify distinct clusters and segment the population based on activity patterns.^{[ 118 ]} Moreover, HCA has been effectively employed in creating classifiers to detect motor activities in patients with chronic obstructive pulmonary disease, utilizing data recorded from wearable sensing systems.^{[ 119 ]}

3.2. Supervised ML Algorithms

In supervised learning, the algorithm is trained on labeled data, which means each input comes with an associated output (or label). The model learns to predict the output based on input data by minimizing the error between its predictions and the actual labeled outcomes. Common supervised learning algorithms in wearable sensors include Support Vector Machine, Decision Tree, K‐Nearest Neighbors, Neural networks etc. These algorithms are widely applied in tasks such as activity recognition,^{[ 120 ]} health monitoring,^{[ 113a ]} and disease diagnosis,^{[ 112a ]} where historical data with known outcomes are used to make predictions for future cases.

3.2.1. Support Vector Machine

Support Vector Machine (SVM) is a powerful supervised learning algorithm widely used for classification and regression tasks. The main idea of SVM is to find the optimal hyperplane that most effectively separates data points belonging to different classes within a high‐dimensional space. This hyperplane is determined by maximizing the margin between the nearest data points from each class, known as support vectors, which are crucial for defining the decision boundary (Figure 6c). SVMs are particularly effective for linearly separable data and can be adapted to handle non‐linear decision boundaries using the kernel method, which employs functions such as polynomial and sigmoid kernels. These methods have diverse applications for intelligent wearable sensing, such as assisted speaking recognition^{[ 120a ]} and object recognition.^{[ 13 ]} For example, Che et al. developed a self‐powered wearable system that enables assisted speaking without vocal fold reliance.^{[ 120a ]} The system captures laryngeal muscle movements and converts them into electrical signals, which are then processed using the SVM algorithm to generate speech signals with high fidelity, achieving an accuracy of 94.68% (Figure 7a). Zhu et al. introduced a haptic‐feedback smart glove as a promising and cost‐effective solution for advanced human‐machine interaction.^{[ 120b ]} This glove enables multi‐dimensional manipulation, provides haptic feedback, and performs real‐time object recognition through an SVM‐based model. Utilizing ML techniques, the smart glove achieves an object recognition accuracy of 96% (Figure 7b). When using SVMs in scenarios where the number of features far exceeds the number of samples, avoiding overfitting becomes critical. This requires careful selection of kernel functions and the regularization term to ensure the model generalizes well to new data. Additionally, SVMs do not natively provide probability estimates, and these probabilities are typically calculated through a computationally expensive five‐fold cross‐validation process. This can add computational complexity and increase the time required for model training and evaluation.

Figure 7.

Smart sensor with ML algorithms based on SVM, RF, and KNN. a) Speaking recognition with SVM algorithm. Reproduced with permission.^{[ 120a ]} Copyright 2024, Springer Nature. b) Object shape recognition with SVM algorithm. Reproduced with permission.^{[ 120b ]} Copyright 2020, American Association for the Advancement of Science. c) Gesture recognition with RF algorithm. Reproduced with permission.^{[ 121 ]} Copyright 2019, Elsevier. d) Facial movements decoding with K‐NN‐based algorithm. Reproduced with permission.^{[ 112a ]} Copyright 2024, Springer Nature.

3.2.2. Tree‐Based Algorithms

Decision Trees (DTs) are algorithms that facilitate decision‐making and prediction using a tree‐like structure. It works by dividing the dataset into subsets based on feature values and repeating this process for each subset until a predefined stopping criterion is reached. Each internal node in the tree represents a decision rule, while each leaf node gives the final classification or regression result. One significant advantage of DTs is its interpretability, as it can be easily visualized and is suitable for explaining the decision‐making process. Random Forest (RF) constructs an ensemble of individual DTs by combining their classification and regression outputs into a single comprehensive result (Figure 6d). This methodology effectively addresses overfitting and decreases the prediction variance that can occur with a single DT. For example, a gesture recognition system for hand movements was developed, utilizing wireless wearable motion capture data combined with an RF algorithm to achieve precise classification (Figure 7c).^{[ 121 ]} Soares et al. developed an electronic tongue based on impedance spectroscopy for detecting bovine mastitis in milk samples. By employing ML techniques with DT models for classification, the system achieved highly accurate results.^{[ 112b ]} Although DTs are popular ML algorithms, they are unstable, as small changes in input data can lead to significant alterations in the tree's structure. Additionally, DTs may favor the majority class in imbalanced datasets, resulting in biased decisions and poor testing performance. They are also susceptible to overfitting, as they tend to capture noise in the training data, which does not generalize well to new, unseen data.

3.2.3. K‐Nearest Neighbors

The K‐Nearest Neighbors (K‐NN) algorithm operates based on the distances among data points in the feature space. For classification tasks, it assigns a new sample to the category most frequently represented by its nearest neighbors. While for classification tasks, it predicts the mean value of its neighbors (Figure 6e). As a non‐parametric algorithm, K‐NN does not assume any specific distribution of the data, allowing it to adapt to a wide range of data patterns and structures. Additionally, this flexibility enables K‐NN to effectively handle multi‐class classification problems. For example, Sun et al. designed an integrated system capable of decoding facial strains and predicting facial kinematics in both healthy individuals and patients with amyotrophic lateral sclerosis (Figure 7d).^{[ 112a ]} This system integrates piezoelectric thin films with classification algorithms to accurately decode facial movements, utilizing a dynamic‐time warping K‐NN (DTW‐KNN) model to predict motions by comparing detected voltage waveforms with those in the training set. Similarly, Yao et al. introduced an innovative tactile sensor with a uniform graphene coating on microstructured elastomers.^{[ 113a ]} The sensor demonstrated applications in epidermal signal monitoring across various arteries and in surface texture recognition, achieving over 95% accuracy when combined with the K‐NN algorithm.

3.2.4. Neural Network‐Based Algorithms

The progress of ML algorithms has brought significant attention to neural network‐based models. Neural networks, characterized by their multi‐layered structures and end‐to‐end learning approach, have shown remarkable performance across numerous tasks. By learning complex feature representations of data, neural networks can capture nonlinear relationships within the data and demonstrate stronger generalization capabilities on extensive datasets. This neural network‐based approach not only automatically learns features from massive datasets but also adapts to different types of data and tasks, providing a more flexible and powerful tool for addressing various problems. Advances in wearable biosensors now allow for the collection of extensive physiological signal data from physical activities. This facilitates the use of large datasets to train artificial neural networks, leading to improved performance of smart wearable devices. Below, we mainly discuss three common types of neural network‐based algorithms in wearable sensor applications, which include feedforward neural networks, recurrent neural networks, and convolution neural networks.

Feedforward Neural Networks (FNNs)

FNN is a type of artificial neural network composed of an input layer, one or more hidden layers, and an output layer. Information in an FNN flows in a unidirectional manner, moving from the input nodes through the hidden layers to the output nodes. Each node in a layer connects to every node in the next layer, with the connections weighted by parameters that are learned and optimized during the training process (Figure 6f). FNNs are commonly trained using backpropagation, an optimization algorithm that employs the chain rule of calculus to compute gradients, thereby minimizing a predefined loss function through iterative updates to the weights of networks. The training process adjusts the weights to reduce the discrepancy between predicted outputs and actual targets. FNNs are commonly used for classification and regression tasks where the relationship between input features and outputs is relatively straightforward. They are applied in areas such as ambulatory cardiovascular monitoring,^{[ 113b ]} object recognition,^{[ 122 ]} and gesture and material cognition.^{[ 123 ]} For example, Fang et al. proposed a cost‐effective, lightweight, and mechanically robust textile triboelectric sensor that converts minor skin deformations from arterial pulsations into electricity, allowing for the accurate and consistent measurement of systolic and diastolic pressures using FNN algorithms (Figure 8a).^{[ 113b ]} In addition, Li et al. developed a robot hand with quadruple tactile sensors for precise object recognition through grasping. By combining tactile data with a multilayer perceptron (MLP) model, the smart hand can accurately identify various shapes, sizes, and materials, demonstrating high accuracy in classifying seven types of garbage (Figure 8c).^{[ 122 ]} Moreover, Niu et al. proposed an AI‐enhanced full‐skin electronic skin (e‐skin) that enables advanced tactile recognition for gestures and robot interactions. The cognitive system combines a six‐layer MLP neural network with full‐skin bionic e‐skin, enabling real‐time recognition of an object's material type and location with touching.^{[ 123 ]}

Figure 8.

Smart sensor with neural network‐based ML. a) Ambulatory cardiovascular monitoring with FNN algorithm. Reproduced with permission.^{[ 113b ]} Copyright 2021, Wiley‐VCH. b) Acute hemodynamic monitoring with RNN algorithm. Reproduced with permission.^{[ 126 ]} Copyright 2021, American Association for the Advancement of Science. c) Object recognition with MLP algorithm. Reproduced with permission.^{[ 122 ]} Copyright 2020, American Association for the Advancement of Science. d) Human activities recognition with CNN algorithm. Reproduced with permission.^{[ 129 ]} Copyright 2021, Springer Nature. e) Sign language recognition with a CNN‐based algorithm. Reproduced with permission.^{[ 130 ]} Copyright 2021, Springer Nature.

Recurrent Neural Networks (RNNs)

RNNs represent a class of neural networks designed to process sequential data by integrating temporal dynamics into their architecture. In contrast to FNNs, RNNs employ recurrent connections to retain information from previous inputs, thereby improving their capacity to represent temporal relationships and capture sequential patterns within the data (Figure 6f). Training RNNs requires extending the standard backpropagation method to handle sequential data by unfolding the network across time steps and computing gradients. However, traditional RNNs often encounter challenges such as vanishing and exploding gradients, which can significantly impair their performance on long sequences. To address these problems, advanced RNN architectures such as Long Short‐Term Memory (LSTM) networks and Gated Recurrent Units (GRU) have been developed. These architectures include specialized gating mechanisms that enhance the network's ability to capture long‐term dependencies and improve performance on complex sequential tasks. RNNs and their variants are extensively used in applications such as natural language processing^{[ 124 ]} and time series analysis.^{[ 125 ]} For instance, a soft sternal patch was developed to monitor acute hemodynamic disturbances and autodetect apneas and hypopneas during sleep. Leveraging an RNN algorithm, the device enables automatic features extraction from the collected data, offering a novel approach to sleep monitoring and cardiovascular assessment. In preliminary at‐home trials with symptomatic patients, the system demonstrated 95% precision, compared to the data scored by certified sleep clinicians (Figure 8b).^{[ 126 ]} Additionally, Khatun et al. proposed a hybrid model that combines convolutional neural networks and long short‐term memory networks (CNN‐LSTM) for human activity detection, achieving higher accuracy in recognition compared to other models.^{[ 127 ]}

Convolutional Neural Networks (CNNs)

CNNs are the most widely used neural network algorithm for various computer vision‐related tasks, including image recognition, object detection, and facial expressions analysis. CNNs are characterized by their hierarchical structure of convolutional layers, pooling layers, and fully connected layers (Figure 6f), which enable them to automatically learn and extract spatial hierarchies of features from input data. Convolutional layers in CNNs use shared filters to detect local features like edges and shapes, efficiently capturing spatial dependencies with fewer parameters. Following this, pooling layers down‐sample the feature maps, reducing dimensionality while retaining key information, which enhances computational efficiency and reduces overfitting. The final stage involves one or more fully connected layers that integrate these features for tasks such as classification or regression. The ability of CNNs to learn complex features and patterns in an end‐to‐end manner has led to their widespread adoption and success in various applications within the field of ML and AI. For example, an advanced system was developed to monitor blood pressure and cardiac function by integrating a conformal, flexible strain sensor array with CNN, enabling accurate measurement of pulse waves from the wrist without precise positioning.^{[ 128 ]} Loke et al. presented fiber‐integrated digital electronics capable measuring and storing physiological parameters, while also incorporating neural networks for analyzing sensory data (Figure 8d). Using a well‐trained CNN model, this fabric‐based device can accurately recognize human activities with an average accuracy of 96.4%.^{[ 129 ]} To decode sensing signals and identify diverse stimuli, a ML model incorporating a 1D CNN and a FNN was implemented, accurately predicting changes in the posture of soft actuators.^{[ 112c ]} Wen et al. proposed an AI‐enabled sign language recognition system, utilizing a CNN model combined with principal component analysis to interpret and reconstruct sentences from individual word elements, achieving a high sentence recognition accuracy of 95% (Figure 8e).^{[ 130 ]}

4. Intelligent Wearable Medical Sensors Based on Smart Materials and ML

Wearable sensors can be worn non/minimal‐invasively on the body to monitor and collect various physiological metrics.^{[ 4f ]} They convert these metrics into electrical or optical signals, which are then transmitted to external devices such as computers, smartphones and watches for recording, processing, and display.^{[ 4} , ^{131 ]} These sensors function as perception units and data grippers in intelligent healthcare, seamlessly incorporating ML to assist in decision‐making.^{[ 131} , ^{132 ]} Depending on their working mechanisms, wearable sensors can be categorized into several types: biochemical sensors, bioelectrical sensors, biomechanical sensors, optical sensors, thermal sensors, and other wearable sensors.

4.1. Biophysical Sensors

Biophysical sensors encompass a broad category of sensors that measure various physiological signals based on physical principles. These sensors play a crucial role in intelligent wearable medical devices by capturing essential biomarkers for health monitoring and disease diagnosis. Depending on the type of signal detected, biophysical sensors can be further categorized into several subtypes. Among them, biopotential sensors focus on detecting electrical signals generated by biological activity,^{[ 133 ]} while electro‐mechanical sensors measure mechanical properties such as pressure and strain.^{[ 134 ]} Additionally, thermal sensors monitor temperature variations,^{[ 135 ]} and optical and colorimetric sensors leverage light‐based techniques to assess biochemical and physiological parameters.^{[ 136 ]} In the following subsections, we will explore these sensor categories in detail, beginning with biopotential sensors.

4.1.1. Biopotential Sensors

Biopotential signals typically reflect the activity of specific groups of cells, originating from the membrane potential and the synchronized polarization/depolarization processes within cells.^{[ 137 ]} Commonly recorded biopotential signals include ECG, electromyograms (EMG), EEG, and electrooculograms (EOG).^{[ 133 ]} Recent innovations in biosensor technology have led to the development of flexible and stretchable patch‐type devices, which are revolutionizing the monitoring of biopotential signals.

These wearable sensors are designed to conform closely to the skin, significantly enhancing both wearability and user comfort. For example, ECG sensors can be fabricated by patterning serpentine electrodes onto flexible or stretchable substrates (Figure 9a)^{[ 138 ]} or by employing conductive inks made from metal nanomaterials, such as nanoparticles, nanosheets, and nanowires, liquid metals, or polymer on stretchable substrates (Figure 9b).^{[ 139 ]} These techniques ensure conformal contact between the skin and electrodes, resulting in low skin‐electrode contact impedance and a high SNR.

Figure 9.

Intelligent biopotential sensors. a) An ECG sensor that integrates stretchable thin‐film circuits with elastomers. Reproduce with permission.^{[ 138a ]} Copyright 2019, Wiley‐VCH. b) A biocomposite conformal and adhesive polymer electrode based on silk fibroin through interfacial polymerization with pyrrole. Reproduce with permission.^{[ 139c ]} Copyright 2020, American Chemical Society. c) A wireless EEG measurement earbud and three open‐mesh, structured, tattoo‐like electrodes and connectors attached to recording sites. Reproduce with permission.^{[ 140a ]} Copyright 2022, Springer Nature. d) A flexible in‐ear bioelectronic device that adaptively expands and spirals along the auditory meatus under electrothermal actuation, ensuring conformal contact without over‐constraining the meatus. Reproduce with permission.^{[ 140b ]} Copyright 2023, Springer Nature. e) A stretchable EMG sensor array designed for recognizing both static and dynamic gestures. Reproduce with permission.^{[ 141b ]} Copyright 2023, Springer Nature. f)A custom‐designed, flexible 16 × 4 electrode array that conforms to the forearm, enabling high‐density, large‐area EMG recordings without the need for individual wires. Reproduce with permission.^{[ 141c ]} Copyright 2021, Springer Nature. g) The aerosol jet–printed EOG electrodes, featuring highly stretchable, low‐profile biopotential electrodes, enable comfortable and seamless integration with a therapeutic VR environment. Reproduce with permission.^{[ 145 ]} Copyright 2020, American Association for the Advancement of Science.

These sensors are also transforming traditional signal acquisition methods. For instance, in EEG monitoring, in‐ear wearable sensors can be worn for extended periods, featuring lightweight and snug designs that significantly reduce motion artifacts, thereby delivering high‐quality EEG signals even during physical activities (Figure 9c,d).^{[ 140 ]} The use of dry electrode technology overcomes the signal degradation issues commonly associated with traditional wet electrodes, which often suffer from electrolyte drying. This advancement enables longer‐term, high‐quality EEG monitoring, making it suitable for a broader range of applications.

Moreover, these devices can support higher‐density electrode arrays, allowing for more detailed signal acquisition. In the context of EMG monitoring, flexible wearable sensors can cover larger skin areas with increased electrode density, capturing a more comprehensive profile of muscle activity (Figure 9e,f).^{[ 141 ]} Such multichannel profiles can be further used to extract features related to multi‐degree hand gesture recognitions, which involves the activation of multiple muscles.^{[ 142 ]} Additionally, they provide raw data for active rehabilitation systems designed for disabled patients,^{[ 143 ]} and deliver control information for the operation of automated prosthetic devices.^{[ 144 ]} Flexible design and skin conformity of these wearable sensors also facilitates seamless integration with other electronic devices and diverse sensing technologies. For example, when combined with virtual reality (VR) systems, flexible electrodes can be used for eye‐tracking and therapeutic applications (Figure 9g).^{[ 145 ]} This combination not only enhances the accuracy of signal detection but also expands the potential for remote therapy in home settings.

The integration of ML further amplifies the capabilities of these advanced wearable sensors, fully leveraging their precision and flexibility to advance personalized medicine and intelligent health monitoring. For example, in ECG monitoring, ML can automatically detect and classify heart rhythm abnormalities;^{[ 146 ]} in EMG, ML can precisely interpret muscle activity patterns for diagnosis and rehabilitation guidance;^{[ 147 ]} in EEG, ML can identify complex brainwave patterns, aiding in the early diagnosis of neurological disorders and the development of brain‐computer interfaces;^{[ 140} , ^{148 ]} and in EOG, ML is utilized for eye‐tracking, improving vision rehabilitation and human‐computer interaction.^{[ 149 ]} Through these applications, ML not only enhances the analysis of biopotential signals but also drives progress in intelligent health monitoring and personalized healthcare.^{[ 139 ]} Through these applications, ML not only enhances the analysis of biopotential signals but also drives progress in intelligent health monitoring and personalized healthcare.

4.1.2. Electro‐Mechanical Sensors

Electro‐mechanical sensors are often used to detect biomechanical signals from humans, such as motion, pulsation, respiration, and heartbeat. Motion detection often involves the use of pressure sensors, strain sensors, accelerometers, gyroscopes, inclinometers, and global positioning systems (GPS). Some of them can be integrated into smartwatches for the monitoring of our exact physical locations. The collected data can be utilized for applications such as rehabilitation after injuries or diseases, general activity, exercise, and fitness monitoring, as well as fall risk estimation, fall warnings, and alerts for the elderly and those who are at a high fall risk. Among all the mechanical sensors that have been proposed to integrate with a wearable sensing system, pressure and strain sensors are of great interest since they offer direct and reliable capturing of human biomechanics. Joint movements,^{[ 134 ]} gestures,^{[ 150 ]} perceptions of stiffness,^{[ 151 ]} and tactile impressions^{[ 152 ]} can be accessed via those wearable sensors. The mechanisms of such sensors are either resistive,^{[ 153 ]} piezoelectric,^{[ 154 ]} triboelectric^{[ 155 ]} or capacitive,^{[ 156 ]} which all provide electrical signals. Owing to their flexibility in forms and designs, they have been integrated into different wearable devices, such as gloves to capture finger movement parameters and assist motor skill assessment,^{[ 157 ]} and prosthetics to monitor pressures applied on bodies.^{[ 158 ]} Almost all strain and pressure sensors that are integrated with ML have sensor arrays. Multichannel signals collected on different positions of the human body are fed into ML models to do classifications of movement modules, calibration of sensor performance, signal interpretations and realize human‐machine interface. For instance, a haptic glove with yarn‐type strain sensors^{[ 159 ]} (Figure 10a) on each finger joint realizes the gesture‐to‐speech translation with the help of CNN. Another example is to utilize a soft and flexible strain sensor to conformably attach to the skin and recognize high‐resolution pulse signals (Figure 10b), which is then converted to blood pressure and other cardiovascular‐related parameters using a deep learning model.^{[ 128 ]} Besides applications in the field of biomechanics, electro‐mechanical sensors, such as accelerometers, can also be applied to detect sound and vibration signals. Piezoelectric responsive materials, such as PVDF^{[ 160 ]} and PZT,^{[ 161 ]} have been utilized to fabricate flexible ultrasound‐based systems. By integrating multiple ultrasound units, wearable ultrasound devices can achieve real‐time imaging of the heart (Figure 10c)^{[ 162 ]} and withdraw blood pressure information (Figure 10d).^{[ 163 ]} Another category of vibration‐based sensors is a mechano‐acoustic sensor that records acoustic waves through accelerometers. Human sounds, such as speech and cough,^{[ 164 ]} and tissue stiffness information^{[ 165 ]} can be captured and analyzed to reflect rehabilitation status and disease progressions. ML‐based classification methods enable precise and fast feature extractions. For instance, a mechano‐acoustic sensor (Figure 10e) is placed near the throat position to provide clinical‐grade ambulatory monitoring of COVID‐19 conditions with the help of SVM on the data processing and feature extraction.^{[ 166 ]}

Figure 10.

Intelligent electro‐mechanical and thermal sensors a) Wearable yarn‐based strain sensor for ML‐assisted gesture‐to‐speech translation. Reproduce with permission.^{[ 159 ]} Copyright 2020, American Association for the Advancement of Science. b) Textile‐based pressure sensor for deep‐learning‐based cardiac signals. Reproduce with permission.^{[ 128 ]} Copyright 2023, American Association for the Advancement of Science. c) Flexible ultrasonic device for imaging of the heart. Reproduce with permission.^{[ 162 ]} Copyright 2023, Springer Nature. d) PVDF flexible sensor for ultrasound‐based blood pressure sensor. Reproduce with permission.^{[ 163 ]} Copyright 2023, Springer Nature. e) Mechano‐acoustic sensor for ML‐assisted vibrational signal recognitions. Reproduce with permission.^{[ 166 ]} Copyright 2019, Springer Nature. f) Temperature sensor‐based skin hydration sensor system. Reproduce with permission.^{[ 173 ]} Copyright 2020, American Association for the Advancement of Science.

4.1.3. Thermal Sensor

Body temperature is one of the most important vital signs, closely related to an individual's health status. Deviation from normal body temperature typically represents an abnormality and the onset of diseases. For example, an elevated core temperature may indicate infection or fever, while decreased temperature may suggest an underactive thyroid,^{[ 167 ]} low blood sugar, or circulation problems.^{[ 168 ]} However, collecting reliable accurate core body temperature data requires placing temperature sensors on specific locations, such as mouth, rectum, and armpit, as exposed skin temperature can be easily interfered with by the environment. To address this, combining three on‐body temperature sensors that captures the thermal condition of an occupant with an external temperature/humidity sensor that monitors the ambient thermal condition can help thermoregulate human body.^{[ 135 ]} Also, together with an ML model (random forest classifier) for thermal comfort prediction, the built human‐in‐the‐loop can achieve energy‐friendly control of indoor temperature and, which can improve the thermal comfort of an occupant, contribute to the occupational health of workers in extreme working conditions.

Local temperature is also an important physiological indicator for tissue or organs such as a wound or transplanted kidney. For instance, increased wound temperature often signals a bacterial infection.^{[ 169 ]} Monitoring wound temperature together with other biomarkers such as pH, uric acid, and cytokines^{[ 170 ]} enables wearable wireless platforms to provide digitalized and quantifiable data for chronic wound management. Additionally, localized kidney temperature (T_kidney) increase is an early indicator of transplant rejection.^{[ 171 ]} A minimally invasive implantable biosensor has been reported to be sutured beneath the renal capsule to ensure stable contact with the kidney and connected via fine wires to a wireless module secured inside the abdominal cavity. This system enables continuous, real‐time monitoring of T_kidney and thermal conductivity (K_kidney), detecting rejection earlier than traditional biomarkers like creatinine.^{[ 172 ]} Wearable temperature sensors can even assess skin hydration levels by tracking the temperature changes induced by an external heat source (Figure 10f).^{[ 173 ]}

4.1.4. Optical and Colorimetric Sensors

Optical sensors usually detect targets by changes in characteristic wavelengths. One prominent example is PPG sensor, commonly used for measuring blood oxygen saturation (SpO₂) and heart rate. Measurement of SpO₂ generally involves employing two light sources at different wavelengths along with a photodetector. The absorbance at these wavelengths is different for hemoglobin bound or unbound to oxygen. Therefore, the percentage of hemoglobin bound to oxygen can be measured by calculating their absorptions. Ultralow SpO₂ usually indicates abnormal oxygen saturation, which can be related to severe pneumonia and respiration problems.^{[ 174 ]}

In addition to SpO₂, other physiological parameters such as heart rate and even blood pressure (BP)^{[ 175 ]} can also be extracted from the PPG data. PPG module can be easily integrated into smartwatches and smart wristbands to realize continuous monitoring. Based on techniques of pulse transit time (PTT),^{[ 176 ]} pulse arrival time (PAT),^{[ 176b ]} and ML,^{[ 177 ]} wearable cuff‐less continuous BP monitoring can also be achieved. The technologies of PTT and PAT for BP monitoring usually require two PPGs or integrating PPG with ECG or pulse sensors.

ML for BP monitoring usually utilizes algorithms to extract features from PPG, ECG, or pulse signals and generates outputs representing the systolic and diastolic BP. For instance, three neural network architectures (AlexNet, ResNet, and customized neural networks) are utilized to predict BP from pure PPG signals.^{[ 136 ]} With the training of the data points of 3750 subjects and subsequent fine‐tuning, the mean average errors are reduced to 12.5 mmHg for systolic BP and 8.5 mmHg for diastolic BP. Another example utilizes a textile triboelectric sensor that replaces traditional optical‐based PPG to withdraw pulse information.^{[ 178 ]} With the help of a supervised feedforward neural network architecture and a published dataset, a very small mean deviation (2.9% and 1.2% for systolic BP and diastolic BP, respectively) is achieved and validated through a comparison experiment with commercial pressure meters. Beyond BP, PPG data can support ML‐based analyses of heart rate, heart rate variability, stress level, sleep patterns, and fitness level.^{[ 179 ]}

4.2. Biochemical Sensor

Obtaining biomolecular‐level information is also essential for the prevention, diagnosis, and management of diseases. Biochemical sensors that measure biomarkers, such as glucose, lactate, and cortisol, offer critical insights into health. Wearable biochemical sensors enable continuous biomarker monitoring, providing temporal pattern analysis during critical events, such as post‐meal glucose spikes or pharmacokinetics. These sensors differ from physical ones as they require direct biofluid contact to access the target molecules. Commonly used biofluids include blood, interstitial fluids, sweat, saliva, and tears. Once the biofluids are correctly sampled, metabolites (such as glucose,^{[ 180 ]} lactate,^{[ 181 ]} uric acid),^{[ 182 ]} nutrients (such as Vitamin C,^{[ 183 ]} amino acids),^{[ 7b ]} drugs,^{[ 182} , ^{184 ]} pathogens^{[ 110} , ^{185 ]} and even proteins^{[ 186 ]} can be detected through wearable biochemical sensors. Concentration information is then converted to electrical or optical signals through electrochemical, colorimetric, and fluorescence methods.

Glucose sensors, linked to diabetes management,^{[ 187 ]} have garnered significant attention.^{[ 188 ]} Wireless continuous glucose monitoring sensors, such as the Abbott Freestyle Libre 3 (Abbott Diabetes Care. Ltd., United States), and Dexcom G6 (Dexcom Inc., United States), utilize electrochemical methods for precise glucose tracking, with mean absolute relative deviations under 10% up to 14 days.^{[ 189 ]} Daily glucose data are collected and uploaded to cloud storage and then used to predict glucose trends via the integration of data models,^{[ 190 ]} including kernel ridge regression and support vector regression. Though achieving commercial success, the sensors’ invasivenes and high costs impede their broader adoption. To solve this problem, some minimally invasive (Figure 11a) and non‐invasive (Figure 11b) wearable platforms have been proposed using sweat analysis,^{[ 191 ]} microneedle,^{[ 192 ]} microwave,^{[ 193 ]} and near‐infrared spectrum.^{[ 194 ]} However, due to the limitation of the sensor performance in terms of stability and accuracy, they have only been demonstrated as prototypes and have not yet been commercialized. ML‐based prediction and calibration methods can be used to improve the accuracy of these newly developed methods, including a supervised learning‐based regression approach for sweat glucose determination and calibration of near‐infrared^{[ 195 ]} and microwave^{[ 196 ]} approaches.

Figure 11.

Intelligent biochemical sensors. a) A microneedle‐based minimally invasive metabolites sensor. Reproduce with permission.^{[ 192 ]} Copyright 2022, Springer Nature. b) A wearable RF‐based non‐invasive glucose sensor. Reproduce with permission.^{[ 213 ]} Copyright 2021, American Association for the Advancement of Science. c) A microneedle patch with glucose‐responsive hydrogels for smart delivery of glucagon and insulin. Reproduce with permission.^{[ 200d ]} Copyright 2022, American Association for the Advancement of Science. d) KNN‐based electrochemical sensors for the detection of tyrosine and tryptophane. Reproduce with permission.^{[ 201 ]} Copyright 2024, Elsevier. e) A flexible microfluidic sensor for the detection of sweat amount and ions based on smartphone pictures and ML methods. Reproduce with permission.^{[ 206 ]} Copyright 2022, Wiley‐VCH. f) An intelligent microfluidic sensor for monitoring wound status. Reproduce with permission.^{[ 208 ]} Copyright 2022, American Association for the Advancement of Science.

Further integration with automatic insulin pumps enables the precise injection of insulin and closed‐loop management. Strategies employing algorithms such as artificial neural networks,^{[ 197 ]} fuzzy logic controller,^{[ 198 ]} and personalized model predictive control,^{[ 199 ]} have been proposed to improve the injection safety and efficiency for type I diabetes. Another strategy for closed‐loop management is to utilize glucose‐responsive materials, which are typically made of Phenylboronic acid‐based hydrogels and lysine‐modified insulins as recognition elements. Once the glucose concentration is above a pre‐set value, the hydrogel matrix swells or collapses due to either the change in pH or chemical properties. Consequently, active molecules, such as insulin and glucagon embedded in the matrix, are released. These hydrogel materials can be integrated with microneedles to further reduce the invasiveness of delivery, thereby improving patient experience and compliance (Figure 11c).^{[ 200 ]}

Metabolites and nutrients other than glucose also play a crucial role in health status and disease monitoring. Quantifying these biomarkers have implications for various abnormal conditions, such as infections, fatigue, pre‐diabetes, and stress.^{[ 7b ]} Except for some metabolites that have corresponding enzymes (majorly oxidases), many analytes lack efficient recognition elements that selectively convert concentration information to electrical/optical signals. Electrochemical‐active species, such as uric acid, ascorbic acid, and tyrosine, can be detected through voltammetry techniques but suffer from peak overlapping and large noise. ML‐based characterization methods can help to distinguish and extract useful information. For example, K‐NN has been used to separate overlapping differential pulse voltammetry peaks of tyrosine and tryptophan and correct the influence of pH variation in sweat sensors (Figure 11d).^{[ 201 ]} Electrochemical inactive species, such as essential amino acids, hormones, and even proteins, require specially designed recognition elements, including aptamers, molecular imprint polymers (MIP), and antibodies. Through electrochemical‐active reporters (Ferrocene, Methylene Blue, Prussian Blue) that are attached or incorporated in these elements, electrical signals are generated to reflect concentration information. Nevertheless, this type of biochemical sensor sometimes suffers from nonlinearity of response and low selectivity.^{[ 202 ]} ML‐based methods can offer more precise calibration and guide the design of the recognition elements (mostly aptamers and MIPs). For instance, an ANN model has been integrated with the MIP‐based palmitic acid sensor to non‐linearly fit the responsive curve.^{[ 203 ]} Additionally, neural network‐guided approaches can significantly improve the aptamer binding affinity toward the target molecules and reduce their length.^{[ 204 ]}

Ions and electrolytes, such as sodium, potassium, ammonium, and hydrogen ions (pH), are also crucial health parameters. Corresponding sensors mainly utilize ion‐responsive or selective materials to generate optical or potential signals based on changes in analyte concentration. Colorimetric‐based ion sensors, which rely on changes in color type or intensity, can be integrated with ML to improve accuracy. ML‐based imaging processing algorithms enable fast and accurate data extraction.^{[ 205 ]} For instance, a computer vision‐assisted app^{[ 206 ]} can evaluate the traveling distance and color intensity of microfluidic sensor for fast and accurate sweat rate and chloride ion detection.^{[ 205 ]} With only a smartphone captured photos at ambient light conditions, sweat rate, and chloride ion concentrations are recorded with accuracy comparable to manual measurements (Figure 11e). Such method has also been extended to the analysis of tear biomarkers using multichannel CNN‐GRU to build a cloud server platform.^{[ 207 ]} A similar approach has been adopted to develop colorimetric wound sensors that are capable of measuring pH, temperature, and other wound‐related biomarkers (Figure 11f).^{[ 208 ]} CNN algorithms assist in analyte concentration determination and improve the classification accuracy of wound status to 97%. Potential signal‐based sensors, which directly output electrical voltage for further analysis, are much easier to couple with the data acquisition model and perform ML‐based signal processing. For instance, a deep artificial neural network (deep ANN) has been utilized to correlate voltage responses of a contactless pH sensor with inflammation and regeneration condition of wounds.^{[ 209 ]}

Drugs and other biomarkers, such as interleukins, proteins, and infectious species, are also crucial in clinical practice for determining therapeutic windows and disease progressions. The capabilities of wearable sensors for detecting inflammation factors in wound exudates,^{[ 210 ]} C‐reactive proteins,^{[ 186 ]} and paroxetine^{[ 211 ]} in sweat, tobramycin, and vancomycin in interstitial fluids^{[ 212 ]} have been extensively demonstrated. Although few have adopted ML as a prediction or calibration method, their data is very crucial for personalized medicine and therapeutic plans, holding great promise to integrate with ML along with other physiological parameters to enable intelligent healthcare.

4.3. Other Sensors

Although most current externally located wearable sensors focus on obtaining a single specific physiological signal, there are advanced wearable sensors that integrate multiple types of sensors into one device.^{[ 140} , ²¹⁰ , ^{214 ]} This integration allows for comprehensive monitoring of a person's health or a particular disease, significantly reducing the inconvenience of wearing multiple sensors. Additionally, it enhances the reliability of each sensor through cross‐validation or multimodal calibration. Typically, a combination of physical and chemical sensors is employed to provide a holistic view of the condition being monitored. For instance, a wound immunosensor patch can measure inflammation‐related proteins (such as tumor necrosis factor‐alpha and interleukins), bacterial levels, and pH to assess inflammation and infection. Simultaneously, physical sensors, like temperature sensors, offer supplementary data to detect potential infections (Figure 12a).^{[ 210 ]} Another example is a system designed to monitor stress responses by detecting physiological signals (pulse waveform, galvanic skin response, and skin temperature), sweat metabolites (glucose, lactate, and uric acid), and electrolytes (sodium, potassium, and ammonium) (Figure 12b).^{[ 214c ]} Additionally, a device that simultaneously measures EEG signals and lactate in sweat has been developed to explore the relationship between brain activity and lactate levels, with potential applications in understanding traumatic brain injuries and daily stress fluctuations (Figure 12c).^{[ 140c ]}

Figure 12.

Other Intelligent wearable sensors. a) A flexible microfluidic multiplexed immunosensor for point‐of‐care quantitative assessment of wound biomarkers. Reproduce with permission.^{[ 210 ]} Copyright 2021, American Association for the Advancement of Science. b) An AI‐enhanced electronic skin with robust, long‐term sensing for stress response monitoring. Reproduce with permission.^{[ 214c ]} Copyright 2024, Springer Nature. c) A fully in‐ear integrated array of multimodal sensors for simultaneous monitoring of brain state and dynamic sweat metabolism. Reproduce with permission.^{[ 140c ]} Copyright 2023, Springer Nature. d) A conformal, stretchable ultrasonic device for non‐invasive, accurate, and continuous monitoring of vital signs from beneath the skin. Reproduce with permission.^{[ 216b ]} Copyright 2018, Springer Nature. e) A wearable sensor for dual epidermal fluid sampling and detection. Reproduce with permission.^{[ 215c ]} Copyright 2018, Wiley. f) A wearable magnetic field sensor with a low detection limit of 22 nT and a wide operation range from 22 nT to 400 mT. Reproduce with permission.^{[ 229 ]} Copyright 2023, Wiley. g) A flexible, battery‐less, near‐field‐enabled sensing tag integrated into an FFP2 facemask for gaseous CO2 detection. Reproduce with permission.^{[ 218 ]} Copyright 2022, Springer Nature.

Most wearable sensors currently operate through passive detection of biochemical or electrical signals to assess health status. However, sometimes the detection can only be conducted in specific scenarios, such as when sweat analysis based on wearable sensors is required during physical activity. To overcome these limitations, some sensors now incorporate active stimulation sources. For example, in sweat analysis, electrical stimulation can trigger sweat production, enabling immediate response when abnormal sweat composition is detected and allowing for regular data collection (Figure 12d).^{[ 215 ]} This approach minimizes data loss or degradation due to external factors. Another example is the integration of ultrasound generation devices into wearable sensors (Figure 12e).^{[ 216 ]} These sensors, with mechanical properties similar to the skin and an ultra‐thin profile, maintain conformal contact with the skin's dynamic curves over time. They enable continuous monitoring of deep blood vessels' central blood pressure (CBP) without the operational difficulties or instability associated with traditional methods. This non‐invasive, continuous, and precise monitoring of deep biological tissues or organs opens new opportunities for diagnosing and predicting cardiovascular diseases using wearable devices.

Some sensors are specifically designed to monitor the surrounding environment, providing personalized health recommendations based on the individual's living conditions. For instance, Li et al. reported a flexible magnetic field sensor with a wide detection range from 22 nT to 400 mT and a low detection limit of 22 nT^{[ 217 ]} (Figure 12f). This intelligent magnetic‐sensitive wearable device has potential applications in geomagnetic navigation, interactive human‐machine interfaces for entertainment and rehabilitation, and safety warning systems for exposure to strong magnetic fields. Another example, prompted by the widespread use of face masks during the COVID‐19 pandemic, is the smart facemask for wireless carbon dioxide (CO₂) monitoring reported by P. Escobedo et al., designed to monitor potential adverse effects from prolonged mask usage (Figure 12g).^{[ 218 ]}

Accurately measuring physiological parameters can be challenging when relying on externally located sensors. In clinical practice, examples include recording pressure inside target organs^{[ 219 ]} or monitoring organ movement.^{[ 220 ]} Clinically relevant pressure recordings for intracardiac, intracranial, or urinary bladder pressure often require invasive procedures, limiting their accessibility. Traditionally, such recordings are conducted in a hospital setting with immobile patients for short periods. The setup involves water‐filled tubes with external transducers or optical cables, posing technical limitations.^{[ 221 ]} Ongoing developments in sensor and miniaturization techniques contribute to the creation of smaller, more efficient, and biocompatible implantable sensors. These technological advancements open up new possibilities for healthcare applications, allowing for short‐ or long‐term implantation of sensors.^{[ 219b ]} Designs such as sensor‐array integrated catheters^{[ 222 ]} and passive arterial‐pulse sensors^{[ 220} , ^{223 ]} have emerged to achieve an internal signal collection that is both accurate and comprehensive.

Options include using a wire to penetrate the skin for power supply and external data logging or fully implanting sensors and electronics, with or without a power source. Fiber‐based sensors have recently attracted attention owing to their flexibility and capabilities to integrate with existing surgical procedures. Flexible fiber‐based sensors can serve as alternatives to subcutaneous implantable sensors that monitor physiological parameters on a long‐term basis.^{[ 224 ]} On the other hand, suture‐like sensors^{[ 225 ]} can monitor the progression of surgical wound healing, leakage, and potential infections. Additionally, patch sensors can also be implanted for wearable wireless monitoring.^{[ 226 ]} For example, early rejection of a transplanted kidney can be detected by continuously and in real‐time monitoring the long‐term local temperature and thermal conductivity.^{[ 172 ]} This monitoring allows for the detection of inflammatory processes associated with graft rejection.

Speaking of communication and data transmission strategies, NFC could eliminate the need for an implanted battery, but signal strength limitations arise with shallow antenna implants. Various communication protocols, such as Bluetooth or Medical Implant Communication System (MICS), are available for implants with power sources but their transmission range is still limited. External power sources (batteries) are still required for implantable sensors and limit their lifetime. Although current techniques such as wireless powering and self‐powered techniques (e.g., triboelectric nanogenerators^{[ 227 ]} and biofluids‐based batteries)^{[ 228 ]} have been proposed, their applications to implantable devices remain constrained by the complexity of human internal environments.

5. Biomedical Applications Enabled by ML and Smart Materials‐Based Wearable Sensors

Smart materials and ML‐enhanced wearable sensors can cater to a wide range of users, including healthy individuals, at‐risk or chronic patients, and healthcare professionals offering specialized services in rehabilitation centers and public health organizations. They enable the early detection of disease risks, continuous biomarker monitoring, and a deeper understanding of individual health conditions, leading to more precise and personalized prevention and treatment strategies.^{[ 230 ]} Wearable sensors, in particular, are instrumental in providing smart preventive healthcare,^{[ 231 ]} improved chronic disease management,^{[ 232 ]} smart medicine,^{[ 233 ]} and intelligent rehabilitation,^{[ 234 ]} significantly improving the delivery of care and enabling real‐time, patient‐centered interventions that are tailored to individual needs.

5.1. Smart Preventive Healthcare

Wearable sensors can also foster a shift in healthcare from reactive to proactive paradigms.^{[ 132} , ²³¹ , ^{235 ]} Wearable sensors can seamlessly and simultaneously track diverse vital physiological parameters, including heart rate, activity levels, sleep patterns, BP, and SpO₂, as well as biochemical markers such as glucose and cortisol levels.^{[ 236 ]} These data can be incorporated with AI algorithms to detect subtle deviations from healthy patterns, identify health anomalies, and recognize early indicators of diseases even before symptoms manifest. The ML models, which are trained on vast datasets, can also evolve during the adaptive learning process, enabling the tailoring of detection capabilities for each individual. Upon the detection of anomalies, wearable devices can immediately alert users, doctors, or caregivers. This will enable timely medical intervention, potentially avoiding disease progression and improving long‐term treatment outcomes.

Wearable sensors can also be integrated with AI to help healthy individuals or patients monitor their health status and provide suggestions and nudges to adopt a healthier lifestyle. For instance, an elastoplastic silk fibroin electrode‐based bioelectronic device, functioning as a wearable personal trainer, was developed to transmit sensory information that can warn users of risky postures, helping to prevent severe muscular injuries, particularly for amateurs.^{[ 114a ]} The unsupervised algorithm of PCA is utilized to reduce the dimensionality of waveforms and statistically evaluate triboelectric outputs during workouts. As illustrated in Figure 13a, wavelet packet transform is employed to classify proper and risky postures, allowing athletes to assess their posture accuracy, avoid injuries, and receive timely warnings to inhibit risky movements.^{[ 114a ]} Furthermore, a wearable system was developed that integrates a conformal, flexible strain sensor array containing six sensors (A to F) placed on the wrist, working alongside a deep learning model for detecting early warning signs of hypertension or cardiac anomalies, allowing for proactive intervention (Figure 13b). The system employs self‐developed preprocessing algorithms for denoising, debasing, segmentation, and classification, ensuring high‐quality pulse signals without motion artifacts. The deep learning module, comprising four convolutional layers, three max‐pooling layers, and two fully connected layers, automates feature extraction and delivers precise, timely outputs of BP and cardiac parameters.^{[ 128 ]} Other than that, people with busy lifestyles or working under high stress or pressure can use wearable sensors (e.g., cortisol sensors)^{[ 237 ]} to continuously monitor their own bodily stress levels and take actions to reduce them. Individuals with a family history of diabetes can wear glucose sensors to monitor fluctuating glucose levels and adjust their food intake accordingly.^{[ 238 ]}

Figure 13.

Intelligent healthcare enabled by ML and wearable sensors. a) Elastoplastic silk fibroin electrode‐based bioelectronics enable warning the risky postures to inhibit severe muscular injuries for amateurs. Reproduced with permission.^{[ 114a ]} Copyright 2022, Wiley. b) A wearable strain sensor array was developed for monitoring blood pressure and cardiac function via a deep learning algorithm. Reproduced with permission.^{[ 128 ]} Copyright 2023, American Association for the Advancement of Science. c) An intelligent prosthetic arm with a robotic hand and an epidermal VR device on his upper arm for personalized rehabilitation. Reproduced with permission.^{[ 247 ]} Copyright 2019, Springer Nature. d) A haptic glove was demonstrated to deliver a dynamic haptic experience for the rehabilitation of hand activities. Reproduced with permission.^{[ 234b ]} Copyright 2023, Wiley. e) A fully portable continuous real‐time auscultation with a soft wearable stethoscope for automated, objective diagnosis of diseases via ML. Reproduced with permission.^{[ 248 ]} Copyright 2022, American Association for the Advancement of Science. f) A miniaturized, wireless neural probe system with the capability of bi‐directional wireless drug delivery and electrophysiology monitoring in socially interacting mice. Reproduced with permission.^{[ 252 ]} Copyright 2022, Springer Nature.

5.2. Intelligent Rehabilitation and Human‐Machine Interfaces

Wearable sensors, particularly tactile sensors and EMG sensors, can be integrated with ML and incorporated into prosthetic and assistive devices through human‐machine interfaces, enabling a wide range of intelligent rehabilitation applications. These systems are particularly beneficial for aiding individuals recovering from conditions such as stroke, spinal cord injuries, or neuromuscular disorders.^{[ 239 ]}

Generally, tactile sensors can help realize functions of tactile feedback, object identification, motion control enhancement, and sensory restoration, while EMG sensors can facilitate gesture recognition, motion control, and rehabilitation evaluation.^{[ 240 ]} For example, scalable tactile gloves equipped with hundreds of tactile sensors, can capture the intricate signatures of human grasp. These gloves, combined with ML algorithms, enable the identification and weighing of objects with high precision.^{[ 241 ]} Similarly, wearable high‐density EMG biosensing systems have been reported to utilize hyperdimensional computing algorithms for in‐sensor adaptive learning, achieving real‐time hand gesture classification with an accuracy of 97.12%. These systems perform training, inference, and model updates locally and in real time, allowing them to adapt to changing situational context.^{[ 242 ]}

Tactile sensors and EMG sensors can integrate with neural stimulation electrodes to fabricate closed loop systems for motion control enhancement, natural interaction, and sensory restoration. For instance, the integration of wearable tactile sensors and EMG sensors with stimulators and soft pneumatic actuators can lead to the development of soft neuroprosthetic hands, which exhibits adaptive grasping, high operational speed, dexterity, and natural interaction.^{[ 243 ]} In such closed‐loop systems, EMG sensors control pneumatic actuation for precise movement, while tactile sensors provide real‐time tactile feedback through electrical stimulation, enhancing user control and interaction. Furthermore, the combination of tactile sensors with implanted neural stimulation electrodes also enable the creation of a bidirectional brain‐computer interface (BCI) that significantly improves robotic limb control, help restore sensations and aid in rehabilitation for individuals with limb loss or sensory deficits.^{[ 244 ]} This system employs tactile sensor arrays on a robotic hand to supplement vision‐based grasping with tactile feedback, while simultaneously recording neural activity from the motor cortex to control the prosthesis. Through intracortical microstimulation of the somatosensory cortex, tactile sensations can also be generated, resulting in improved task performance for individuals with tetraplegia, bringing their abilities closer to those of able‐bodied individuals. Similarly, a closed‐loop hand prosthesis has been developed by combining tactile sensors, position sensors, EMG sensors, and implanted nerve stimulation for somatotopic tactile feedback and remapped proprioception through sensory substitution, restoring pressure perception and providing richer, multimodal sensations.^{[ 245 ]} Additionally, neuromorphic e‐skin has shown great promise in restoring the sense of touch.^{[ 246 ]} For instance, a prosthesis with neuromorphic multilayered e‐dermis and transcutaneous nerve stimulation can perceive touch and pain, closely mimicking the natural sensory functions of human skin. This technology not only enables prosthetic limbs to detect tactile stimuli but also allows them to respond to potentially harmful stimuli, such as pain, thereby enhancing both safety and functionality for users.^{[ 246b ]}

Furthermore, wearable sensors integrated with VR or augmented reality (AR) enable immersive virtual rehabilitation with remote guidance, enhancing patient adherence and creating a motivating therapeutic environment. For example, a wireless, battery‐free platform of electronic systems based on haptic interfaces was developed, capable of softly conforming to the curved surfaces of the skin and communicating information through spatiotemporally programmable patterns of localized mechanical vibrations.^{[ 247 ]} A key application of this technology is providing tactile feedback for robotic prosthetic devices. As shown in Figure 13c, a man with a lower arm amputation uses a prosthetic arm with an epidermal VR device on his residual limb. Sensors on the prosthetic detect object shapes, transmitting data to the VR device to create a virtual haptic representation on his upper arm, enhancing interaction.^{[ 247 ]} Similarly, a haptic glove was demonstrated to deliver variable stiffness force feedback, along with fingertip force and vibration feedback (Figure 13d). This functionality enables users to interact with virtual environments by touching, pressing, grasping, squeezing, and pulling various objects, offering a dynamic haptic experience for the rehabilitation of hand activities.^{[ 234b ]}

5.3. Intelligent Disease Diagnosis

Wearable sensor technology has the potential to assist individuals with chronic diseases (e.g., diabetes, chronic wounds) in self‐managing of their conditions.^{[ 238 ]} Wearable sensors offer a paradigm shift by providing continuous, real‐time monitoring for individuals, offering a stark contrast to the episodic and constrained nature of hospital‐based assessments.^{[ 233b–d ]} This can help patients to conveniently self‐manage the diseases^{[ 238 ]} even while at work or home, thereby contributing to empowerment, independence, and thus improved quality of life. For example, for the diagnosis and management of patients with chronic obstructive pulmonary disease and cardiovascular disease, a soft wearable stethoscope system has been developed for continuous cardiopulmonary monitoring, allowing real‐time cardiovascular and respiratory assessment during daily activities. This system accurately collects cardiorespiratory data, enabling the diagnosis of various pulmonary abnormalities. Utilizing a wavelet denoising algorithm, the soft wearable stethoscope shows an enhanced SNR, improving the detection of lung diseases. With further integration of deep learning algorithms, the system can automatically detect and diagnose four types of lung diseases—crackle, wheeze, stridor, and rhonchi—with ≈95% accuracy across five classes. A user‐friendly mobile application accompanies the device, enabling heart and lung sound recordings, real‐time signal display, automatic diagnosis of abnormal lung sounds, and secure data upload to synchronized local storage (Figure 13e).^{[ 248 ]}

Furthermore, patients with hypertension and glaucoma can wear BP sensors and wearable tonometers to continuously detect BP and intraocular pressure with a reminder to take their medication.^{[ 232a ]} Wearable sensors play a crucial role in improving healthcare for the disabled or the elderly by facilitating seamless contact with doctors, family members, and caregivers. Wearable sensors can provide timely alarms or reminders, ensuring proactive health management. For example, integrated wearable sensors can enable monitoring of physiological signs (e.g., heart rate^{[ 232a ]} and respiratory rate,^{[ 126 ]} sleep‐related signals),^{[ 249 ]} which can be used for patients with conditions such as congestive heart failure or chronic obstructive pulmonary disease. With user‐friendly interfaces and built‐in alarm functions, such systems can provide timely feedback to caregivers managing patients with limited self‐care capabilities.^{[ 249 ]}

5.4. Smart Medicine

Wearable sensors can intricately decipher in situ personal datasets such as glucose level, temperature, pH, respiration rate, and heart rate.^{[ 250 ]} Based on the dynamic health status, leveraging advanced analytics and ML, data‐driven insights, more precise and accurate therapeutic dosage can be predicted. For example, wearable continuous glucose sensors can leverage ML to predict insulin bolus requirements and execute self‐calibration and correction for insulin dosage recommendations for patients with diabetes.^{[ 232b ]} Based on wearable sensors and ML, efficient drugs can be screened by real‐time monitoring of their pharmacokinetics. Traditional investigations on pharmacokinetics to assess the therapeutic effects of drugs often involve intermittent invasive blood draws. However, wearable biochemical sensors can monitor in real‐time the metabolism and concentration of drugs, such as vancomycin,^{[ 233 ]} methylxanthine,^{[ 184 ]} and doxorubicin^{[ 251 ]} in different organs, through analyzing diverse biofluids such as sweat, saliva, tears, or interstitial fluid. For instance, a miniaturized wireless neural probe system was developed to deliver drugs and simultaneously monitor their real‐time effects on both neural activity and behavior in socially interacting mice. Equipped with a bi‐directional wireless communication module, the system allows precise control over drug infusion while transmitting neural signals from multiple animals as their behavior is observed (Figure 13f). The recorded neural data is analyzed using PCA and k‐means clustering to detect neural spikes from individual neurons, providing insights into how the drugs influence brain activity and behavior. This system represents a significant advancement in smart medicine, offering new possibilities for personalized drug administration and neural monitoring for improved therapeutic outcomes.^{[ 252 ]}

6. Challenges and Future Prospectives

Though significant advances have been made, many challenges and limitations in intelligent wearable sensing systems empowered by smart materials and ML still need to be addressed to realize their full potential. For successful adoption by individuals in both the wellness and healthcare sectors (e.g., healthcare professionals and patients), emphasis must remain on usability design, friendly interface, data integrity, clinical readiness, alignment with clinical needs, affordability, and integration with existing medical systems. Key considerations include addressing data privacy and security concerns, overcoming technical limitations of wearable sensors such as accuracy and reliability, meeting user demands, fostering motivation and compliance, and navigating regulatory and ethical challenges. These factors must be carefully managed to ensure effective implementation.

1. Data integrity: Ensuring the integrity of the data provided by intelligent wearable sensors involves overcoming several challenges.^{[ 253 ]} Calibration errors can occur during manufacturing or user setup, leading to systematically skewed data. User errors, such as incorrectly positioning the sensor or failing to maintain the device, can also introduce variability. Additionally, environmental conditions such as temperature, humidity, or even electromagnetic interference can affect sensor readings. Noise introduced in any of these ways reduces the reliability of the data, making it less useful or even hazardous for clinical applications. Factors such as sensor placement and user variability can affect data accuracy.^{[ 254 ]} For example, inaccuracies in heart rate measurements obtained through PPG may arise from factors such as different skin types, motion artifacts, and signal crossover.^{[ 255 ]} Additionally, ML models are susceptible to bias if trained on inaccurate, incomplete, or unrepresentative data sets.^{[ 256 ]} For example, if a model is trained primarily on data from younger individuals, but then applied to an older population, the likelihood of erroneous predictions increases substantially. Similarly, a lack of diversity in training data on ethnicity, gender, or other demographic variables can lead to biased results that may not be universally applicable. These biases can have serious implications that can lead to a misdiagnosis or ineffective treatment plan. Therefore, performance calibration, reliability validation, user education, and diverse models are crucial to ensure data integrity.

2. Demand‐driven innovation: A paradigm shift is urgently needed in the development of wearable sensors, prioritizing performance and application requirements before engineering materials to meet these specific needs. Unlike traditional practice that first develop materials and then seek suitable applications, this demand‐driven approach is crucial for advancing and adopting wearable devices. For instance, self‐healing materials should first be designed to rapidly and reliably repair mechanical damage under the specific conditions in which the device operates,^{[ 257 ]} while metamaterials need to be modified for specific modalities like strain or temperature.^{[ 45} , ⁵⁴ , ^{258 ]} Likewise, responsive sensing materials must be tailored for distinct sensing modalities. Although there is progress in enzyme‐based glucose electrodes to realize long‐term applications, the effective management of diabetes is still beyond satisfactory. Phenylboronic acid‐based sensors have been developed to offer an alternative to enzyme‐based glucose sensors due to their high operational stability, resistance to biofouling, and adaptability to a wide glucose concentration and pH range, making them suitable for long‐term continuous glucose monitoring.^{[ 259 ]}

3. Integration of smart materials and ML: Although the integration of advanced smart materials into wearable devices presents exciting opportunities, there are still significant challenges. For example, in continuous glucose monitoring for diabetes management, the responsive sensing materials that react to glucose change in biofluids can enable non‐invasive, real‐time glucose monitoring.^{[ 260 ]} However, due to exposure to dynamic physiological environments, the stability of the sensing materials may be affected in some severe conditions, such as fluctuating pH levels, varying temperatures, or mechanical stress from body movement.^{[ 261 ]} These factors can lead to material failure, reduced sensitivity or inaccurate readings, compromising the accuracy and reliability of glucose monitoring. When integrated with ML, the potential of wearable devices leveraging smart materials can be further enhanced. For example, ML algorithms can improve the reliability of glucose monitoring by compensating for drift in the signal acquired by wearable devices.^{[ 262 ]} However, integrating ML into wearable devices presents significant challenges, such as developing ML models that are adaptable to diverse populations.^{[ 263 ]} A promising future direction for the combination of ML and wearable devices lies in personalized healthcare, which aims to deliver individualized medical treatment by accounting for a person's unique health profile, shaped by genetics, lifestyle, and environmental factors. Wearable sensors collect critical data on these factors, while ML algorithms analyze the data to reveal patterns and insights that may not be immediately apparent to human analysts. Together, wearable sensors and ML empower healthcare providers to design tailored treatment plans, advancing the potential of personalized care.^{[ 264 ]}

4. Performance: Key performance indicators, such as sensitivity, sensing range, linearity, response time, and hysteresis, are the primary focus of sensors. However, these properties are often interrelated, meaning that improving one can limit another—for instance, enhancing sensitivity may reduce sensing range while reducing hysteresis can affect softness.^{[ 265 ]} Hence, it is critical to balance the optimization of specific properties with the practical requirements of applications.

5. Power sources and sustainability: Many wearable devices have limited battery life, which requires frequent charging or battery replacement. This can disrupt continuous monitoring, particularly for long‐term applications.^{[ 10a ]} While significant progress has been made in self‐powered electronics that integrate energy harvesting, the energy generated often falls short of meeting the power demands of most devices.^{[ 266 ]} As a result, most wearable devices rely on bulky and rigid batteries as their primary power source, due to their superior stability and sufficient energy output.^{[ 267 ]} Although soft batteries and supercapacitors have been developed to replace rigid ones, they often suffer from low power and energy density due to compromises between electrochemical and mechanical properties. Beyond focusing on improving power sources, another strategy is to develop energy‐efficient sensors and processing circuits that can collect data without excessive power consumption.^{[ 268 ]} Furthermore, the environmental impact of the production and disposal of these devices is a concern that is increasingly being discussed. As a result, there remains significant room for improvement in making wearable devices more energy‐efficient, sustainable, and environmentally friendly.

6. Packaging issues: As wearable devices become increasingly sophisticated, researchers face the challenge of integrating multiple functions into a compact and efficient design. Maintaining a small footprint is crucial to ensuring that these devices remain unobtrusive and comfortable for users.^{[ 269 ]} However, miniaturization must be achieved without compromising sensor performance or battery life.^{[ 270 ]} This necessitates careful optimization of power‐efficient components and the development of effective energy management strategies. Moreover, wearable devices must be robust enough to function reliably in diverse conditions, including exposure to moisture, temperature fluctuations, and mechanical deformation.^{[ 271 ]} The packaging materials and design must also be biocompatible, durable, lightweight, aesthetically pleasing, and cost‐effective.^{[ 272 ]} Therefore, continuous innovation in packaging technologies is essential for advancing wearable sensor systems.

7. User compliance and behavior: The adherence of users to the protocols of wearable devices significantly impacts the integrity and utility of the data collected. For wearable sensors to function optimally, consistent usage and proper maintenance by the users are imperative. Inconsistencies such as neglecting to wear the device, failing to charge it, or omitting to synchronize it can lead to significant data gaps.^{[ 273 ]} Moreover, improper usage by some users may compromise the precision of the data. Consequently, ensuring sustained user engagement in long‐term monitoring presents a considerable challenge. To mitigate this, the development of an intuitive user interface, coupled with comprehensive initial training, is crucial to enhance user compliance and facilitate the adoption of the technology. Additionally, the implementation of pedagogical strategies and motivational incentives is vital for fostering user engagement and adherence. The level of user compliance may also be influenced by the intended application of the collected data, such as whether it is used to directly inform and modify individual or patient behavior, or whether it is collected without direct interaction with the sensor user.

8. Security, social, and ethical issues: Except for the technical limitations, the implementation of wearable sensors also faces social and ethical issues. One major issue is data privacy and security, as these devices collect vast amounts of sensitive health information, making them vulnerable to hacking and unauthorized access.^{[ 274 ]} Compliance with regulations like the EU GDPR and HIPAA adds complexity, especially in different jurisdictions. Another challenge is user adoption and compliance. The high cost of wearable sensors can limit access, and inconsistent use by users, such as forgetting to wear or charge the devices, can lead to data gaps.^{[ 8 ]} Additionally, regulatory hurdles, such as meeting FDA and EU MDR standards, are necessary for wearable sensors to be approved as medical devices.^{[ 275 ]} Ethical concerns also arise, particularly around data ownership, transparency, and the potential for wearable technology to widen healthcare disparities.^{[ 276 ]} Moreover, wearable sensors and intelligent healthcare have the potential to redefine both health insurance and public health policies. Conversely, health insurance and public health policies can strategically promote the adoption of wearable devices to encourage individuals to maintain a healthier lifestyle.

9. Implementation: An urgent challenge in wearable sensors is achieving efficient transmission of large amounts of sensor data while enabling power‐efficient in‐sensor ML with local computing. Conventional low‐power multichannel microcontrollers (MCUs), such as ESP32 and STM32, lack the necessary computational capacity to support deep learning, transfer learning, and other advanced ML models, limiting their ability to perform real‐time, on‐device processing. As a result, data preprocessing and ML algorithms in wearable sensors are typically restricted to offline analysis, preventing real‐time decision‐making.^{[ 113} , ^{277 ]} To overcome this limitation, edge computing with dedicated ML chips has emerged as a promising solution for high‐accuracy motion monitoring and real‐time inference. A novel integration approach—combining wearable sensors with edge‐computing chips—has been developed to enable a full‐body motion monitoring system that leverages edge data processing and in‐sensor ML for full‐body motion classifications and avatar reconstruction.^{[ 278 ]}

While these challenges and limitations present significant obstacles, they also open avenues for ongoing research, innovation, and collaboration. As technology advances and stakeholders collaborate to tackle these issues, intelligent wearable sensors hold immense potential to revolutionize healthcare, paving the way for more personalized, proactive approaches to disease management and improved overall health.

Conflict of Interest

The authors declare no competing interests.

Author Contributions

C.T.L, V.H.T.H., and A.C.T.M. initiated this paper based upon institutional collaboration and meetings in Singapore and Norway in 2023. C.T.L., M.A.R. conceived the outline, wrote and edited the contents. S.W.C., S.C.F., Z.Q., Z.X.W., B.B. L., Z.J.L., and C.T.L. wrote and revised the paper. S.W.C., S.C.F. Z.Q., and Z.X.W., contribute equally to the work. M.A.R., M.Y.H.W., A.O., O.K., K.M.N., T.G., A.C.T.M., and V.H.T.H. contributed text and text revisions to various versions of the manuscript. All authors approved the final manuscript.

Biographies

Shuwen Chen obtained her Ph.D. in 2019 from the Wuhan National Laboratory for Optoelectronics at Huazhong University of Science and Technology (HUST). From 2020 to 2024, she was a Research Fellow in the laboratory of Professor Chwee Teck Lim at the National University of Singapore. She is currently an Associate Professor at the Institute of Medical Equipment Science and Engineering (IMESE), HUST. Her research focuses on wearable flexible sensors for healthcare applications.

Vibeke H. Telle‐Hansen is professor in Nutrition at Oslo Metropolitan University (OsloMet), Norway. Prof Telle‐Hansen's research focuses on individual response and precision nutrition in the prevention of non‐communicable diseases. She has extensive experience in conducting human randomized controlled trials, using high‐throughput technologies (metabolomics and transcriptomics) and wearable sensors to collect real‐world data. Since 2022, Prof Telle‐Hansen has academic responsibility for Intelligent Health at OsloMet, an interdisciplinary, strategic initiative promoting research, innovation, and education within health and technology.

Lim Chwee Teck is the NUSS Chair Professor and Director of the Institute for Health Innovation and Technology at the National University of Singapore. He has coauthored over 500 journal publications and cofounded six startups with one public listed in 2018. Prof Lim is an elected Fellow of nine academies and organizations including the Royal Society UK, US National Academy of Inventors, IUPESM, AIMBE, IAMBE, Singapore National Academy of Science, and the Academy of Engineering Singapore. He has garnered numerous research awards including the Nature Lifetime Achievement Award for Mentoring in Science, Highly Cited Researcher, Asia's Most Influential Scientist Award, Wall Street Journal Asian Innovation Award (Gold), and the President's Technology Award.

Chen S., Fan S., Qiao Z., Wu Z., Lin B., Li Z., Riegler M. A., Wong M. Y. H., Opheim A., Korostynska O., Nielsen K. M., Glott T., Martinsen A. C. T., Telle‐Hansen V. H., Lim C. T., Transforming Healthcare: Intelligent Wearable Sensors Empowered by Smart Materials and Artificial Intelligence. Adv. Mater. 2025, 37, 2500412. 10.1002/adma.202500412