Recent advances in skin-interfaced pressure sensors and time series classification of their signals

Abstract

Recent advances in machine learning have seen a wide range of applications across many fields. When combined with developments in flexible, skin-interfaced pressure sensors, these technologies are driving a new generation of personalized health monitoring. From preventing diabetic foot ulcers to tracking respiratory rate and other vital signs, these systems are advancing smarter and more responsive healthcare solutions. This review presents a comprehensive overview of the latest developments in skin-interfaced flexible pressure sensing, starting from their physical mechanisms and ending at microscale material structure. In addition, advanced machine learning approaches for sensor data processing and interpretation is explored, ranging from the fundamental concepts to more recent deep learning models such as temporal convolutional networks for time series classification. Moreover, a systematic review of recent literature is presented, highlighting the application of machine learning in analyzing signals from flexible pressure sensors. Emerging applications leveraging machine learning techniques to facilitate smart health monitoring and human-machine interfaces are explored. A concluding section outlines the challenges and outlook for these emerging technologies as it relates to the biomedical field. To the best of our knowledge, this is the first review which evaluates the potential integration between skin-interfaced flexible pressure sensors with cutting-edge machine learning models, offering a synergistic perspective on next-generation biomedical applications.

Introduction

The convergence of advanced materials engineering and state-of-the-art machine learning models has catalyzed the development of advanced skin-interfaced pressure sensor systems. These systems, designed to be flexible and conform to the contours of the human body, leverage specific pressure sensing mechanisms and engineered micro/nano structures to achieve specific sensor characteristics (Nie et al., 2024). Particularly sought after sensor characteristics typically include high sensitivity, low hysteresis, good stability, and excellent mechanical compliance (Barhoum and Altintas, 2023; Huang et al., 2019). These capabilities are particularly relevant for biomedical applications where non-invasive monitoring of physiological or biomechanical signals is required. Pressure sensing mechanisms, ranging from piezoresistive, piezoelectric, and capacitive to more recently explored approaches such as triboelectric and iontronic, have demonstrated many use cases including cardiovascular monitoring (Chowdhury et al., 2023), gait analysis (Li et al., 2024; Beigh et al., 2023), and posture recognition (Ma et al., 2024; Ran et al., 2021).

In parallel, breakthroughs within the field of machine learning have had meaningful impacts on other fields. In particular advancements in time series classification (TSC) methods have become a powerful tool for extracting meaningful patterns from time series sensor data. Traditional machine learning algorithms continue to see wide use; however, recent progress in deep learning and ensemble-based methods (e.g. InceptionTime, ROCKET, HC2) have dramatically improved performance (Middlehurst et al., 2024). This review aims to synthesize the most recent advancements in skin-interfaced pressure sensors and TSC techniques, providing a comprehensive overview of sensor characterization, sensing mechanisms, engineered materials innovations, and machine learning frameworks for TSC, while highlighting important considerations and challenges of these synergistic systems within the context of biomedical applications.

Skin-interfaced, flexible pressure sensing fundamentals

This section covers the foundational concepts necessary to understand the function and performance of skin-conformable biomedical sensors. It outlines key sensing behaviors and performance metrics, such as sensitivity, response time, detection limit, and stability, that are critical for evaluating sensor functionality in physiological monitoring. It then discusses the primary transduction mechanisms commonly employed in these systems, including piezoresistive, piezoelectric, and triboelectric effects. Finally, the section explores how material selection and structural design at the micro- and nanoscale influence these mechanisms, enabling the development of flexible, high-performance pressure sensors for emerging wearable healthcare applications.

Sensor behavior

Important characteristics for defining the performance of flexible pressure sensors along with their associated deviations are illustrated within Figure 1. These performance metrics are (a) linearity and hysteresis, (b) Stability and drift, (c) Response and recovery time (d) Repeatability, and (e) Sensitivity and dynamic range.

Figure 1.

(a) Linear response and hysteresis, (b) stability and drift, (c) response and recovery time, (d) repeatability, (e) sensitivity and dynamic range.

Sensor performance depends on several interconnected characteristics. Linearity describes the proportionality of the sensor output compared to it’s input over a given range. Within this interval both the maximum and minimum detectable values establish the operating boundaries of the sensor (Barhoum and Altintas, 2023). Over extended use, stability becomes an important metric as it reflects the degree to which a sensor’s output is maintained over time. On the other hand, drift is the gradual divergence of the sensor output despite there being no change in the stimuli being applied to the sensor. High stability is crucial for long-term monitoring particularly for biomedical applications.

Hysteresis poses a challenge in which the sensor’s output not only varies with its current input but also depends on its past inputs (Barhoum and Altintas, 2023). This can be very undesirable in dynamic sensing applications where the stimuli (e.g. pressure change) may vary rapidly, as it introduces uncertainty and unreliability in measurement.

Temporal characteristics such as response time and recovery time dictate how quickly a sensor detects and responds to a change in input signal upon application of a stimulus (Barhoum and Altintas, 2023). This characteristic is important in real-time monitoring applications where delays in signal detection can compromise decision making or system response. Moreover, repeatability quantifies the sensor’s ability to produce identical outputs when under the same conditions across consecutive trials.

Lastly, a sensor’s sensitivity is the rate at which it responds to changes in input signals (Barhoum and Altintas, 2023). This is typically characterized as the slope of the response curve. Sensitivity is directly related to gage factor, a higher gage factor typically implies greater sensitivity, allowing one to detect minute changes in the parameter being measured.

Mechanisms for pressure sensing

Mechanisms for pressure sensing can vary widely depending on the requirements of the incident application. Additionally, the selection parameters used to determine the sensing mechanism depends on a myriad of design factors including but not limited to sensitivity, flexibility, biocompatibility, response time, stability, and manufacturability. Recent advancements in flexible, skin-interfaceable pressure sensors primarily employ one of the following sensing mechanisms: piezoresistive, piezoelectric, triboelectric, capacitive, and iontronic. In this section each of these mechanisms are discussed, outlining key principles and notable materials.

Piezoresistive sensing

The piezoresistive effect constitutes the change in a material’s electrical resistivity when subjected to mechanical deformation. The piezoresistive effect can be explained further by understanding what is occurring on the atomic level within the piezoresistive material. However, it’s important to note that the specific phenomena which occurs depends on the intrinsic material properties and the structures of the material. There are broadly four categories of piezoresistive materials, these include metal conductors, semiconductors, conductive composites and conductive polymers. Typically, metal based (e.g. Platinum, Nickel) and intrinsically semi-conductive materials (e.g. Silicon) are not the first choice as the active material within flexible sensors primary due to their rigidity among several other factors. Instead, approaches using conductive composites or conductive polymers are employed.

Conductive composites

Conductive composites are the most popular choice for flexible piezoresistive sensing. Conductive composites consist of an elastic matrix which encompasses conductive fillers, the structure of the elastic matrix may vary (e.g. aerogels, sponges, films (Camlibel et al., 2023; Duan et al., 2020). Common material choices for the elastic matrix includes numerous polymers outlined in Figure 2. Moreover, common material choices for the conductive fillers can be seen within Figure 3.

Figure 2.

Common elastic polymer materials.

Figure 3.

Common conductive filler materials.

The piezoresistive effect occurs within conductive composites through percolation theory. Prior to deformation, the conductive composite exhibits high resistivity. However, as deformation brings conductive fillers into closer proximity or direct contact, conductive pathways form within the polymer matrix allowing for quantum tunneling effects, increasing electrical conductivity; this point is known as the percolation threshold (Duan et al., 2020; Munson-McGee, 1991). The conductivity of a conductor-insulator near the percolation concentration is expressed using equation (1).

$$\sigma ={\sigma }_0{\left(P-P_c\right)}^t$$

(1)

Where, σ is the conductivity of the system, ${\sigma }_0$ is the conductivity of the filler, P is the concentration of the conductive filler, $P_c$ is the percolation threshold of the conductive filler, and t is a dimensionless index which describes the increase in conductivity once the percolation threshold is reached (Duan et al., 2020).

Conductive polymers

Conductive polymers posses a de-localized π-electron system within their conjugated backbone structure, resulting in a wide band gap and intrinsically conductive molecular structure (Hatchett and Josowicz, 2008). Through a doping process, the conductive properties of these polymers can be further enhanced (Chauhan et al., 2019). Thus, in conductive polymers, the piezoresistive effect manifests due to changes in conductivity resulting from the deformation of the polymer’s conjugated backbone structure. Common conducting polymers used in pressure sensing include poly(3,4-ethylenedioxythiophene):poly-(styrene sulfonate; PEDOT:PSS; Camlibel et al., 2023; Ding et al., 2018), polyaniline (PANI; Camlibel et al., 2023; Ge G et al., 2018), and Polypyrrole (PPy; Camlibel et al., 2023; Pan et al., 2014).

Piezoelectric sensing

The piezoelectric effect occurs due to electrical polarization within a dielectric material when subjected to mechanical deformation. This electrical polarization is caused by the deformation of the materials lattice structure. Inversely, mechanical deformation can be induced using an external electro-magnetic field. The charge polarization induced within the piezoelectric material will generate a voltage across the material, this potential difference can then be measured and related to the amount of external pressure applied to the material (Safari and Akdoğan, 2008). Due to these properties piezoelectric sensing will generates its own voltage potential thus not requiring an external power source (Park et al., 2018; Wang, 2010). This makes piezoelectric particularly useful for lightweight applications. Additionally, piezoelectric sensors are ideal for handling high frequency (high force) signals and have relatively high response rates (Button, 2015). However, piezoelectric sensing is not suitable for long-term static pressure measurements due to charge leakage in the absence of sustained mechanical loading (Sandwell et al., 2020) and is sensitive to large temperature fluctuations (Button, 2015). Moreover, piezoelectric sensing struggles when attempting to detect low frequency (low force) signals due to their response being less sensitive to this type of stimuli (Fu et al., 2025). This makes piezoelectric sensing most suitable for applications which necessitate the detection of rapid force changes such as vibrational or impact sensing (Fu et al., 2025). In the context of biomedical pressure sensing, piezoelectric sensors are well-suited for detecting pressure fluctuations associated with rhythmic physiological activities such as respiration (Mahbub et al., 2017; So et al., 2021) and pulse (Park et al., 2017a), as well as mechanical movements of joints (Li et al., 2023).

Crystals

Typically, active materials used for piezoelectric sensing fall into three classifications: crystals, ceramics, and polymers (Nie et al., 2024). Additionally, hybrid materials which combine these base materials can be fabricated to further enhance piezoelectric performance. Single crystals such as quartz ( ${\mathrm{SiO}}_2$ ) can effectively retain energy within resonant systems over time which is represented by their high mechanical quality factor (Zhang et al., 2018). The piezoelectric coefficient will vary depending on the crystal material. For instance, quartz exhibits a relatively low piezoelectric coefficient ( $d_{33}$ ) of 2.3 pC/N while lithium niobate ( ${\mathrm{LiNbO}}_3$ ) demonstrates much higher values, typically ranging from 300 to 400 pC/N (Momen et al., 2025). However, the primary challenge of using crystals alone is simply that they exhibit poor flexibility and are inherently brittle, making them unsuitable for the dynamic topography of human skin (Wang et al., 2023). Therefore, for the purposes of skin-interfaced sensing, crystals are used as a filler within a polymer matrix to create flexible piezoelectric composite material (Xu et al., 2017a).

Ceramics

Piezoelectric ceramics can be further subdivided into hard and soft ceramics which are dictated by their doping constituents. Hard ceramics are doped using acceptor ions (e.g. ${\mathrm{Fe}}^{3+}$ ) whereas soft ceramics are doped using donor ions (e.g. ${\mathrm{La}}^{3+}$ ). Soft ceramics are the class of piezoelectric ceramics which are relevant for sensor applications, they generally exhibit higher piezoelectric coefficients and electromechanical coupling coefficients than hard ceramics (Chen et al., 2022; Wang et al., 2024a). These properties allow them to excel at sensitivity and signal generation but experience greater energy leakage. Commonly used piezoelectric ceramics, such as Barium Titanate ( ${\mathrm{BaTiO}}_3$ ) and Lead Zirconate Titanate (PZT), exhibit high piezoelectric coefficients ( $d_{33}$ ; ranging from 190 to 220 pC/N (Acosta et al., 2017; Jankowska-Sumara et al., 2017; Renteria et al., 2021) and 200–600 pC/N (Yan et al., 2016a; Yang et al., 2022b) respectively) allowing for extremely efficient conversion of mechanical stress into electrical charge. However, they suffer from similar disadvantages as single crystals, in that they are brittle, rigid, and lack the mechanical flexibility necessary to conform to the dynamic surface of human skin. Despite their shortcomings, piezoelectric ceramics have successfully been integrated into flexible bio-compatible films using polyimide encapsulation methods and advanced fabrication techniques such as electrospinning (Dagdeviren et al., 2017; Fu et al., 2020).

Piezoelectric polymers

Piezoelectric polymers are a favorable material class for flexible pressure sensing due to their high flexibility and bio-compatibility. However, piezoelectric polymers exhibit relatively lower piezoelectric coefficients when compared with commonly used piezoelectric ceramics (Nie et al., 2024; Yang et al., 2020); for reference, polyvinylidene fluoride (PVDF), a commonly used piezoelectric polymer exhibits a piezoelectric coefficient ( $d_{33}$ ) value of approximately 30 (Jain et al., 2015). These polymers derive their piezoelectric properties primarily from mechanisms that either induce a molecular dipole alignment or arise from an inherently non-centrosymmetric molecular structure. For instance, PVDF exhibits crystalline dipole alignment in its β-phase (Gregorio and Ueno, 1999); polyurethane (PU) demonstrates piezoelectric behavior through hydrogen bonding-induced polarization and crystallization in its hard domain (Zhang et al., 2010); and poly-L-lactic acid (PLLA) possesses an inherently helical backbone, giving rise to a non-centrosymmetric molecular structure (Mishra et al., 2021a).

Piezoelectric composites

Lastly, composite piezoelectric materials can be fabricated to balance both piezoelectric and mechanical properties (Nie et al., 2024). For instance, dispersing crystals or ceramic nano particles in a piezoelectric polymer matrix will allow for more favorable flexible properties while bolstering piezoelectric properties. Some examples of these composites include PVDF/PZT (Jain et al., 2015) and ${\mathrm{BaTiO}}_3$ /PVDF (Yang et al., 2020). Moreover, piezoelectric particles can be mixed with passive polymers to create piezoelectric-passive polymer composites (Hou et al., 2020; Newnham et al., 1978). The composite connectivity is represented by a two-digit code, where the first digit indicates the number of dimensions in which the piezoelectric component is interconnected, and the second digit represents the connectivity of the passive polymer. A value of “3” signifies interconnection along all three spatial axes (x, y, and z; Newnham et al., 1978; Nie et al., 2024), this concept is further illustrated in Figure 4(b). A wide range of fabrication methods are utilized in the development of piezoelectric composites, a detailed discussion of these techniques falls outside the scope of this review, for a comprehensive overview refer to Jain et al. (2015).

Figure 4.

(a) Pressure sensing mechanisms, including piezoresistive, piezoelectric, triboelectric, capacitive, and iontronic. (b) Composite connectivity configurations. (c) Phases of percolation leading up to percolation threshold.

Capacitive sensing

Capacitive sensors function by detecting capacitance changes in response to mechanical deformation. Generally, a capacitive sensor consists of two conductive electrodes which are separated by a dielectric layer: when external pressure is applied the distance between the electrodes decreases resulting in an increase in capacitance. Typically, for skin-interfaced applications the dielectric layer consists of a soft polymer due to their flexible properties and their ability to be easily molded into porous sponge/foam like structures (Mishra et al., 2021b). Moreover, polymer composites enhanced with high-dielectric-constant filler materials have been adopted to improve overall sensitivity. Popular materials include PDMS (Hwang et al., 2021; Kou et al., 2018; Lei et al., 2012; Yang et al., 2022a), Ecoflex (PBAT; Park et al., 2017b), and PVDF (Luo et al., 2021; Yang et al., 2019b).

The capacitance of the sensor system can be calculated using equation (2) where, ϵ is the permittivity of the dielectric layer, A is the overlapping area, and d is the distance between the electrode layers.

$$C=\frac{ϵA}{d}$$

(2)

Capacitive sensing is a particularly popular mechanism used for flexible pressure sensors due to their low power consumption and ability to detect long sustained pressure while maintaining good stability. However, capacitive sensors systems still face various challenges which can interfere with their performance. The primary limitation of capacitive sensors is their susceptibility to electromagnetic interference and parasitic capacitance which may contribute additional undesired capacitance to the system. This can result in signal offset or distortion making it more difficult to detect the changes in capacitance resulting from induced pressure.

Triboelectric sensing

Triboelectric sensing, like piezoelectric sensing is capable of generating its own voltage potential negating the requirement of an external power source. Triboelectric sensors (also commonly referred to as triboelectric nanogenerators, “TENG”) convert mechanical pressure changes to electrical signals using two dissimilar materials which are brought into contact and then separated, resulting in charge separation owing to contact electrification and electrostatic induction (Fan et al., 2012a; Wang, 2013; Zhu et al., 2012). The effect, relying on difference in electronegativity and surface roughness of the two dissimilar materials (Fan et al., 2012b), allows for a wide range of material choices. However, it is important to select materials with high surface charge density to ensure a greater number of charges are available during frictional contact. A typical TENG consists of two materials that come into frictional contact: one with high electronegativity, such as polyethylene terephthalate (PET), which readily gains electrons, and the other with low electronegativity, such as Kapton, which readily loses electrons (Fan et al., 2012b). TENGs can operate in two common modes, those being contact-separation mode and contact-sliding mode. The contact-separation mode involves energy conversion driven by repeated cycles of physical contact and separation perpendicular to the material interface. Whereas the contact-sliding mode involves rotational or translational motion by the two material interfaces rubbing against each other without separation (Yu et al., 2023). Additionally, engineered micro- and nanostructures that increase the surface area can enhance the functional surface charge density of the material, thereby improving its sensitivity (Fan et al., 2012a; Nie et al., 2024). The electronegativity of the constituent materials can also be influenced by temperature and humidity (Lu et al., 2017; Nguyen and Yang, 2013).

The open circuit voltage, $V_{\mathrm{OC}}$ developed under contact-separation mode can be expressed using equation (3) where, σ is the triboelectric charge density, ${ϵ}_0$ is permittivity of vacuum, and d is the perpendicular gap distance between the two material surfaces (Wang et al., 2023).

$$V_{\mathrm{OC}}=\frac{\sigma d}{{ϵ}_0}$$

(3)

TENGs have emerged as a promising mechanism for skin-interfaced pressure sensing due to the flexibility, stretchability, biocompatibility, and self powered-nature of their constituent materials (Fan et al., 2012b). However, like piezoelectric sensing, TENGs are unable to detect long sustained pressure due to continuous charge generation requiring ongoing motion. Instead, they are best used for detecting highly sensitive pressure fluctuations and have been used for many of the same applications as piezoelectric sensing (Chen et al., 2014; Lin et al., 2017; Shen et al., 2021). Lastly, the signal output of TENGs is sensitive to fluctuations in temperature and humidity, which can hinder their ability to accurately and quantitatively measure pressure (Yu et al., 2023).

Iontronic sensing

Iontronic sensing (also referred to as interfacial supercapacitive sensing (Chang et al., 2021)) is a relatively recent development compared to the previously discussed mechanisms. While it can be viewed as a subcategory of capacitive sensing, its unique characteristics warrant its own discussion. Iontronic sensing as it relates to pressure sensing primarily refers to a specific mechanism which utilizes an established interfacial phenomenon known as the electrical double layers (EDLs). This sensing mechanism utilizes ionic liquids, hydrogels, or ionogels which are sandwiched between electrode layers, typically the ionic liquid is deposited as a droplet or suspended within a sponge or fabric structure (Chang et al., 2021; Chhetry et al., 2019; Lin et al., 2020). At the interface between the electrode layers and the ionic liquid/gel, there exists an extremely thin film (theoretically the thickness of 1 nm) called an electrical double layer that can hold a charge. Leveraging the interfacial capacitive effect as pressure is applied to the EDL allows for changes in the interfacial contact area, resulting in extremely high unit area capacitance (on the order of several $\mu F·{\mathrm{cm}}^{-2}$ ; Xiong et al., 2022).

This substantial increase in capacitance allows for ultra high sensitivity while being relatively immune to noise generated by electromagnetic interference and parasitic capacitance. However, the primary challenge regarding this sensing mechanism lies in its environmental instability, as ionic liquids/gels are prone to degradation when exposed to certain temperature and humidity levels (Xiong et al., 2022).

Material structures and characteristics for pressure sensing

Most flexible pressure sensors leverage specific engineered material structures to achieve enhanced properties. Whether aiming to optimize sensitivity or tailor mechanical properties, a wide range of material structures have been employed to meet specific functional requirements.

The performance of pressure sensors is strongly influenced by their structure and morphology. Morphological features such as porosity, alignment, or cellular and fibrous structures govern the mechanical flexibility, compressibility, and strain transfer efficiency to the active sensing layer (Taromsari, 2024). The overall configuration, for example, whether the sensor is designed as a film, fiber, or 3D network, affects its ability to conform to curved or dynamic surfaces, such as human skin, and enables large-area responsiveness. Together, these structural and morphological parameters directly influence key performance metrics and mechanical durability (Ma et al., 2018; Yang et al., 2017).

The final structure and large-scale layout of a sensor are shaped not only by the materials used but also by the fabrication method, which governs how the sensor’s morphology evolves during processing. These elements work in tandem, highlighting the interconnected nature of design parameters that collectively influence sensor architecture and performance (Wang et al., 2022a). While material choices were addressed in the previous section, the following discussion focuses on the key macroscopic layouts and morphological features commonly adopted in these sensors. From a morphological standpoint, various microstructures have been explored for resistive sensors, including compact, layered, cellular, planar, curved, wrinkled, fibrous, and porous forms. Among these, morphologies with high surface-to-volume ratios are particularly advantageous due to their enhanced mechanical adaptability and improved capacity for strain accommodation (Ma et al., 2017, 2018; Taromsari, 2024; Wang et al., 2022a, Zeng et al., 2021).

A widely adopted design feature is the incorporation of a porous structure within the sensor’s active layer and are mostly used for piezoresistive and iontronic sensing. In the case of piezoresistive sensing porous foam structures are often used to alter the sensitivity of the sensor, as the porosity of the structure dictates the surface area of the active layer and, consequently, alters the percolation threshold required for conductive pathways to form (Ding et al., 2019; Xu et al., 2017b). With regards to iontronic sensing, porous materials such as sponge-like structures are used to retain ionic liquids and gels, moreover, hydrogels themselves with crosslinking polymers may be employed to achieve a porous structure (Li et al., 2017; Zhao et al., 2022).

Depending on the processing strategy, targeted microstructural features, and intended performance outcomes, flexible pressure sensors can be engineered into a variety of large-scale formats. Based on existing studies, three principal macroscopic designs are commonly utilized: (i) one-dimensional (1D) fibers, (ii) two-dimensional (2D) films, and (iii) three-dimensional (3D) interconnected networks.

1-D fibers

Among the various formats explored for pressure sensors, one-dimensional (1D) fibers are particularly promising for skin-conformable applications due to their excellent mechanical flexibility and high specific surface area (Uzun et al., 2019). These fibers can also be readily integrated into wearable platforms through techniques like knitting and pleating, providing a seamless path to textile-based electronics (Seyedin et al., 2020).

Several fabrication approaches have been employed to produce 1D fiber structures. Common methods include wet spinning, electrospinning, and coating flexible yarns with sensing materials. In wet spinning, a suspension of the active material is extruded into a coagulation bath where it solidifies via phase separation. While this technique enables continuous fiber production, it demands careful control over multiple processing parameters such as suspension concentration, flow rate, and bath composition and temperature to maintain fiber uniformity and structural integrity (Li et al., 2019a; Zhang et al., 2020).

Alternatively, fiber-like sensors can be prepared by depositing active layers onto commercial yarns, often made from nylon-based polymers-using techniques like dip coating, spin coating, or spray coating. This approach offers strong compatibility with textile processes but typically requires interfacial binders or stabilizers to ensure the active material adheres securely to the yarn surface (Huang et al., 2021; Li et al., 2019a).

Electrospinning has emerged as a highly adaptable method for generating fibrous networks with tunable properties. It supports a broad range of polymer types and allows precise control over fiber diameter, porosity, mat thickness, and alignment (Taromsari, 2024). Electrospun fibers can be arranged randomly or in aligned arrays, and their porous, flexible nature provides excellent strain accommodation and high surface-to-volume ratios. Compared to wet spinning, electrospinning offers more accessible parameter tuning and easier structural customization. The resulting mats can be functionalized through additional coating processes, cut into 2D films, layered into 3D structures, or twisted into yarns suitable for textile integration, further enhancing their design flexibility (Jiang et al., 2019; Levitt et al., 2020; Ma et al., 2018; Taromsari, 2024; Xue et al., 2019).

Fabrics and mesh-based material structures possess permeable interwoven fibrous networks while allowing for a soft and flexible interface for comfortable skin contact. Fabric-based sensing materials have been developed leveraging the previously discussed sensing mechanisms, including, piezoresistive (Gong et al., 2014; Liu et al., 2017; Miyamoto et al., 2017; Yang et al., 2025; Yu et al., 2022; Zheng et al., 2021), piezoelectric (Dong et al., 2020; Kim et al., 2022a; Zhi et al., 2023), triboelectric (Dong et al., 2020; Meng et al., 2019; Sim et al., 2016; Wang et al., 2021b), and capacitive (He et al., 2018; Lee et al., 2015; Lin et al., 2018). With regards to iontronic sensing the use case for fabric/mesh-based material is slightly different as it effectively helps to suspend ionic liquids/gels sandwiched within the sensor stack-up and are typically achieved using electrospinning fabrication methods (Li et al., 2021; Lin et al., 2020). What distinguishes fabric-based material structures is their capability to be fully integrated into a wearable sensor garment, allowing for a system which achieves exceptional reusability, breathability, and integration into everyday clothing. A standout example is the “3DKnITS” smart sock system, seen in Figure 5(b), uses a piezoresistive nylon/spandex textile as the active sensing layer, sandwiched between two conductive nylon textile layers serving as the electrodes. The piezoresistive material was fabricated by coating polyester knit fabric with the conductive polymer Polypyrrole (PPy), resulting in a knit fabric which exhibits a piezoresistive response under compressive load (Wicaksono et al., 2022).

Figure 5.

(a) MoS2/HEC/PU porous sponge structure (Chen et al., 2023); (b) 3DKnITS piezoresistive textile-based sensor sock (Wicaksono et al., 2022); (c) Micropatterned Pyramidal microstructure for ionic gel-based sensor to enhance sensitivity (Cho et al., 2017); (d) Folding 2D pattern into 3D structure for multimodal sensing (Won et al., 2019); (e) Sensor array configuration with high spatiotemporal resolution (Ouyang et al., 2024). Reprinted with permission from: Chen et al., ACS Appl. Mater. Interfaces (2023), © ACS Publications; Wicaksono et al., EMBC (2022), © IEEE; Cho et al., ACS Appl. Mater. Interfaces (2017), © ACS Publications; Won et al., ACS Nano (2019), © ACS Publications; Ouyang et al., Biosensors and Electronics (2024), © Elsevier.

2-D films

In addition to electrospun films, other two-dimensional (2D) sensor configurations include coated interdigitated electrode patterns fabricated on polyimide (PI) substrates. These designs enable precise control over sensor geometry, avoid complex or abrasive processing steps, and are well-suited for scalable and reproducible manufacturing (Cao et al., 2019; Chen et al., 2021; Niu et al., 2021; Zheng et al., 2023). Micropatterns incorporate micro-engineered structures throughout the sensor to gain more control over key sensor characteristics, including sensitivity, response time, and hysteresis. Micropatterns serve a dual function: firstly, they form air gaps that effectively define the working response range of the sensor; secondly, they concentrate compressive forces at the tips of microstructures, allowing for a more dynamic contact area when compared to a planar surface (Cui et al., 2022; Li et al., 2016, 2019b). Additionally, micropatterns can introduce a directionally sensitive microstructure which when used in conjunction with a sensor array, enhance the detection of shear forces (Ma et al., 2015). Micropatterns can be incorporated with any of the previously discussed sensing mechanisms with common structure morphologies being bumps (Wang et al., 2016a), domes (Park et al., 2014; Zhang et al., 2017), cones (Qiu et al., 2018), pyramids (Cho et al., 2017; Yang et al., 2019a), and pillars (Chen et al., 2020; Luo et al., 2019) among others (Wu et al., 2024).

3-D networks

Three-dimensional (3D) sensor architectures are typically realized through methods such as template-assisted fabrication, self-assembly strategies, and additive manufacturing techniques (Yu and Koltun, 2015).

Three-dimensional mesostructures can be incorporated into sensor design as geometries which deviate from a planar structure (Guo et al., 2018; Huang et al., 2023). These 3D structures are used to amplify the sensor’s deformability to achieve a broader dynamic range or a high sensitivity. However, where these structures truly excel is demonstrated in their ability to simultaneously and distinctly measure multiple modes of mechanical stimuli such as bending, shear, and normal forces (Won et al., 2019). Common methods for achieving 3D structures for sensing include origami constructs (Misseroni et al., 2024), mechanically guided (Won et al., 2019), and multilayer stacking (Tang et al., 2023). Origami constructs are fabricated by folding two-dimensional patterns along specific creases to create 3D geometry (Ahn et al., 2010; Misseroni et al., 2024). A particularly interesting example of origami structures is using them to create morphing structures for capacitive sensing, wherein the structure is further folded in on itself, developing an increase in capacitance (Ray et al., 2024). Mechanically guided structures is a general term used for any structures which use mechanical forces to induce out-of-plane deformations of 2D patterns into 3D structures (Xu et al., 2015; Yan et al., 2016b). An advantage of these structures is their ability to be formed into intricate 3D architectures allowing for application tailored directional sensitivity (Kwak et al., 2020; Park et al., 2020). Moreover, multilayer stacking methods can be employed to create 3D sensing structures, typically leveraging 3D printing technology (Tang et al., 2023). The drawbacks of these 3D sensing structures are that they can be bulky due to their added thickness, leading to reduced flexibility and a lack of conformity to the skin.

Individual pressure sensors can be arranged into a grid array to produce a pressure map, which can be used to visualize the pressure distribution over large surface areas. This configuration is particularly useful for monitoring dynamic pressure patterns (i.e. gait monitoring (Wang et al., 2020) or tactile sensing (Sundaram et al., 2019)) and allows for the detection of subtle changes in force distribution. Typically, for a sensor array system the top and bottom electrode layers are arranged in rows perpendicular to each other. On the electronics side of the system, multiplexers can be used to rapidly measure the electrical signal at each node. Moreover, sensor arrays do not need to be constrained to a planar arrangement; alternatively, sensor nodes can be arranged in a three-dimensional space to form a point cloud system. A point cloud pressure system can be especially useful for mapping the pressure along the surface of three-dimensional objects, enabling spatially resolved pressure readings for complex geometries and irregular surfaces.

Time series classification fundamentals

The engineered materials discussed within the previous sections directly govern the electrical response of skin-interfaced pressure sensors. They influence characteristics such as signal amplitude, noise, and stability among others. Consequently, sensor outputs are intuitively represented as time series data, typically expressed as changes in resistance, voltage, or capacitance over time. These temporal signatures motivate the application of machine learning-based time series classification techniques.

This section seeks to explain fundamental concepts relating to time series classification, including the distinction between univariate and multivariate time series, approaches to labeling time series data, and various classification techniques ranging from traditional methods to the most recent breakthroughs. Understanding these time series concepts is particularly relevant for skin-interfaced pressure sensors, as time series data represents the primary format in which these devices output their measurements. Lastly, prominent state-of-the-art models will be explored.

Time series data

Time series data refers to a sequence of data points collected over a period of time, with each data point corresponding to a measurement taken at a specific time; in the context pressure sensing the data points are pressure measurements. In comparison to static n-dimensional feature datasets (e.g. age, height, weight), a standard univariate time series dataset only has two features: the measured data point and the time at which it was recorded. Whereas a multivariate time series can be viewed as a collection of univariate time series, where each of the univariate time series are their own dimension, examples of both univariate and multivariate time series can be seen in Figure 6(a) Additionally, time series datasets are serial dependent meaning that they are inherently sequential, with each measurement depending on the one preceding it (Bock et al., 2021).

Figure 6.

(a) Multivariate time series data composed of univariate time series with dimensions denoted as M. (b) Whole series classification, the whole time series X(t), is associated with one label y. (c) Timepointwise classification, the time series X(t) and the label y annotates a specific point. (d) Window-based classification, time series X(t) is given a time-stamped label from $y_1$ to $y_2$ .

Labeling time series datasets

Pressure time series data can use both supervised and unsupervised machine learning methods for classification, with supervised approaches requiring labeled datasets (Ethem, 2020). There are three primary strategies for labeling time series data (Figure 6(b) and (c)), these being global labels, window-based labels, and time-point-wise labels (Bock et al., 2021).

The global label method assigns a single label to an entire time series, this method does not attempt to label a specific event within the time series, rather it only aims to assign a class to the time series as a whole. As an example, consider classifying an entire pressure sensor pulse recording as “regular” or “irregular,” we are not necessarily identifying a specific event within the time series which constitutes an irregular pulse, instead we are simply assigning the label based on outcome (Bock et al., 2021; Kadous and Sammut, 2005). Although, this method is relatively simple and straightforward, it lacks the ability to identify specific events within the time series, disregarding temporally dependent patterns. Lastly, this method lacks the ability to make predictions within the time series data, as the time series is being classified as a whole.

Time pointwise labels annotate specific points in time indicating the occurrence of events within the time series. This method is almost the exact opposite to a global label in the sense that we are no longer concerned with classifying the time series as a whole, instead we are concerned with specific events within the time series. Continuing with our example of an irregular pulse, consider labeling the point in time right before each irregular pulse waveform within the time series. Moreover, this method can be extended to where every time step within the dataset is labeled, resulting in a sequence of labels which match the length of the time series (Kadous and Sammut, 2005). The clear advantage associated with a time pointwise approach is the ability to focus on time points leading up to an event, enabling more impactful predictions. Additionally, since only time points prior to the event are considered, it eliminates data leakage from post-event time points which would otherwise be considered when using a global label. However, it can be challenging to determine the ”start time” of an event, especially regarding biomedical applications where there is no universally agreed upon start time for certain phenomena (Bock et al., 2021). Therefore, consistently labeling the start of events relies heavily upon domain experts.

Window-based labels segment a time series into discrete windows of time which can be either overlapping or non-overlapping, these windows are then assigned labels. This method allows for local analysis within the time series, allowing the model to learn from smaller segments of the time series. Thus, window-based labeling is particularly good at capturing local patterns which may otherwise be missed when using a global label approach. However, challenges arise when defining the time windows, it could be the case that a particular window contains multiple events, moreover, how do we exclude the transition points between class windows? These factors can lead to label noise resulting in the model learning false correlations.

Time series classification methods

When discussing methods for classifying time series data there are generally six categories: feature-based, distance-based, shapelet-based, dictionary-based, convolution/kernel-based and deep learning-based methods. Each method will briefly be introduced and discussed. The most traditional approaches encompass feature-based and distance-based methods; the middle generation of approaches include shapelet-based and dictionary-based methods; and the most recent developments include convolutional/kernel-based techniques along with deep learning architectures (Sun et al., 2023). However, there have been recent algorithmic developments in each of the categories to further improve their performance, additionally, hybrid methods have been employed which combine two or more methods in order to achieve higher classification accuracy.

Feature-based methods

Feature-based methods aim to extract meaningful features from time series data. These features can range, for example they can be statistical (e.g. mean, minimum, maximum values), shape-based (e.g. capturing changes in slope or number of peaks), or cyclic patterns derived from the frequency domain using Fourier transforms (Wu et al., 2018; Fulcher, 2017). A feature space can be created using these compiled features, where then traditional machine learning models such as Support Vector Machines (SVM; Cortes et al., 1995), decision trees (Quinlan, 1986), or Artificial Neural Networks (ANN; Guo et al., 2010) may be applied.

Feature-based approaches are generally considered to be the easiest to interpret because the chosen features which are being used for classification are clearly outlined. However, a severe shortcoming of this method is the task of feature extraction, which typically entails manual extraction with the input of a domain expert (once automatic feature extraction comes into play it is likely a neural network being used; Chernikov et al., 2022). Moreover, an excess number of features may introduce high dimensionally resulting in the model overfitting. In addition, due to the fact that an entire time series is essentially being summarized by a set of extracted features, this approach can result in a loss of detail.

However, interval-based approaches circumvent problems which arise due to extracting features from the entire time series. This approach, commonly known as a time series forest classifier (TSF), essentially segments the time series into phase dependant intervals and then extracts features from each of these intervals which are used to train a decision tree (Middlehurst et al., 2024; Deng et al., 2013).

Distance-based methods

Distance-based methods classify time series data based on the similarity between multiple time series, this is done by calculating a distance metric that quantifies how the series differ over time. A classical approach to applying this method would be by measuring the Euclidean distance between time series. However, this method can only work if both time series are aligned and have the same number of time steps (are the same length). Given that most time series data collected from real-world environments will vary in length, more sophisticated approaches are required to classify the data.

Dynamic time warping with nearest neighbor algorithm

Dynamic Time Warping (DTW) coupled with a Nearest Neighbor (NN) classification algorithm address the variation in time series length and alignment. DTW aims to align two time series regardless of differences in length and synchronicity, effectively accounting for temporal distortions within the time series dataset and can be seen within Figure 7(a). The NN classifier (specifically the 1-NN classifier) then predicts the class of the time series based on the closest measurement from the training time series dataset (Abanda et al., 2019). DTW with NN is generally considered to be the standard algorithm for time series classification (TSC) due to its respectable accuracy scores achieved on benchmark datasets (Abanda et al., 2019). However, it is safe to say that the DTW with NN approach has started to be outperformed by newer distance-based and non-distance-based algorithms alike. This is particularly apparent when comparing the results from the 2016 TSC ‘bake-off’ with those of the 2024 evaluation (Bagnall et al., 2017; Middlehurst et al., 2024).

Figure 7.

(a) Dynamic Time Warping (DTW) being used to temporally align time series data. (b) Shapelet-based classification method being used on univariate time series. (c) Dictionary (bag-of-words) pipeline (Middlehurst et al., 2024). (d) Convolutional Kernel pipeline. (e) Fully Convolutional Network (FCN) architecture demonstrating an example of a deep learning-based approach. Reprinted with permission from: Middlehurst et al., Data Mining and Knowledge Discovery (2024), © Springer Nature.

Shapelet-based methods

Time series shapelets are phase-independent, highly discriminative segments of the time series training data. Shapelets serve as a feature in themselves to be used for classification, their presence, absence, or similarity throughout newly introduced time series is predictive of its class label (Middlehurst et al., 2024; Ye and Keogh, 2009). To identify matching shapelets throughout the time series, a sliding window approach is employed, wherein the shapelet subsequence is compared to each window subsequence of equal length throughout the time series by calculating the z-normalized Euclidean distance between them, as seen in Figure 7(b). The calculated distance is then used as a feature typically within a decision tree classifier for evaluation (Ye and Keogh, 2009).

A key advantage of using shapelets for TSC problems is their clear interpretability, the waveform of the shapelet is usually clearly identifiable and therefore the model’s decisions can be linked to relevant patterns within the time series. Additionally, shapelets have the advantage of being inherently local features, which enhances the model’s robustness to noise within the dataset; contrast to approaches which only consider global features. Lastly, optimized shapelet-based methods can achieve considerably lower classification time complexity when compared to benchmark distance-based approaches such as DTW with NN (Ji et al., 2018).

Dictionary-based methods

Dictionary (also known as bag-of-words) approaches convert time series data into a sequence of symbolic representations, analogous to letters in a text document. This series of letters is then segmented into “words,” which can be understood functionally as time windows in the sense that they are phase-independent subsequence of the time series (Figure 7(c)). The time series data can be discretized using the Symbolic Fourier Approximation (SFA) method (Schäfer and Leser, 2017). Once discretization is complete, a feature vector comprised of the “words” is created, upon which a standard machine learning classification method (i.e. SVM, K-NN, ANN) is applied.

One advantage of a dictionary-based approach is its high discriminative power, the process of representing the time series as discrete segmented “words” yields strong class separation. Additionally, this method is well suited for effectively capturing the overall structure of the whole time series while remaining robust to noise (Wang et al., 2013).

However, a major drawback of this technique is that it ignores temporal order within the time series, this is due to the series being collapsed into a bag of “words” which ignores the order of local patterns. This limits the utility of this approach for tasks which require accurate time-dependent sequencing, particularly those requiring precise identification of recurring patterns (motifs) within the dataset (Wang et al., 2013; Patel et al., 2002). Additionally, the number of “words” must be predefined; if the chosen number of words is too little it will yield a course, under-discriminative feature vector, whereas choosing too many can introduce noise (Wang et al., 2013).

Convolutional/Kernel-based methods

A convolutional, or kernel-based approach applies a one-dimensional kernel (also known as a filter) in a sliding window fashion across the time series, performing a dot product operation between the kernel and each equally sized series subsequence (Goodfellow et al., 2016). This operation produces a series of activation maps which indicate local patterns within the time series (single dimension), this concept is easier to visualize when applied to images (two dimensions; Zhou et al., 2015). The activation maps then undergo pooling which essentially extracts a summary a single summary feature from each of them (Gholamalinezhad and Khosravi, 2020). The resulting features are then concatenated into a single feature vector, which serves as input into a standard classifier; this pipeline is illustrated in Figure 7(d).

A recent development within the area of convolutional/kernel-based approaches has been Random Convolutional Kernel Transform (ROCKET). This approach takes the previously discussed fundamental concept and intensifies it, ROCKET works by initiating a large number of randomly parameterized kernels (parameters include length, dilation, padding, and bias). Following the same convolutional sliding window approach mentioned earlier, two pooling operations are performed: global max pooling and proportion of positive values (PPV) pooling (Gholamalinezhad and Khosravi, 2020). The features extracted from each operation are concatenated into a single feature vector to be used in a RIDGE regression classifier (Dempster et al., 2020; Middlehurst et al., 2024).

Kernel-based approaches have the benefit of automatically learning features directly from raw data, limiting the need for manual feature engineering. Additionally, the convolutional sliding window allows the model to detect localized patterns and structures within the data, which may be indicative of a specific class. Moreover, this approach introduces shift-invariance allowing patterns to be recognized regardless of their temporal position. The pooling operations aggregate kernel activations, thereby reducing the model’s sensitivity to temporally shifted patterns, which may vary across time series within the dataset. This results in a more robust model which can better handle slight misalignments of time series data within a given dataset, ultimately reducing the need for extensive preprocessing.

Despite its advantages, kernel-based methods suffer from a lack of interpretability. While the sliding-dot-product and pooling operations are understandable, it is still less intuitive compared to shapelet or hand-picked feature-based methods, which are inherently more intuitive. Additionally, hyperparameters (the parameters which define the kernel itself) can require extensive tuning, although random-kernel methods can help minimize this process.

Deep learning-based methods

Deep learning-based methods are the most prominent methods currently being explored within recent literature. One of the fields that deep learning methods first had a substantial impact on was computer vision, with the advent of AlexNet, which achieved unprecedented classification accuracy on the LSVRC-2010 ImageNet dataset (Krizhevsky et al., 2012). Since this catalyzation there has been a sentiment that deep learning methods may be able to revolutionize TSC given a large enough labeled dataset, as AlexNet did using the ImageNet dataset (Ismail Fawaz et al., 2019a; Russakovsky et al., 2015).

In its most basic form deep learning architecture consists of a sequence of interconnected layers of ”neurons”. The network begins with an initial input layer, followed by a series of hidden layers, and finally ending in an output layer. Each layer applies a linear transformation followed by a non-linear activation of their inputs by applying a non-linear activation function (e.g. ReLU, Sigmoid, Tanh (Agarap, 2018; Szandała, 2021)). By propagating these transformed inputs through successive hidden layers, the network develops a more abstract representation of the data until ultimately outputting a prediction at the output layer (LeCun et al., 2015). The most basic architecture is known as a Multilayer Perceptron (MLP) which is comprised of a series of fully connected (FC) layers (Popescu and Balas, 2009). However, Artificial Neural Network (ANN) architectures can become increasingly complex, comprising various types of layers such as convolutional (O'Shea and Nash, 2022), pooling (Gholamalinezhad and Khosravi, 2020), recurrent (Meher et al., 2013a), and dropout layers (Meher et al., 2013b), to name a few. Finally, these networks undergo a process known as training during which a loss function is used to compare the prediction with the ground-truth. This loss is then used in backpropagation to iteratively update each parameter within the network (LeCun et al., 2015).

Convolutional neural networks (CNN)

According to Ismail Fawaz et al. (2019a), the most effective baseline deep learning architectures for TSC are Fully Convolutional Neural Networks (FCNs) and Deep Residual Networks (ResNet); both being variations of discriminative end-to-end Convolutional Neural Networks (CNN). The architecture of FCNs mainly consist of convolutional layers while omitting pooling layers. Additionally, the final Fully Connected (FC) linear layer typically found at the end of Convolutional Neural Network (CNN) architecture is replaced with a Global Average Pooling (GAP) layer, followed by a SoftMax classifier. The second baseline architecture appropriated for TSC is ResNet. This architecture’s claim to fame is its use of residual connection (also known as a skip connection) which allows the input to a layer to bypass intermediate transformations and be added directly to the layer’s output. As a result, the network learns the residual, or the difference between the input and the output. These connections allow for a deeper network architecture by mitigating the effects of vanishing gradients. The proposed FCN architecture used in Wang et al. (2016b) is seen in Figure 7(e).

CNN are effective at automatically extracting features from the dataset by recognizing discriminative subsequences within the time series limiting the need for manual feature engineering. Additionally, CNN benefit from weight sharing, allowing the same set of weights to be applied across the entire sequence, reducing the number of overall parameters. Additionally, the use of inception-style filters (Szegedy et al., 2016), dilated convolutions (Yu and Koltun, 2015), and residual connections can improve the model’s performance by improving feature extraction ability, capturing long-term dependencies in the data, and mitigating vanishing gradients respectively. Typically, when these more advanced techniques are not used, CNN architectures will struggle with capturing long-term dependencies and suffer from the vanishing gradient problem.

Although CNN-based architectures are extremely promising for TSC, there are multiple drawbacks. Firstly, CNN typically require large, labeled datasets, which can be rare in the context of biomedical time series data. When the dataset is too small, it often leads to the model overfitting and failing to generalize effectively. Secondly, interpretability of these models is often difficult, due to the black box nature of Deep Neural Network (DNN) architecture (Zhang et al., 2021; Zhang and Zhu, 2018). Additionally, excessive pooling layers can lead to information loss surrounding specific temporal events within the dataset. This is due to the data undergoing successive layers of abstraction such that brief but significant patterns are lost in place of a more generalized trend. Lastly, these models may require considerable hyperparameter tuning to achieve optimal performance, which can take a substantial amount of time.

Sensor characteristics and their influence on machine learning design

The physical characteristics of skin-interfaced pressure sensors influence machine learning architectures and pipeline design. Non-linear sensor output caused by hysteresis or viscoelastic drift introduce baseline shifts and path-dependant responses that must be addressed using preprocessing techniques. Moreover, sensor noise can vary significantly depending on the sensing mechanism and constituent materials. Consequently, baseline correction, detrending, and normalization are all common preprocessing techniques used to mitigate drift and sensor-to-sensor variability (Tawakuli et al., 2025). Low-pass or moving-average filtering may also be used to improve the signal-to-noise ratio (SNR). Lastly, the sample rate of the sensor system and subsequent time series data must be considered. Low sample rates may mask physiological responses whereas high sample rates typically increase noise sensitivity, motivating the use of downsampling.

These constraints caused by the sensors themselves make model selection an important design consideration. Certain signals warrant specific models; for instance, linear classifiers may be inadequate for sensors exhibiting strong hysteresis or drift prompting the use of non-linear approaches ranging from conventional methods such as SVM to more advanced approaches such as CNNs, kernel, or dictionary-based models. As another example, when periodic or quasi-periodic signals are observed, a shapelet approach is likely the best option. Lastly, as will be discussed in the following section, ensemble models can be particularly powerful by combining multiple models together into a single pipeline, thereby improving robustness to noise, variability, and non-linear behavior.

Prominent machine learning algorithms for time series classification

An extremely promising approach has been to combine multiple of the previously discussed methods to create a hybrid model. A recent study by Middlehurst et al. (2024) compared various state-of-the-art TSC models, it was found that some of the best performing models were ones which combined two or more of the previously discussed approaches.

A particularly well performing hybrid approach was the Collective of Transformation Ensembles (COTE) family of models (Bagnall et al., 2015). Specifically, one of the top performing models within this family is HIVE-COTE version 2 (HC2; Middlehurst et al., 2021b). HC2 uses an ensemble of the following methods: Shapelet Transform Classifier (TSC; Hills et al., 2014), an ensemble of dictionary-based approaches known as Temporal Dictionary Ensemble (TDE; Middlehurst et al., 2021a), an ensemble of ROCKET-based approaches known as Arsenal, and an interval-based model called Diverse Representation Canonical Interval Forest (DrCIF; Middlehurst et al., 2021b).

Another extremely well performing model which must be highlighted was an ensemble approach known as Hybrid Dictionary-ROCKET Architecture (Hydra) and its extension Multi-ROCKET Hydra (MR-Hydra), which combines Dictionary and ROCKET-based approaches (Dempster et al., 2023). On average, MR-Hydra performed similarly to HC2 in terms of accuracy and distinguished themselves from the other models evaluated.

Although both HC2 and MR-Hydra are generally considered to be the best performing models, they still struggle with scalability when applied to large amounts of data. Notably, the main model identified in the review that demonstrated scalability was QUANT while simultaneously being an exceptionally fast model. This model employs an interval feature extraction technique on four representations of the data: the raw time series data, first-order differences, second-order differences, and Fourier coefficients (Dempster et al., 2024).

Lastly, the model known as InceptionTime is a deep learning-based approach which is widely regarded as one of the leading approaches for TSC, consisting of an ensemble of five deep CNN models (Fawaz et al., 2019b). Moreover, an extension of InceptionTime was proposed, called H-InceptionTime, which adds hand-crafted one-dimensional convolutional kernels which work in tandem with the existing InceptionTime architecture, further improving the model’s performance (Ismail Fawaz et al., 2019a).

Applications of machine learning for flexible skin-interfaced pressure sensing

There has been a plethora of research papers published within the past decade which have explored applying machine learning algorithms to biomedical pressure sensing. Many of the latest solutions have been leveraging both the recent advancements in pressure sensing materials along with a range of machine learning algorithms to enhance health monitoring. Table 1 provides an overview of recent studies exploring the topic, with a primary focus on classification tasks and the corresponding machine learning models applied. Additionally, Figure 8 highlights prominent applications where machine learning models were used in conjunction with flexible pressure sensor systems.

Table 1.

Catalog of recent flexible sensor studies which integrate machine learning techniques.

Study	ML architecture	Task	Sensing mechanism	Active material(s)	Sensor structure
Ma et al. (2024)	CNN+ViTs (Transformer)	Classification of posture	Piezoresistive	Coaxial wet-spinning of MXene/PVA@ANFs	Fabric/Fibers
Babu et al. (2024)	Dual-input branch autoencoder using FC layers	Denoising and reconstruction of pulse signal	Piezoresistive	CNT/Carbon	Sponge
Sun et al. (2024)	Moving average pre-processing layer+FC layers	Classification of finger positioning	Triboelectric	Vertical graphene (vG) + fluorinated ethylene propylene	Substrate+Sandwiched between FEP film
Tang et al. (2024)	FC layers (ANN)	Classification of hand gestures	Piezoresistive	Reduced graphene oxide/cellulose nanocrystalline (rGO/CNC) composite film	Substrate+Sandwiched between PDMS film
Guo et al. (2024a)	FC layers (ANN)	Classification of biomechanical movement	Piezoresistive	MXene-FPDMS	Sponge
Li et al. (2024)	CNN	Classification of plantar pressure/gait	Iontronic	PI/IL fibers as dielectric material	Substrate+Sandwiched between polyimide film
Yang et al. (2024a)	CNN	Classification of activities and gestures	Piezoresistive	polyimide fiber (PIF)/carbon black (CB)	Aerogel
Zhang et al. (2024)	KNN	Classification of wrist posture	Piezoresistive	(PEDOT:PSS)/cellulose nanocrystals	Substrate+Sandwiched between PDMS film with dome like structure
Guo et al. (2024b)	LSTM	Classification of biomechanical movement	Piezoelectric	PVDF fibrous film+MWCNT layer	Substrate+Sandwiched between spin coated PDMS
Beigh et al. (2023)	CNN	Classification of gait	Piezoelectric	BTO/SU-8	Substrate+Sandwiched between Kapton film
Fang et al. (2023)	Grid Search+SVM	Classification of subvocalization	Piezoelectric	Polyvinylidene fluoride (PVDF)	Substrate+Sandwiched between PDMS film
Lin et al. (2022a)	XGBoost, KNN, SVM	Classification of subvocalization	Piezoresistive	Radial anisotropic porous silver fiber (RAPSF)	Substrate + encapsulated
Wang et al. (2022b)	FC layers (ANN), Decision Tree, Random Forest	Classification of parkinsonian and hemiplegic patients	Triboelectric	Graphene Oxide	Hydrogel using Polyacrylamide
Chen et al. (2023)	Support Vector Machine (SVM)	Classification of subvocalization	Piezoresistive	Molybdenum disulfide (MoS2) and hydroxyethyl cellulose (HEC)	Sponge
Fang et al. (2022)	CNN	Classification of breathing states	Triboelectric	PVDF	Textile
Kim et al. (2022b)	CNN	Classification of subvocalization	Piezoresistive	Silicon Nanomembrane (SiNM)	Substrate+Sandwiched between Polyimide film
Xu et al. (2021)	CNN	Classification of biomechanical movement	Iontronic	Polyvinyl alcohol (PVA)-KOH ionic film	Substrate+Sandwiched between silicone
Wang et al. (2021a)	GMM	Identification of speaker based on audio signals	Piezoelectric	PZT	Substrate+Sandwiched between PET film
Lin et al. (2022b)	DTW	Identification of speaker based on audio signals	Piezoelectric	PTFE	Substrate+Sandwiched between PET film
Long et al. (2020)	CNN	Classification of subvocalization	Piezoresistive	Molybdenum-carbide-graphene (MCG) composite	Porous microstructure+Paper substrate
Bokka et al. (2020)	KNN, DT, RF, ET	Classification of force stimuli	Piezoresistive	SnS2 Quantum Dots	Substrate+Sandwiched between PVA
Zhou et al. (2020)	Single Layer Perceptron (SLP)	Classification of music clips	Piezoresistive	Porous metal-organic framework (MOF)	Mesh
Chu et al. (2018)	DTW + unspecified classifier	Classification of pulse normality	Piezoelectric	Ecoflex	Substrate+Sandwiched between FEP film
Ramírez et al. (2018)	K-means clustering+SVM, AdaBoost, Custom L1-distance algorithm	Classification of swallowing activity	Piezoresistive	Palladium (PdNi) + Graphene	Substrate+Sandwiched between Polymide film
Han et al. (2018)	GMM	Identification of speaker from audio signals	Piezoelectric	PZT	Substrate+Sandwiched between PU film

Figure 8.

(a) Human pulse diagnosis system for medical assessments using a wearable piezoelectric sensing system with a DTW + classifier approach (Chu et al., 2018). (b) A PVDF and PDMS-based flexible sensor for decoding subvocalization using a classical machine learning approach (SVM; Fang et al., 2023). (c) Vertical graphene triboelectric pressure sensor array as a tactile sensing system for classifying finger actions (Sun et al., 2024). (d) Iontronic pressure sensor for biophysical monitoring using deep learning-based approaches for knee injury rehabilitation (Xu et al., 2021). Reprinted with permission from: Chu et al., Adv. Funct. Mater. (2018), © Wiley; Fang et al., Sensors & Actuators A: Physical (2023), © Elsevier; Sun et al., Nano Energy (2024), © Elsevier; Xu et al., Microsyst. Nanoeng. (2021), © Springer Nature.

The survey conducted in Table 1 shows that deep CNN models are dominant for applications which require spatial awareness (especially with multi-element sensor arrays) such as gesture or posture recognition. Whereas, simpler classifiers such as SVM/DTW are used for low-frequency one-dimensional repeating signals such as pulses or audio signals. For example, an immediately apparent correlation is between piezoresisitve sensors that generate two-dimensional pressure maps and the adoption of CNNs, as these models are well suited to leverage the inherent spatial structure of such data (Krizhevsky et al., 2012; Figure 4(Right)).

General health monitoring

Health monitoring is an umbrella term used to describe the tracking of physiological data (e.g. heart rate, blood pressure, respiratory rate) of individuals to detect abnormalities which may indicate potential health concerns. Recent developments in thin-film pressure sensing materials have allowed for a convenient mode of collecting physiological data from the surface of the skin. Additionally, applying machine learning techniques to analyze sensor signals can improve the efficiency of health monitoring by providing real-time anomaly detection.

A common application is the heart rate monitoring, Jia et al. (2024) developed a non-invasive, skin-integrated pulse sensing system which uses deep learning autoencoder technique to denoise and reconstruct heart rate. While this is not a classification-based task, it acts as a powerful demonstration of the usefulness of leveraging machine learning techniques. Additionally, this method would need to be coupled with a classification model to eventually be useful for anomaly detection. Chu et al. (2018) study which demonstrates the usage of a DTW algorithm to classify abnormal pulse signals addresses this gap by supplying the classification component necessary for effective anomaly detection. Another exemplary use case is monitoring respiratory rate, Fang et al. (2022) developed an on-mask sensor network which uses CNN architecture to classify different respiratory signals (e.g. normal breathing, deep breathing, occasional cough) to aid in real time respiration monitoring and illness detection. Moreover, Babu et al. (2024) developed an electronic-skin respirometer (eSR) which can be attached to the chest to detect breathing patterns for early detection of chronic obstructive pulmonary diseases. They tested multiple regression techniques (Gradient Boosting Regression (GBR), Linear Regression (LR), Support Vector Regression (SVR)) to classify respiratory diseases, specifically emphysema and chronic bronchitis.

Subvocalization and swallow detection

A prominent application which is being explored extensively is the placement of thin-film sensors on the throat (across the larynx region) or around the mouth. These sensors detect subtle changes in pressure along the surface of the skin whenever the user mouths a word (subvocalization). The signals from the sensors can be preprocessed and then fed into a machine learning model to classify the signal into the respective word. This application is particularly useful for individuals who are mute or have limited verbal communication abilities. In a study carried out by Fang et al. (2023), a flexible thin-film sensor using Silver-PVDF as the active material was placed across the throat. The signal collected by the sensor underwent feature extraction preprocessing, using wavelet packet energy (an extension of Discrete Wavelet Transform) as time-frequency dynamics features while wavelet entropy and LZ complexity were used as non-linear dynamics features. An SVM classification model was then used to classify the subvocalization of 26 different letters, achieving an accuracy of 90.55%.

An additional application using a similar implementation approach is swallow detection. By using a machine learning algorithm, subtle variations in signal waveform corresponding to different swallowing activity can be distinguished. Using this premise, Ramírez et al. produced a piezoresistive thin-film sensor, using Palladium nanoislands (PdNI) as the active material to monitor physiological activity while swallowing. They tested various different machine learning algorithms to classify severity of dysphagia (difficulty swallowing), determining that their L1-distance-based model outperformed both the SVM and AdaBoost models (Ramírez et al., 2018). This demonstrates the power of distance-based approaches for discriminating different time series waveforms.

Hand gesture recognition

A popular demonstration within literature involves attaching strain sensors on regions of the hand and applying machine learning algorithms to classify different hand gestures. Various implementation approaches have been explored, ranging from attaching the sensors directly to the skin to embedding the sensors to an electronic-textile-based glove system (Pan et al., 2021; Sun et al., 2024; Yang et al., 2024a; Wang et al., 2024b). This application is particularly promising for deciphering hand motion in virtual/augmented reality systems, without requiring external motion capture cameras.

A study performed by Sun et al. developed a wearable triboelectric sensor system featuring sensors that wrap around each finger, allowing for precise recognition of individual finger positioning and enabling classification of complex hand gestures. To demonstrate the effectiveness of the system they used an FCN machine learning model to classify different table tennis technical actions. The machine learning model was trained on a fairly substantial time series dataset of 3200 samples across 16 classes of table tennis actions (200 repetitions for each class). Each time series sample was multi-channel, composed of 10 distinct channels corresponding to different positions along the fingers. This study successfully captured the advancements in deep learning approaches for classification of multi-channel biomedical time series data, attaining an accuracy score of 98.1% (Sun et al., 2024).

Motion analysis and posture recognition

Tracking the biomechanical movement of individuals lends itself to be useful in many different fields such as motion capture, sports science, and medical rehabilitation to name a few. The ability for thin-film sensor arrays to be placed along areas of the body to map pressure distribution allows for the collection of large amounts of quantitative data. This data can be used in applications ranging from aiding in performance optimization for sports science to personalized treatment plans for medical rehabilitation. Moreover, the placement of strain sensors at points of interest along the body such as joints allow for specific movements to be detected and monitored.

Gait analysis uses sensor arrays at the bottom of the foot to map plantar pressure during activities (Beigh et al., 2023; Li et al., 2024; Wicaksono et al., 2022). Li et al. (2024) developed fibrous iontronic sensors which were integrated into an insole-based system to map the plantar pressure distribution during phases of static standing and dynamic walking, with the pressure data visualized as a heat map. A one-dimensional Convolutional Neural Network (1D-CNN) was used to classify 9 distinct plantar stances (e.g. Neutral, Pronator, Supinator), using a dataset of 300 heatmap images and achieving an accuracy of 99.8%. Similarly, sensor array systems paired with machine learning models can be integrated into furniture (e.g. office chairs, wheelchairs, medical beds) to detect and prevent pressure ulcers or correct posture (Ran et al., 2021; Wang et al., 2020).

Thin film sensors positioned at joints have been an effective method for tracking motion (Xu et al., 2021; Zhang et al., 2024). Xu et al. (2021) devised an iontronic thin film sensor which was placed across the knee to detect different knee joint postures depending on the knee bending angle. An FCN architecture was used to classify the various joint positions based on the stress induced on the sensor system, a relatively small dataset of 80 samples comprised of four classes, each of which representing a different bending angle range. The model achieved a high accuracy rate of 93.27% however, it must be noted that due to the small sample size it is very likely that the model is overfit and would struggle to generalize when tested on a more diverse set of data.

Conclusion and Outlook

Recent advances in nanomaterial engineering have enabled a new generation of skin-interfaced sensors, particularly driven by their innovative engineered micro/nano structures and characterized by their thin nature and exceptional application-relevant sensitivity. Particularly, sensors which leverage the iontronic mechanism show promise for furthering the field of skin-interfaced pressure sensing, especially relating to biomedical applications. Uniquely, iontronic sensing has been demonstrated to provide an ultrasensitive response while remaining immune to noise. Moreover, micropatterns composed of micro-engineered structures with dome or pyramid morphologies have been especially useful for fine-tuning the sensor’s response. Array based sensor systems have come to dominate much of the recent research within this field due to their ability to create pressure mappings and significant application to problems which require a measurement of non-uniform pressure distribution across an area (e.g. gait analysis, posture recognition). I suspect that as research within this field progresses, sensor arrays will only become higher fidelity allowing for granular pressure distribution imaging.

However, despite significant progress in materials engineering aspect of skin-interfaced pressures sensors, several fundamental problems remain. Many mechanisms designed for high sensitivity applications (e.g. iontronic, triboelectric, and micro-structured capacitive designs) exhibit trade-offs between sensitivity, hysteresis, and environmental stability. Additionally, for many of these cutting-edge sensors, long-term mechanical fatigue and signal drift have been insufficiently characterized and tested under repeated loading for extended periods of time. Lastly, very few of these sensors have been deployed and evaluated in clinical environments, and as a result, many have not yet been exposed to the full range of environmental and challenges encountered during real-world use.

The advent of AlexNet brought about a new era of machine learning research, one dominated by deep neural networks, extremely large datasets, and state-of-the-art Graphical Processing Units (GPUs) to train the models. It was inevitable that these advancements would spill over into other fields of research to be used synergistically. Time series’ are the primary form in which pressure data is visualized and stored, within only the past 5 years there have been significant advancements within the field of TSC. With dictionary and deep learning-based approaches demonstrating significant gains in performance, leading to models such as ROCKET and InceptionTime respectively. However, the main advancements have come from the development of ensemble models, such as HC2, which combine multiple approaches to achieve state-of-the-art results.

Many material sensor studies have started to use machine learning techniques to process and analyze their sensor signal data and this is no different for many of the skin-interfaced sensor studies which have been published recently. It is becoming more common to use some form of machine learning model to perform analysis, whether it be a more traditional approach (e.g. SVM, RF) or newer approaches such as deep learning-based approaches. With respect to Table 1, many recent studies have been exploring deep learning approaches to tackling time series pressure data. However, there is still a severe lack of state-of-the-art TSC models being used in conjunction with these innovative new sensor technologies. Additionally, many of the models being used within the cataloged studies are working with either a small dataset or a dataset which is not very diverse. This is primarily because the dataset usually only consists of data, which was collected solely by the authors themselves, thus they typically don’t have access to a wide range of participants. It is likely that many of the classification models developed within the cataloged papers are overfit and would have a hard time generalizing to a more diverse dataset. A solution to some of these problems would be to use open access datasets online for model training to ensure a more generalized model. Even so, the majority of the cataloged papers have their signal in terms of change in resistance, voltage, or capacitance; thus, there would need to be a conversion into common units in order to make use of any centralized database.

Recent advances in skin-interfaced pressure sensors have shown desirable response characteristics for continuous pressure monitoring. Future advancements seek to further improve the sensitivity and dynamic range, while maintaining adequate stability. Furthermore, there is significant research potential in studying the integration of these sensor materials into user-friendly mediums such as textiles. Lastly, despite the aforementioned shortcomings, integrating TSC models into pioneering skin-interfaced sensor systems proves to be extremely useful for rapidly processing data for human-machine interfaces. As this field of research matures, there exists a growing opportunity to establish standardized datasets and protocols to facilitate comparisons between models and sensing technologies.

Currently, there is a lack of centralized large scale public datasets for skin-interfaces pressure sensing. This arises from a combination of technical, regulatory, and domain specific challenges. From a technical standpoint, novel skin-interfaced pressure sensors can employ a diverse combination of mechanisms, materials, and device architectures. As a result, they produce vastly different and non-uniform signal formats, responses, units, and calibration protocols which ultimately hinders dataset aggregation and standardization. Additionally, unlike image pixels in ImageNet, pressure data is predominantly task specific time series data which often requires more complex preprocessing and labeling strategies (Bock et al., 2021). Moreover, the diversity of pressure sensing applications impedes the creation of a unifying benchmark task, in contrast to computer vision which was able to drive progress through standardized datasets and challenges (Russakovsky et al., 2015). Lastly, there exists a vast amount of regulatory constraints which further limit data sharing, as skin-interfaced pressure signals are often considered sensitive biomedical data subject to privacy regulations (Yadav et al., 2023).

Having a centralized database can accelerate the development of generalized models, while also identifying the most promising engineered materials and model combinations for specific biomedical applications. Looking forward, the continued convergence of integrated machine learning models with next-generation pressure sensors is encouraging. Ultimately, coordinated progress across material design and machine learning frameworks will hopefully improve real-time clinical insights directly from non-invasive wearable systems, further enabling responsive healthcare technologies.