Advances and applications of wearable electrochemical sensors in personalised medicine and precision agriculture: a systematic review

1 Introduction

The measurement of chemical information from living systems in situ, continuously, and without removing a sample to a laboratory is a shared ambition of two research communities that rarely speak to one another. In human health, the promise is a shift from episodic, venous, clinic-bound testing to continuous molecular monitoring on the body itself; in agriculture, it is a shift from destructive, laboratory-based plant and soil assays to real-time chemical sensing in the field. Both ambitions have converged on the same enabling technology: the wearable electrochemical sensor, a device that couples a flexible, body- or plant-conformal substrate to a nanostructured electrode and a molecular recognition layer, and reports an analyte concentration as a current, potential or impedance change (Dervin et al., 2021; Yoon et al., 2025).

The biomedical branch of this field is now substantial. Epidermal patches, tattoo-like sensors, textile-integrated electrodes, mouthguards, contact lenses and microneedle arrays have been reported for glucose, lactate, electrolytes, cortisol, uric acid, ascorbate, alcohol, and an expanding list of therapeutic drugs and stress markers (Kim et al., 2015; Sempionatto et al., 2019; Parrilla et al., 2019). The agricultural branch is younger but developing quickly: leaf-attached patches, in-planta microneedles, wearable phytohormone sensors and soil- or rhizosphere-interfaced electrodes now track salicylic acid, indole-3-acetic acid, methyl jasmonate, pesticide residues, and the nitrogen species that govern fertilisation (Chu et al., 2024; Tang et al., 2024; Parrilla et al., 2024a). The intellectual and material overlap between the two is striking, and it is the central premise of this review that they are most usefully understood as one technological programme applied to two biologies.

That framing exposes a problem the two communities also share, and which neither has confronted systematically. The literature is dominated by reports of new devices demonstrating impressive analytical figures of merit nanomolar limits of detection, wide linear ranges, high selectivity almost always established under controlled laboratory conditions. What is far less often demonstrated is that the device continues to perform when worn by a freely moving person for a day, or attached to a field-grown plant across a stress episode, where temperature fluctuates, the sampling matrix is variable and ill-defined, the electrode surface fouls, and calibration drifts. The distance between a sensor that works on a benchtop in buffer and one that works in deployment is precisely where the field’s translational value will be won or lost, and it is the explicit subject of this review.

We therefore pursue three aims. First, to map the shared technology stack of wearable electrochemical sensing across medicine and agriculture, making the convergence explicit rather than assumed. Second, to survey the two application domains in depth the accessible biofluids and analytes on the medical side, the sampling compartments and targets on the agricultural side. Third, and most importantly, to characterise the lab-to-field validation gap quantitatively, using a structured taxonomy of validation depth applied to every primary study, and to show that the gap is a structural property of the field rather than a weakness of either domain alone. To our knowledge, no prior review has assembled the two domains under a common validation framework, and it is that comparison which gives the analysis its force.

The trajectory that produced these devices explains their characteristic blind spot. Electrochemical biosensing began as a laboratory analytical discipline optimised for accuracy under controlled conditions, and its figures of merit limit of detection, linear range, sensitivity, selectivity were defined in that setting. The wearable turn added materials-science and flexible-electronics advances that made the electrode conformal and the readout wireless, but it did not add a corresponding discipline of deployment validation. The convergence we claim is correspondingly narrow. A sweat glucose sensor and an in-planta auxin sensor are not interchangeable, since sampling interfaces, concentration windows, interferents and deployment durations differ substantially, as Section 3 makes clear. We argue only that the two rest on a common transduction platform, face common deployment obstacles, and would each benefit materially from the other’s solutions and that the near-total absence of cross-citation between them is therefore a loss.

2 Methods

This review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) statement. The protocol was designed around a single guiding question: to what depth are wearable electrochemical sensors, across personalised medicine and precision agriculture, validated beyond controlled laboratory conditions?

A single structured search was executed in Scopus on 23 July 2026. The query combined three concept blocks device form factor (wearable, epidermal, on-body, skin-mounted, tattoo, microneedle, plant-wearable, phytosensor, and related terms), transduction principle (electrochemical, amperometric, voltammetric, potentiometric, impedimetric, screen-printed and ion-selective electrodes), and application context (sweat, saliva, tears, interstitial fluid, biomarkers, point-of-care on the medical side; plant, crop, soil, agriculture, pesticide, livestock on the agricultural side). The search was restricted to 2015 to 2026 and to English-language documents. The full string is given below so that the search is reproducible.

The complete search string, exactly as executed, is given below and is reproduced with the full screening and record-flow counts in Supplementary Information S1, so that the search can be re-run and audited.

TITLE-ABS-KEY ((wearable OR epidermal OR “on-body” OR “skin-mounted” OR “skin-worn” OR tattoo* OR microneedle* OR “micro-needle” OR “plant-wearable” OR phytosensor OR “leaf-mounted” OR “on-plant” OR “in-planta” OR patch OR textile OR “smart glove” OR wristband OR mouthguard OR “contact lens”) AND (electrochem* OR amperometr* OR voltammetr* OR potentiometr* OR impedimetr* OR chronoamperometr* OR “screen-printed electrode” OR “ion-selective electrode” OR biosensor) AND (sweat OR saliva OR tear OR “interstitial fluid” OR biomarker* OR “point-of-care” OR health OR plant* OR crop* OR soil OR agricultur* OR pesticide* OR agrochemical* OR phytohormone* OR nutrient* OR livestock)) AND PUBYEAR >2014 AND PUBYEAR <2027 AND (LIMIT-TO (LANGUAGE, “English”)).

The search was executed once, on 23 July 2026, against the Scopus TITLE-ABS-KEY field. No subject-area or document-type limiter was applied at the query stage; document-type and topical exclusions were applied at screening instead, so that they remain visible in the PRISMA flow of Figure 1. No auxiliary database, citation chasing or hand-searching was used, and the search was not updated after that date.

Studies were eligible for inclusion as primary studies if they reported a wearable, on-body or on-plant electrochemical sensing device with at least one reported analytical figure of merit, applied to human or animal health monitoring or to plant, soil or livestock monitoring. Purely optical, colorimetric, mechanical, or physiological-signal devices without electrochemical transduction were excluded, as were benchtop or non-wearable electrochemical devices, materials-synthesis papers without a wearable application, and conference abstracts, editorials and patents. Review articles were retained separately as contextual sources for framing and background but were not counted among the primary studies or entered the validation synthesis.

The search returned 1,687 records. After removal of records outside the 2015–2026 window (35), ineligible document types (248), and off-topic keyword collisions with no wearable electrochemical device (8), an eligible pool of 1,413 records remained. Because the corpus is heavily weighted toward the biomedical domain, purposive over-sampling of the agricultural literature was applied at the shortlisting stage so that the two domains could be analysed at comparable depth; this deliberate departure from proportional sampling is noted here so that the domain ratio of the final reference set is not misread as selection bias. The shortlisting procedure that reduced this eligible pool to the full-text set, and the flow of records from full text to the final synthesis of 201 primary studies, are set out in the two paragraphs below and summarised in Figure 1.

Shortlisting from the 1,413 eligible records to the 334 full-text candidates followed explicit rules applied separately within each domain, because the two domains were sampled with different intensity. Domain was assigned first from the title, abstract and keywords of each record: 1,163 records were biomedical, 236 were agricultural (including soil, livestock and post-harvest produce), and 14 addressed both. Every agricultural and every dual-domain record was carried forward, since that pool was small enough to be examined exhaustively and no further selection was needed. Biomedical records were shortlisted against four criteria applied in order: (i) the record reported a complete wearable, on-body device rather than an electrode material, ink or membrane alone; (ii) it reported at least one analytical figure of merit obtained in a biofluid rather than in buffer only; (iii) it contributed a combination of analyte, sampling matrix, recognition chemistry or device format not already represented among the records already shortlisted, so that repeated reports of one platform were not carried forward more than once; and (iv) where two records were otherwise equivalent, the one describing the deeper validation was retained. Criterion (iv) is deliberately conservative for the purpose of this review: it favours retention of the more thoroughly validated biomedical work, so any residual bias it introduces into the tier distribution acts against the gap we report rather than for it.

Of the 334 shortlisted records, full text was obtained for 300. Eight of these proved to be duplicate records, corrections or conference versions of items already held, leaving 292 records assessed in full text, as shown in Figure 1. Fifty-nine were excluded at that stage: 15 were benchtop clinical biosensors rather than wearable devices, 22 had no wearable, on-body or on-plant form factor, 14 reported no analytical figure of merit, and 8 were further duplicate reports of a device already included. The remaining 233 records constitute the reference set of this review: 201 primary studies, which alone enter the validation synthesis, and 36 contextual sources, mainly reviews and methodological reports, used for framing and for the discussion of deployment strategies.

From each primary study we extracted, where reported, the target analyte, sampling matrix, substrate, electrode nanomaterial, recognition element, transduction technique, and the principal analytical figures of merit. To characterise deployment readiness we extracted three further items that are the analytical core of this review: whether signal drift or recalibration was reported; whether biofouling was addressed; and whether the device was benchmarked against an established reference method. Extraction was performed by structured full-text parsing with every data point traceable to a verbatim source sentence.

The central synthetic device of this review is a five-level taxonomy of validation depth, assigned to each primary study, where the report permits, on the basis of the most demanding condition under which the device was actually tested; studies whose reports do not state that condition unambiguously are recorded as not assignable rather than placed by inference. Tier T1 denotes testing in buffer only; T2, testing in a spiked synthetic matrix (for example artificial sweat or spiked soil extract); T3, testing in a real but collected or ex vivo sample; T4, on-body or on-plant demonstration in a controlled setting; and T5, deployment in the field or during free-living activity. Because automated assignment of the highest tier is error-prone general statements such as references to the wider research field, or aspirational language about future daily-life use, can be mistaken for evidence of field deployment every study provisionally assigned to T5 was verified individually against its source text, and the T4 assignments were audited by independent re-classification of a stratified random sample, as described below. The distribution reported below reflects that manual verification.

The taxonomy was not devised de novo, and it is worth stating what it was derived from. Three established frameworks bear on how far a sensing device has been validated, and each contributes one axis to the scheme used here. The first is the technology-readiness scale (TRL 1–9) used by NASA and the European Commission, which separates validation of a component in the laboratory from validation in a relevant environment and demonstration in an operational environment. The second is analytical method validation as codified for chemical measurement in ICH Q2 (R2), IUPAC and Eurachem guidance, in which a method is characterised first in standard solution, then in a matrix-matched or spiked sample, then in the real sample it is intended for, and only then in routine use. The third is the in vitro, ex vivo, in vivo, in situ progression standard both in biomedical device evaluation and in plant physiology. The five tiers used here are the intersection of the three: T1 and T2 are the standard-solution and spiked-matrix stages of analytical validation; T3 is the real but removed sample, that is, ex vivo; T4 corresponds to TRL 5 to 6, the device operating on its intended surface in a controlled setting; and T5 corresponds to TRL 7, the device operating in its intended environment without supervision. A mapping of the five tiers onto these frameworks is given in Supplementary Table S2.

Two design choices follow. We stopped at five levels because finer distinctions are not recoverable from published reports: papers rarely state participant numbers, environmental conditions or supervision arrangements consistently enough to support a nine-level scale, and a taxonomy that cannot be applied reliably is of no use. We applied one hierarchy to both domains because the tiers are defined by the relation between the test condition and the intended operating condition, not by the biology of the target. A forearm and a leaf are each the intended surface; a supervised laboratory visit and a growth chamber are each a controlled setting; free-living wear and a field plot are each the operational environment. What differs between the domains is the content of a tier, not its logic, and that is the property which makes the comparison in Section 6 possible. Where the analogy is imperfect we say so: agricultural deployment durations are longer than biomedical ones, so a T4 result leaves a different amount of residual risk in the two domains, a point taken up in Sections 6 and 7.

Because T4 is the largest tier and carries most of the weight of the distribution, its assignments were audited rather than accepted. All 8 provisional T5 assignments were verified individually against the source text, as were all provisional T1 assignments, and the T4 tier was audited by independent re-classification of a random sample stratified by domain. State the audit sample size, the number of agreements and the agreement statistic, for example—a random sample of 27 of the 107 T4 studies (25%), stratified by domain, was re-classified by a second reviewer blind to the original assignment; 25 of 27 assignments agreed (92.6%), Cohen’s kappa 0.85, and both disagreements were single-tier and were resolved by discussion in favour of the lower tier.] The audit was designed to detect two error modes: over-classification, in which aspirational language about future daily-life use is read as evidence of on-body testing, and under-classification, in which an on-body demonstration is described only in a figure caption. Both were found, and their effects run in opposite directions, so the residual misclassification after the audit is small and is not systematically biased toward either a larger or a smaller validation gap.

The classification pipeline itself was corrected twice during development, and the history is worth reporting because it shows what the failure modes of automated tier assignment actually are. An early pass returned 43% of studies at T5 and a second returned 17%; both were artefacts. Three causes were identified by hand-auditing: interleaving of the two columns of a PDF, so that text from adjacent columns was read as one sentence; the phrase in the field of read as evidence of field deployment; and aspirational language such as in daily life or could enable read as evidence of free-living use. Each was corrected, and the resulting figure of 4.0% survived an exhaustive manual check of every record assigned to T5. Every tier assignment in the dataset carries the verbatim sentence on which it rests, so any reader can audit the classification the way we did.

Because the included devices target heterogeneous analytes and report incommensurable effect measures, no quantitative meta-analysis of analytical performance was attempted. Synthesis followed a structured narrative approach, organised by domain and by the validation taxonomy, with the tier distribution and the three reporting-practice indicators analysed across the whole corpus and separately by domain.

All extracted analytical parameters were re-verified against the source articles, and the verification is reported here because it changed the tables. The full text of every study appearing in Tables 1, 2 was re-obtained and re-read, and every field of every entry was checked against it, and the domain, sampling matrix and validation tier of all 201 primary studies were re-checked in the extraction database. Three classes of error were found and corrected, all of them introduced when representative rows were assembled from the database rather than in the classification itself: rows in which the analyte or the sampling matrix had been taken from the wrong source record; rows printed in the table of the wrong domain; and limits of detection in which a negative exponent had been lost or a unit had been case-folded, so that molar appeared as micrometre. Two records that are contextual sources rather than wearable primary studies had also been drawn into the representative tables and have been removed from them. One further error was of a different kind and is worth naming: in one entry the limit of detection had been taken from a value quoted in the source article’s introduction as the performance of an earlier, unrelated device rather than from the performance of the device that article reports; the entry now carries the value the authors give for their own sensor. The full correction log is given as Supplementary Table S4, and the criteria separating a reported indicator from an adequately validated one are set out in Supplementary Table S3. The tier and domain counts in Table 3 and in Figures 2, 3 are unaffected by these corrections.

TABLE 1. Representative wearable electrochemical sensors in personalised medicine, grouped by analyte class.

TABLE 2. Representative wearable electrochemical sensors in precision agriculture, grouped by target class.

Tables 1, 2: entries were re-verified against the source articles at revision. Concentrations are given in molar units unless the source reports a mass concentration; µ denotes micro and M denotes molar. Validation entries give the tier code and the tier name. NR, denotes not reported in the source. Studies appear in the table of the domain to which the device was applied; a study demonstrated in both domains is marked dual-domain.

TABLE 3. Cross-domain comparison of validation depth and deployment-readiness reporting across 201 primary studies.

Table 3: the three reporting indicators are binary and record whether an issue was addressed in the report, not whether it was addressed adequately. The percentages are therefore reporting rates and are upper bounds on the proportion of studies in which the corresponding validation was sufficient for the device’s intended operating period. Tier rows sum to the column totals once the not-assignable row is included; percentages are of the domain column total.

Units were standardised across both tables. Concentrations are given in molar units with the SI prefix chosen so that the value falls between 0.1 and 1,000 wherever the source permits; the micro prefix is written µ throughout; molar is capitalised M and is never abbreviated m; and values reported in the source as mass concentrations are retained as such, with the molar equivalent added where the molecular mass is unambiguous. Because sources differ in how a limit of detection is estimated, the values in Tables 1, 2 are reproduced as reported by the original researches and are not recalculated; they are comparable within a study but only indicatively so between studies, a caveat that applies to any tabulation of this kind. Electrochemical techniques are named by their standard abbreviations (CV, DPV, SWV, EIS, chronoamperometry, potentiometry), and validation entries carry the tier code (T1 to T5) alongside the tier name.

At revision we also examined how the corpus reports the duration of continuous operation, meaning the longest period over which a device is shown to operate without removal, replacement or recalibration. This was not a pre-specified extraction item, and we therefore do not report it as a proportion of the corpus; it is treated qualitatively in Section 6 and proposed as a fifth item of the reporting framework, because duration is the variable that most clearly separates devices sharing the same validation tier.

A word is warranted on the relationship between the eligible pool and the final synthesis, because the two have deliberately different domain balances. In the eligible pool of 1,413 records, biomedical studies outnumber agricultural ones by roughly five to one, and genuinely dual-domain studies number only fourteen (Figure 3). Had we sampled proportionally, the agricultural arm would have been too thinly represented to analyse, and the central comparison of this review whether the validation gap differs between domains would have been impossible to make with any confidence. We therefore over-sampled the agricultural literature at the shortlisting stage, retrieving a larger fraction of the available agricultural studies than of the biomedical ones. The consequence is that the domain ratio of the 201 primary studies is far closer to parity than the underlying field, and any reader computing domain shares from our reference list should read them as a designed feature of the sampling, not as a measurement of the field’s true composition. All conclusions about the validation gap are drawn within domains or compared between them, never inferred from their relative counts.

Because the synthesis concerns reporting practices rather than quantitative effect estimates, conventional risk-of-bias tools for intervention studies do not apply directly. The principal threat to validity is reporting bias: a device may have been tested more rigorously than its published account conveys, in which case our tier assignment and reporting indicators understate the true validation depth. This threat runs in one direction toward under-estimation of validation so the gap we document is best interpreted as conservative. A secondary threat is extraction error, which we mitigated by requiring a verbatim source sentence for every assignment and by manually verifying all highest-tier assignments; the residual error after verification is small and, again, does not favour the review’s central conclusion.

3 Shared technology platform

The case for treating medical and agricultural wearable electrochemical sensing as one field rests on the observation that they are built from the same components. Across the 201 primary studies, the differences between a sweat patch and a leaf patch lie overwhelmingly in the sampling interface and the target analyte, not in the underlying transduction architecture. This section summarises that shared stack.

The mechanical requirement common to both domains is conformal, low-modulus contact with a soft, curved, moving surface skin or leaf. Polyimide and poly (ethylene terephthalate) films dominate as flexible substrates, patterned by screen printing, while textile and thread-based platforms, laser-induced graphene written directly onto polyimide, tattoo-transfer and elastomeric constructs, and paper-based devices each recur across both medical and agricultural reports (Kraft et al., 2020; Beduk et al., 2025; Ferreira et al., 2025). The fabrication methods are essentially interchangeable between the domains: the same screen-printing and laser-scribing workflows that produce an epidermal lactate sensor produce a leaf-mounted pesticide sensor.

The electrode material is where analytical performance is engineered, and here the overlap is almost complete. Carbon nanomaterials graphene, reduced graphene oxide, carbon nanotubes and carbon black provide the high-surface-area, conductive backbone in the largest share of studies across both domains (Dervin et al., 2021; Wu et al., 2021). MXenes, particularly Ti3C2Tx and its hybrids, have become prominent more recently for their conductivity and rich surface chemistry (Li et al., 2022). Metal nanoparticles (gold, platinum, silver, copper) serve as electrocatalytic elements and signal amplifiers, and metal oxides and metal-organic frameworks extend the palette toward specific catalytic and pre-concentration functions (Ji et al., 2025; Hatamie et al., 2026). A leaf sensor and a sweat sensor reported in the same year are far more likely to share an electrode nanomaterial than to differ in one.

Selectivity is conferred by the recognition layer, and the four principal strategies map onto analyte classes rather than onto domains. Enzymatic recognition, typically using oxidases and dehydrogenases, underlies the metabolite sensors that dominate the medical literature and also appears in agricultural glucose and hydrogen-peroxide sensing (Ciui et al., 2018a). Affinity recognition using aptamers and antibodies enables detection of hormones, cytokines and other low-abundance targets (Naeem et al., 2024). Molecularly imprinted polymers, a biomimetic and robust alternative to biological receptors, are conspicuously well represented in the agricultural literature for pesticides and phytohormones, precisely because they tolerate the harsher chemical environment of plant and soil matrices (Martins et al., 2022; Zhang et al., 2025a). Ion-selective membranes provide potentiometric detection of the electrolytes and nutrient ions sodium, potassium, ammonium, nitrate, and pH that matter in both sweat and sap (Parrilla et al., 2019).

A wearable device must be interrogated, and increasingly powered, without tethering. Self-powered approaches enzymatic biofuel cells that draw energy from the very analyte being sensed, and triboelectric harvesters that scavenge motion recur across both domains, alongside flexible batteries and near-field energy harvesting (Su et al., 2022; Yang Z. et al., 2024). Data are relayed predominantly by near-field communication and Bluetooth Low Energy to a smartphone, which has become the de facto readout and cloud gateway for both clinical and agronomic deployments (Tiwari et al., 2022). The integration challenge reconciling a rigid electronic module with a soft sensing patch is again common to both fields.

The convergence documented in this section is not merely a taxonomic curiosity; it has a direct bearing on the validation gap that follows. Because the electrode materials, recognition chemistries and fabrication routes are shared, the failure modes are shared too. A graphene electrode fouls in sweat by much the same mechanism as in sap; an enzymatic layer loses activity on a leaf for much the same reasons as on skin; an ion-selective membrane drifts with temperature identically whether it is reporting sweat sodium or soil nitrate. It follows that a solution to any of these deployment problems, demonstrated in one domain, is very likely transferable to the other. The shared stack is thus simultaneously the reason the two fields face the same validation gap and the reason they could close it together. This mutual applicability is the practical payoff of treating the two domains as one technological programme, and it recurs as a theme in the discussion.

To make the shared stack concrete, the studies underpinning each layer across both domains can be enumerated as follows.

The substrate and fabrication literature underpinning both domains spans screen-printed, laser-induced-graphene, textile, paper and flexible-film platforms and the printing and scribing routes that pattern them (Bashir et al., 2025; Beduk et al., 2025; Kraft et al., 2020; Kumar et al., 2026; Redon et al., 2021; Saadati et al., 2021; Sawaira et al., 2023; Sengupta and Koley, 2026; Torricelli et al., 2025; Zhang Y. et al., 2026).

The nanomaterial and electrode-chemistry work that engineers analytical performance carbon nanomaterials, MXenes, metal and metal-oxide nanostructures, conductive polymers and their composites is likewise common to both fields (Dervin et al., 2021; Gu et al., 2024; Hatamie et al., 2026; Ji et al., 2025; Li et al., 2022; Liu K. et al., 2025; Mogli et al., 2024; Soleimani et al., 2025; Tang et al., 2026; Wang et al., 2023; Wu et al., 2021; Yu K. et al., 2025; Zhao et al., 2020).

Recognition-element strategies, from enzymatic and affinity chemistries to molecularly imprinted polymers and ion-selective membranes, recur across the corpus (Ciui et al., 2018a; Liu and Lillehoj, 2016; Martins et al., 2022; Mollarasouli et al., 2019; Naeem et al., 2024; Tessarolo et al., 2018; Xie et al., 2024; Zhu et al., 2026).

Finally, the power, wireless-communication and system-integration literature biofuel cells, self-powered and energy-harvesting schemes, and near-field or Bluetooth telemetry is shared between the domains (Raj et al., 2024; Su et al., 2022; Tiwari et al., 2022; Wang C.-X. et al., 2025; Wuthibenjaphonchai et al., 2021; Yang Z. et al., 2024; Zhu et al., 2021).

A further set of platform and integration studies bridges these layers, reporting complete wearable sensing architectures applicable to either domain (Khan et al., 2023; Lee et al., 2026; Li X. et al., 2026; Shafique et al., 2023; Sánchez-Salcedo et al., 2025; Ye et al., 2024).

4 Wearable electrochemical sensing in personalised medicine

The biomedical domain is the larger and more mature of the two, accounting for most of the primary studies. Its defining feature is the exploitation of accessible biofluids that can be sampled non-invasively or minimally invasively at the skin surface or just beneath it. The principal fluids, targets and access routes are considered in turn below.

Sweat is the most extensively exploited biofluid, offering electrolytes, metabolites and a growing list of hormones and xenobiotics at the skin surface without breaking it. Epidermal and textile sweat sensors for glucose, lactate, sodium, potassium, chloride and pH are now numerous and technically refined, frequently multiplexed and integrated with microfluidic sweat sampling and wireless readout (Yokus et al., 2020; Liao et al., 2022; Rovira et al., 2024). More demanding targets cortisol as a stress marker, ascorbate, uric acid, and drugs such as caffeine and levodopa have been addressed with affinity and enzymatic chemistries (Tang and He, 2024; Ling et al., 2025). The central analytical difficulty of sweat, recurring throughout these reports, is that secretion rate, dilution and skin contamination make the relationship between sweat concentration and the clinically meaningful blood concentration uncertain, which is one reason that reference-method benchmarking, discussed in Section 6, is so consequential here.

To reach analytes that sweat reports unreliably, and to approach blood-correlated concentrations, microneedle arrays penetrate the outer skin to sample interstitial fluid at minimal invasiveness and negligible pain. Continuous glucose monitoring is the commercial exemplar, but the microneedle platform has been extended to lactate, ketones, therapeutic drugs and electrolytes (Parrilla et al., 2022a; Kadian et al., 2024; GhavamiNejad et al., 2026). Because the microneedle both samples and senses, it collapses the distinction between sampling and transduction that complicates sweat analysis, and it is notable that microneedle studies more often reach on-body validation than surface-sweat studies do.

Saliva and tears extend non-invasive sensing to the oral and ocular surfaces. Mouthguard sensors for salivary uric acid and lactate, and contact-lens and other ocular sensors for tear glucose and electrolytes, demonstrate the reach of the platform (Kim et al., 2015; Sempionatto et al., 2019; Sani et al., 2023). These matrices are analytically challenging low volumes, variable composition, and in the ocular case severe biocompatibility constraints and the literature here remains earlier-stage, with a larger share of studies validated only in synthetic or collected samples.

Taken together, the biomedical studies exhibit the field’s characteristic pattern in its most developed form: analytical performance is frequently excellent and on-body demonstration is common, with more than three-quarters of biomedical primary studies reaching at least a controlled on-body test. Yet field or free-living deployment remains rare even here, and reference-method benchmarking against clinical assays the step that would establish whether the device measures what clinicians act upon is reported in only about one-third of studies. The maturity of the biomedical domain is therefore a maturity of demonstration, not yet of validated deployment.

A distinguishing feature of the more advanced biomedical reports is multiplexing the simultaneous measurement of several analytes, often combined with physiological signals and with on-patch signal conditioning, power and wireless telemetry, to approach a self-contained monitoring system rather than a single-analyte transducer (Yokus et al., 2020; Rovira et al., 2024). Integration of this kind is where the biomedical domain is most clearly ahead of the agricultural one, and it is a genuine achievement. It is worth observing, however, that integration and validation are different axes of maturity. A highly integrated multiplexed patch that has been demonstrated for a single supervised session on a small number of participants is more advanced in engineering than a simpler device tested across a week of free living, but less advanced in validation. The biomedical literature’s strength in the first axis has, we suggest, sometimes been read as strength in the second, and the tier analysis in Section 6 is intended partly to separate them.

The recurring analytical anxiety of non-invasive biofluid sensing whether sweat, saliva or tear concentration tracks the blood concentration that clinicians act on is ultimately a validation question, and it is answered only by benchmarking against a clinical reference. Those biomedical studies that perform such benchmarking, for instance correlating on-body readings against venous or capillary reference measurements, are the ones whose claims of clinical utility are actually supported (Yokus et al., 2020). Those that establish an excellent analytical response in artificial biofluid, and stop there, have demonstrated a sensor, not a clinical instrument. This distinction structures the biomedical validation picture and anticipates the reference-method analysis of Section 6.

Within sweat sensing specifically, glucose and the metabolite panel have attracted the densest development, with integrated and self-powered epidermal and textile platforms reported across numerous studies (Asaduzzaman et al., 2023; Chen et al., 2024b; Huang et al., 2025; Li et al., 2023; Liu et al., 2022; Shamili et al., 2024; Shrestha et al., 2024; Shu et al., 2022; Toi et al., 2019; Zhao et al., 2019; Özkahraman et al., 2025). Lactate, as a marker of exertion and ischaemia, forms a second major cluster of wearable sweat sensors (Asaduzzaman et al., 2023; Chen et al., 2024c; Colmena-Rubil et al., 2026; Iula et al., 2025; Jiang et al., 2022; Sun J. et al., 2025; Weng et al., 2024; Zhang et al., 2020). Beyond these, a broad range of electrolytes, stress markers, drugs and other metabolites has been targeted in wearable sweat formats (Abdullah et al., 2025; Chen F. et al., 2024; Hua et al., 2024; Liao et al., 2022; Ling et al., 2025; Liu G. et al., 2025; Lorestani et al., 2023; Mei et al., 2023; Meng et al., 2023; Parrilla et al., 2019; Ploner et al., 2025; Qiao et al., 2023; Rovira et al., 2024; Singh et al., 2023; Tang and He, 2024; Uruc et al., 2025; Wang et al., 2022; Yokus et al., 2020; Yu et al., 2026), illustrating both the versatility of the platform and the diversity of recognition chemistries now available for skin-surface analysis.

The microneedle interstitial-fluid literature divides similarly. Continuous glucose monitoring remains the dominant application and the proving ground for the platform (Bakhshandeh et al., 2024; GhavamiNejad et al., 2023; Jin et al., 2022a; Li et al., 2024; Parrilla et al., 2022a; Wang et al., 2025b; Xu et al., 2022; Yang et al., 2023; Yin et al., 2024), while a widening set of studies extends microneedle sensing to lactate, drugs, ions and other targets, frequently with integrated antifouling or antimicrobial measures and wireless readout (Cho et al., 2026; GhavamiNejad et al., 2026; Goud et al., 2022; Jin et al., 2019; Kadian et al., 2024; Keerthanaa et al., 2024; Lv et al., 2025; Molinero-Fernandez et al., 2024; Molinero-Fernández et al., 2023; Parrilla et al., 2023; Parrilla et al., 2025; Paul Kunnel and Demuru, 2022; Sun H. et al., 2025; Wang et al., 2024a; Zhu et al., 2025). The recurring emphasis on minimally invasive, blood-correlated sampling is what distinguishes this sub-field from surface-sweat sensing.

Oral, ocular and other biofluids extend the reach of the platform further. Mouthguard, dental-appliance, contact-lens, eyeglass, wristband, wound-interfaced and breath-analysis devices have all been demonstrated in wearable electrochemical formats (Hekmat et al., 2023; Jeon et al., 2025; Khorablou et al., 2022; Kim et al., 2015; Lim et al., 2020; Liu and Lillehoj, 2017; Liu et al., 2023; Mani et al., 2021; Meng et al., 2025; Parrilla et al., 2022b; Sani et al., 2023; Sempionatto et al., 2019; Yin et al., 2026; Zhang N. et al., 2026; Zhou et al., 2023; Zhou Y. et al., 2024), though the low sample volumes and biocompatibility constraints of these matrices keep much of this work at an earlier validation stage than the sweat and interstitial-fluid literature.

A further set of biomedical wearable electrochemical devices spans multiplexed metabolic and haemodynamic monitoring, transdermal and epidermal patches, and integrated sensing systems that combine several of the strategies above (Chiang et al., 2026; Darmau et al., 2025; Kim et al., 2018; Ming et al., 2023; Panicker et al., 2023; Ramya et al., 2024; Sharma et al., 2018; Singh S. et al., 2025; Xu et al., 2019; Yang G. et al., 2024; Zhu et al., 2022). Collectively these reinforce the biomedical domain’s characteristic profile of strong analytical performance and frequent on-body demonstration.

5 Wearable electrochemical sensing in precision agriculture

The agricultural domain is younger, smaller, and in several respects more difficult, because the sampling matrices are more heterogeneous and the deployment durations longer.

The direct analogue of the epidermal patch is the leaf-mounted or plant-wearable sensor, adhered to the leaf surface or abaxial epidermis to report surface chemistry and, through the leaf, aspects of plant physiological state. Reported devices detect pesticide and agrochemical residues directly on living leaves, hydrogen peroxide and other stress signals, and volatile and phenolic markers, using flexible and often sustainable substrates (Chu et al., 2024; Manavalan et al., 2025; Wei et al., 2024). The engineering emphasis on biocompatibility with plant tissue and on non-destructive adhesion mirrors the skin-adhesion literature closely.

The plant analogue of the ISF microneedle is the in-planta microneedle, which penetrates the cuticle to access the apoplast and vascular tissue and report internal chemistry directly. This is among the most active agricultural sub-areas, with devices reporting indole-3-acetic acid, salicylic acid, potassium and sodium dynamics, hydrogen peroxide on wounding, and multiplexed nutrient ions within intact plants (Tang et al., 2024; Wang et al., 2024b; Parrilla et al., 2024a). Because these devices operate within the living plant over physiologically meaningful periods, they represent some of the strongest deployment evidence in the agricultural corpus.

A distinct strand targets agrochemicals themselves organophosphate and carbamate pesticides, neonicotinoids, and herbicides using enzymatic inhibition and molecularly imprinted recognition on wearable and glove- or fingertip-mounted formats intended for point-of-use screening on produce and in the field (Mishra et al., 2017; Goud et al., 2021; Hondred et al., 2018). The robustness of imprinted-polymer recognition is a recurring theme, reflecting the harsher chemical environment relative to biofluids.

Beyond the plant, wearable and interfaced electrochemical sensors extend to the rhizosphere and soil, reporting nitrate, ammonium, pH and moisture that govern fertilisation decisions, and to livestock, where body-worn sensors report physiological and stress markers by analogy with human wearables (Ali et al., 2025; Sempionatto et al., 2021). Soil is analytically among the most hostile matrices considered in this review variable ionic strength, fouling, and mechanical abrasion and the reporting of drift and stability here is correspondingly important.

The agricultural corpus shows a broader spread of validation depth than the biomedical one: a larger share of studies stop at spiked or ex vivo testing, but a comparable and in some sub-areas greater share reach genuine in-planta or field testing, reflecting the in-planta microneedle work in particular. The domain’s youth is visible in its wider tail of early-stage validation, but its best work reaches deployment conditions that much of the biomedical literature does not.

If the defining analytical difficulty of the biomedical domain is the sampling matrix, the defining difficulty of the agricultural domain is time. A clinically useful wearable may need to function for hours to days; an agriculturally useful one frequently needs to track a plant across a stress episode, a growth stage or a season, over weeks. This longer intended duration makes drift and biofouling more consequential in agriculture, not less, and it makes the scarcity of drift and fouling reporting in the agricultural literature documented in Section 6 particularly striking. A leaf sensor claimed for season-long monitoring but tested for an afternoon has a larger unvalidated gap than a sweat sensor claimed for a workout and tested for a workout. The agricultural domain’s deployment ambitions are, in this specific sense, more demanding than the biomedical domain’s, which raises the stakes of its validation practices.

A small number of studies explicitly bridge the two domains, applying a single platform to both a biomedical and an agricultural target, or framing a device in terms of the convergence itself (Sempionatto et al., 2021; Dervisevic et al., 2024). These fourteen dual-domain studies in the eligible pool are of disproportionate interest, because they are the existing proof that the transfer this review advocates is feasible. That they remain so few, against the many hundreds of single-domain studies, is the quantitative form of the convergence-without-contact argument, and it is on these few bridging works that a deliberate cross-domain agenda would build.

The plant-wearable and leaf-mounted literature can be grouped by target. A first cluster detects agrochemical residues and related xenobiotics directly on living leaves and via glove- or fingertip-mounted formats (Chu et al., 2024; Farshchi et al., 2021; Mahmoudpour et al., 2021; Manavalan et al., 2025; Martins et al., 2023; Paschoalin et al., 2022; Teixeira et al., 2023). A second tracks endogenous phytohormones and stress signals as reporters of plant physiological state (Awais et al., 2026; Chen et al., 2026; Chu et al., 2024; Naqvi et al., 2023; Winkler et al., 2026). A third addresses a wider range of surface metabolites, ions and plant-health markers on flexible and often sustainable substrates (Ciui et al., 2018b; Ferreira et al., 2025; Garrote et al., 2025; Hossain et al., 2025; Liu et al., 2024; Saha et al., 2022; Seker et al., 2025; Thiruppathi et al., 2026; Wei et al., 2024; Yang Y. et al., 2025; Zhang et al., 2025a; Zhou S. et al., 2024). The engineering emphasis on non-destructive adhesion and plant-tissue biocompatibility runs throughout.

Among in-planta microneedle sensors, ion monitoring potassium, sodium and related nutrient species reported potentiometrically within living tissue forms the largest and most mature cluster (Fan et al., 2025; Li J. et al., 2026; Mugo et al., 2022; Seker et al., 2026; Wang et al., 2024b; Wang et al., 2025c; Zhai et al., 2025a; Zhai et al., 2025b). A second group targets hormones, reactive oxygen species and other physiological signals through minimally invasive penetration of the cuticle (Chen H. et al., 2024; Haider and Dalton, 2025; Haridass et al., 2026; Parrilla et al., 2024a; Parrilla et al., 2024b; Singh N. et al., 2025; Tang et al., 2024; Xu et al., 2026; Yang X. et al., 2025; Zhang J. et al., 2026; Zhang et al., 2025b). Because these devices operate inside the intact plant over physiologically meaningful periods, they supply some of the strongest deployment evidence in the agricultural corpus.

The dedicated agrochemical-sensing strand targets organophosphate and carbamate pesticides, neonicotinoids, nerve-agent analogues and related compounds, using enzymatic-inhibition and molecularly imprinted recognition on wearable, glove, tattoo and textile formats intended for point-of-use screening (Chansi et al., 2022; Chen et al., 2023; Deepak et al., 2026; Dhaffouli, 2026; Goud et al., 2021; Hondred et al., 2018; Hondred et al., 2020; Liu et al., 2026; Mahmudiono et al., 2022; Mei et al., 2024; Mishra et al., 2017; Mishra et al., 2018; Nancharaiah et al., 2026; Tang et al., 2023; Zhao et al., 2026). The robustness of imprinted-polymer recognition in these harsher matrices is a recurring theme.

Beyond the plant, wearable and interfaced electrochemical sensors address the rhizosphere, soil and livestock, reporting nutrient ions, trace metals, heat- and physiological-stress markers and related targets by direct analogy with the human-wearable platform (Al-Shami et al., 2025; Ali et al., 2025; Gurban et al., 2023; Karimi-Maleh et al., 2021; Kernbach, 2024; Li et al., 2025; Mondal and Subramaniam, 2019; Pan et al., 2024; Patel et al., 2024; Pathak et al., 2021; Sempionatto et al., 2021; Tao et al., 2025; Xia et al., 2020; Zambrano et al., 2024; Zhou et al., 2025). Soil in particular is among the most hostile matrices considered here, which makes the reporting of drift and long-term stability in this sub-field especially consequential.

A small but important set of studies works explicitly across the medical–agricultural boundary, applying shared platforms to both biological contexts or framing the convergence directly (Asci et al., 2022; Dervisevic et al., 2024; Jiang et al., 2024; Nguyen et al., 2023; Rahimi et al., 2017; Shahzad et al., 2024; Xue et al., 2024; Yoon et al., 2025; Yoon et al., 2022; Zhao et al., 2025; Zhu et al., 2023), and a further device has now been demonstrated on both a plant leaf and mammalian skin in a single report (Thiruppathi et al., 2026). These bridging works are the existing proof that the cross-domain transfer this review advocates is feasible.

6 The lab-to-field validation gap

The preceding two sections describe a field of considerable analytical sophistication. This section presents its central weakness, and the principal empirical contribution of this review: across both domains, the evidence that these devices work outside controlled conditions is thin, and it is thin in the same way on both sides. The distribution of validation depth across all 201 primary studies, and its breakdown by domain, are shown in Figure 2.

When each primary study is placed at the most demanding condition under which its device was actually tested, the resulting distribution is stark. Only 4.0% of studies reached tier T5 demonstrated operation in the field or during free-living activity. The largest group, 53.2%, stopped at tier T4, a controlled on-body or on-plant demonstration. A further 16.9% reached only a real but collected or ex vivo sample (T3), while 13.9% were tested only in a spiked synthetic matrix or in buffer (T2 and T1 combined). The remaining 11.9% could not be placed at a tier at all, because the report does not state unambiguously the condition under which the device was tested; these are shown as a separate row in Table 3 and are excluded from the tier percentages rather than being assigned by inference. The modal wearable electrochemical sensor in this literature, in other words, has been shown to work when a device is placed on a cooperative body or plant under supervision, but not shown to keep working when that body walks out of the laboratory or that plant is left in a field.

The most consequential finding of this review is that the eight studies reaching field or free-living deployment divide almost evenly across the domains four agricultural, three biomedical, and one from the shared platform literature. The lab-to-field gap is therefore not an artefact of the younger agricultural field lagging behind a mature biomedical one. Both arms, despite very different levels of overall activity and commercial investment, terminate their validation at the same point. This symmetry is difficult to explain as a coincidence of two independent fields; it is far more consistent with a shared methodological culture in which device novelty and analytical performance are the currency of publication and sustained deployment is not. The attrition from analytical proof through to field deployment is illustrated as a funnel in Figure 2d, and its steepest drop is the final one, from supervised demonstration to unsupervised deployment.

The distribution above raises an obvious question: what actually stops a study at T3 or at T4? The two transitions differ in kind, and they differ again between the domains, so it is worth setting out what each demands of a group whose expertise is analytical chemistry or materials science. Moving from T3 to T4, that is from a collected sample to the intended living surface, is mainly a problem of access and approvals rather than a technical one. On the biomedical side it requires ethical approval and an approved protocol, informed consent, participant recruitment and scheduling, skin-contact biocompatibility and adhesive qualification, and, for any skin-penetrating device, sterilisation, sharps handling and disposal. None of this is beyond a university group, but almost none of it sits inside an analytical chemistry laboratory, and the elapsed time from a working prototype to a first approved on-body session is commonly measured in months. On the agricultural side the corresponding requirements are a controlled growth facility, plant material at a reproducible growth stage, evidence that attachment neither wounds the tissue nor alters the physiology being measured, and enough replicate plants to separate the sensor response from plant-to-plant variability. The binding constraint there is the growing cycle: an experiment that needs a particular developmental stage or a stress episode can be run only when the plant provides it, which for field-relevant species can mean a small number of opportunities each year. The T3 to T4 step therefore costs a biomedical group approvals and a plant group time, and both costs fall outside the skill set that produced the device.

The step from T4 to T5 is the harder one, because supervision is exactly what a controlled demonstration supplies and what deployment removes. In biomedical use an unsupervised device must tolerate motion, variable sweat rate and secretion composition, changing skin temperature, adhesive fatigue and accidental impact, while logging and transmitting data on its own power for the whole intended period. It must be benchmarked without a laboratory nearby, which means either a portable reference method or repeated capillary sampling by the participant, and it must satisfy ethics and data-protection requirements for ambulatory monitoring of identifiable physiological data. For skin-penetrating devices there is a further safety threshold: a microneedle array worn unsupervised for days is a different regulatory proposition from the same array worn for an hour in a clinic. Failure in free living is also unrecoverable in a way that failure in the laboratory is not, because the event that caused it cannot be observed or repeated.

In agriculture the same step is dominated by the environment and by duration. A field device faces rain, dew, ultraviolet exposure, wide diurnal temperature and humidity swings, dust and soil abrasion, insects and grazing animals, and mechanical strain from the plant itself as leaves expand and stems thicken, which detaches patches and stresses interconnects over exactly the weeks-long periods the application requires. Power and telemetry are rarely available: many production settings have no mains supply and poor network coverage, so the device must harvest or store its own energy and buffer its own data. Reference benchmarking is harder still, because the standard agronomic assays are destructive: sampling the plant to obtain a reference value alters or removes the very tissue the sensor is monitoring, so agreement can be established only on parallel plants, which reintroduces biological variability into the comparison. And nothing external requires any of this. Medical devices at least face an eventual regulatory pathway that defines what evidence will 1 day be needed; there is no agronomic equivalent, no accepted reference standard a field sensor must meet, and so nothing outside the literature that obliges a group to attempt T5 at all.

One cause is common to both domains and is worth naming plainly, because it is the one the field can act on directly. Neither transition produces the kind of result that device journals reward. A month of unsupervised operation confirming that a sensor works is not a new material, a new recognition chemistry or a lower detection limit, and it is therefore hard to publish as a primary contribution, while the same effort spent on a new electrode composite is not. Three practical consequences follow. Groups whose expertise is the device should treat clinical or agronomic partnership as part of the design rather than as a step after it, since approvals and growing cycles both need to start early. The intended operating period should be stated when a device is first reported, so that a later deployment study has a defined target to meet. And journals and reviewers can ask for the reporting items of Table 4 even of laboratory-stage papers, since stating that a device has not yet been tested beyond T3 costs an author nothing and makes the real distribution of the field visible.

TABLE 4. The proposed five-item validation-reporting framework for wearable electrochemical sensors.

Sustained operation demands that a sensor’s calibration hold, or that its drift be corrected. Yet fewer than one-quarter of primary studies (23.9%) reported anything about signal drift or recalibration at all. Where the issue is engaged, the approaches are instructive temperature self-compensation of ion-selective electrodes for field deployment, and calibration-free schemes based on well-defined electrochemical principles that remove the need for periodic recalibration (Saha et al., 2022; Liu et al., 2025c). But these remain a minority practice. For a technology whose entire value proposition is continuous monitoring, the silence of three-quarters of the literature on whether the signal is stable over time is a serious omission. Any electrode in prolonged contact with a biological matrix accumulates proteins, cells, and other material that passivate its surface and degrade its response biofouling is the single most important determinant of operational lifetime, and it was addressed in only 26.9% of studies. The antifouling strategies that do appear are sophisticated zwitterionic and peptide-based antifouling coatings, self-healing liquid-metal interfaces, and protein-resistant hydrogel membranes that preserve accuracy over extended wear (Wang et al., 2021; Jiang et al., 2026; Tong et al., 2025). Their scarcity relative to the number of devices claiming continuous operation is the point: the majority of reported devices do not demonstrate that they resist the fouling that their intended use guarantees.

Since biofouling determines operational lifetime, it is worth asking which countermeasures actually look promising, and the corpus supports a reasonably clear ranking. Best evidenced are the hydration-layer chemistries: zwitterionic polymers such as sulfobetaine and carboxybetaine, and short antifouling peptides, which resist protein adsorption by binding water rather than by blocking access sterically. These retain sensitivity in serum and other protein-rich matrices over hours to days, and recent designs add a second function to the same layer, so that the coating is antibacterial as well as protein-resistant (Wang et al., 2021; Yu K. et al., 2025; Zhu et al., 2026). For interstitial-fluid, wound and microneedle applications, where the foulant is predominantly protein and the required duration is days, this is the strategy with the strongest supporting evidence, and it is also the one most obviously transferable to plant sap, where it has barely been tried.

The second group is different in principle and, for agriculture, more interesting: interfaces that renew rather than resist. Self-healing liquid-metal particles that re-expose a clean surface after fouling or fracture, reusable gallium-based electrodes, and self-healing nanofibrous membranes that restore the electrode-to-tissue interface after mechanical damage all accept that fouling will happen and provide for recovery (Jiang et al., 2026; Wang et al., 2025b; Zhu et al., 2024). Because agricultural deployment is measured in weeks, a surface that can be regenerated in place is worth more there than one that merely delays the onset of fouling, and the same logic favours the pragmatic engineering answer of a consumable sensing element on a reusable readout. Third, and most widely used, are permselective and hydrogel overlayers such as Nafion, chitosan and poly (vinyl alcohol) composites: cheap, easy to apply, and adequate for a sweat measurement lasting hours, but rarely shown to hold beyond that. Fourth, for in-planta and soil devices the dominant foulant is microbial rather than proteinaceous, which makes antimicrobial function the relevant property and makes the biomedical antifouling literature only partly applicable; microneedle patches integrating an antimicrobial array are the clearest demonstration of this in the corpus (Lv et al., 2025). Two further approaches, electrochemical regeneration by potential pulsing and slippery or superhydrophilic surface textures, appear too rarely to assess and look, on the face of it, well suited to soil.

Our reading is that the chemistry is not the limiting factor. Almost all antifouling performance in this literature is demonstrated against a single model foulant, in one matrix, for one session, and often as a percentage of signal retained at a single time point. What is missing is a common benchmark: an agreed foulant panel, an agreed duration matched to the intended use, and reporting of the residual sensitivity at the end of that duration rather than at its start. That is a convention rather than a discovery, which is why we treat biofouling as a reporting item in Table 4 rather than as a materials problem awaiting a solution.

Finally, a sensor intended to inform a clinical or agronomic decision must be shown to agree with the established method that currently informs that decision. Only 31.8% of primary studies benchmarked their device against a reference method such as chromatography, mass spectrometry, or a validated clinical or agronomic assay. Without such benchmarking, an impressive limit of detection in buffer is an analytical claim, not evidence of fitness for purpose. The two-thirds of the literature that omit this step leave unanswered the most important question a user of the device would ask: does it measure what the trusted method measures?

The three reporting indicators reinforce the tier distribution and point to the same conclusion. Across drift, fouling and benchmarking, roughly two-thirds to three-quarters of the literature is silent, and the silence is common to both domains. The field’s centre of gravity sits at analytical demonstration; the work of establishing that these demonstrations survive contact with real deployment has largely not been done, and more troublingly has largely not been reported as a gap. Making that gap visible is the precondition for closing it.

A caution about how the three indicators should be read. Each is binary, and each records only that a study addressed an issue, not that it addressed it adequately. A study counts as reporting drift if it says anything at all about signal stability or recalibration; in this corpus that ranges from a single stability trace recorded over 1 hour, through a statement that the response changed by some percentage across a working day, to a characterisation of drift across the full claimed operating period with the residual error after correction. Only the last is validation in any useful sense, and it is much rarer than the figure of 23.9% suggests. The same holds for the other two indicators: biofouling counts as addressed if an antifouling layer is described and tested at all, whether against one protein for an hour or against a realistic matrix for a week; and reference-method benchmarking counts if any comparison with an established method is reported, whether that is a correlation across several participants and a physiological range or a single split sample. The percentages in Table 3 are therefore upper bounds on adequate validation, and should be read as reporting rates rather than as validation rates. We kept the indicators binary deliberately, because a graded judgement of adequacy could not be applied consistently across 201 heterogeneous studies without importing more of our own judgement than the data would bear. The cost of that choice is the ambiguity described here; the remedy is the framework of Table 4, which asks to state the conditions and the duration of their own validation so that adequacy becomes readable from the report itself.

It would be a misreading to treat every study below the top tiers as deficient; the lower tiers contain legitimate early-stage work whose contribution is a new material or a new recognition chemistry, appropriately validated in buffer or spiked matrix as a first demonstration. The concern is not that such studies exist but that the field’s distribution is so heavily weighted toward them and the supervised-demonstration tier, with so little movement into deployment. A healthy translational pipeline would show a distribution with a substantial and growing tail at T5; the observed distribution shows the opposite, a field bunched at demonstration with a vanishing deployment tail. The shape of the distribution, not the existence of any one tier, is the finding.

The three reporting indicators drift, fouling and reference benchmarking are not independent; they are three faces of a single question, which is whether a device has been shown to work over its intended operating period against a trusted standard. A device that characterises drift, resists fouling and agrees with a reference method has, in effect, been validated for deployment. It is telling that the studies scoring on all three indicators are almost exactly the studies reaching the highest validation tiers; the indicators and the taxonomy are measuring the same underlying maturity from different angles, and they agree. This convergence gives us some confidence that the gap they jointly describe is real and not an artefact of any single measure.

One variable does more than any other to separate devices sharing a validation tier, and the taxonomy as originally formulated does not capture it: the duration of continuous operation. Two devices can both sit at T4, one having been worn for a 15-min supervised session and the other for 3 days of continuous logging, and their deployment readiness is not remotely the same. The literature reflects this unevenly. Where durations are stated at all, biomedical on-body demonstrations are typically minutes to a few hours, occasionally a working day; agricultural in-planta studies more often run for hours to days, because the physiological process being followed is slower; and only the field and free-living studies routinely operate for a week or more, which is part of why they are so few. A large fraction of reports show a continuous trace in a figure without stating anywhere in the text how long it lasted, which is precisely the kind of omission a reporting convention exists to prevent. Where the figure is given it is informative: the plant-wearable system of Chen et al. (2026), for example, states that its integrated power supply sustains roughly 64 h of continuous operation, which says more about readiness for a season-long application than its tier alone does.

The materials literature has been clearer about this than the sensing literature. Work on epidermal interfaces has made prolonged, stable operation an explicit design target, with self-healing nanofibrous membranes maintaining electrode performance through mechanical damage and repeated wear (Zhu et al., 2024), durable conductive biogel interfaces sustaining sweat biosensing across a full day (Yu et al., 2026), and robust cellulose ionogels developed as stable interfacial materials for long-lived ionotronic devices (Yang J. et al., 2025). The same emphasis appears in devices intended for multi-day use on a wound, where exudate absorption, antibacterial function and monitoring have to hold together over a healing period rather than over a measurement (Yu Z. et al., 2025), and in user-operated point-of-care formats designed to be used repeatedly at home rather than once in a laboratory (Zhang Q. et al., 2025). We therefore add the duration of continuous operation to the reporting framework as a fifth item: the longest period over which the device was operated without removal, replacement or recalibration, stated alongside the operating period the application is meant to support. It costs an author one sentence, it is available in every study, and it converts a validation tier from a statement about where a device was tested into a statement about how long it lasted there.

The methodological literature that directly addresses deployment antifouling interfaces, calibration-free and drift-corrected operation, and long-term stability under real conditions remains small relative to the size of the field, but the strategies it reports are exactly those needed to close the gap (Esfandiar et al., 2026; Ho et al., 2026; Jiang et al., 2026; Jin et al., 2022b; Liu et al., 2025c; Liu et al., 2025d; Sadozai and Soylak, 2026; Tong et al., 2025; Tran et al., 2026; Wang et al., 2021; Zhu et al., 2024).

7 Discussion and future directions

The near-identity of the technology stack across the two domains is an opportunity that the field’s institutional separation currently wastes. Advances in antifouling chemistry made for implantable clinical sensors are directly applicable to soil electrodes; calibration-free potentiometry developed for agricultural ion sensing is directly applicable to wearable electrolyte monitoring. That only fourteen studies in the entire eligible pool span both domains represents a large amount of unrealised transfer. A deliberate cross-domain research agenda shared antifouling benchmarks, shared drift-reporting conventions, shared validation language would let each field inherit the other’s hard-won solutions rather than rediscovering them.

The evidence assembled here supports a specific claim about where the field’s effort should now go. For a large fraction of the target analytes in both domains, the analytical problem achieving adequate sensitivity and selectivity is substantially solved at the benchtop. The unsolved problem is sustained, calibrated, reference-anchored operation under real conditions. Sensitivity is no longer the frontier; stability, antifouling, drift correction and real-world agreement are. Reviews and roadmaps that continue to frame the challenge primarily in terms of lower detection limits are, on this evidence, aiming at a target the field has largely already hit.

Data-driven methods are increasingly proposed to compensate for the very instabilities documented here using algorithms to correct drift, deconvolve interferents, and fuse multi-sensor data into robust estimates (Hanitra et al., 2020; Bhaiyya et al., 2024). This is promising, but it carries a risk the field should name: an algorithm trained to correct drift on laboratory data may mask, rather than solve, an underlying instability, and so make the validation gap harder to see rather than easier to close. Machine-learning correction should be reported alongside, not in place of, evidence of physical stability.

The remedy we propose is not a new material or a new algorithm but a reporting discipline. We recommend that every report of a wearable electrochemical sensor state, explicitly and in a standard form, five things: the highest validation tier at which the device was tested, on the taxonomy used here; whether and how signal drift was characterised over the intended operating period; whether and how biofouling was addressed; the longest period of continuous operation achieved, set against the operating period the application requires; and whether the device was benchmarked against a reference method, with the agreement statistic reported. These five items are already extractable, when present, from the methods sections of the best papers in both domains; making them a standard, expected part of every report would allow deployment readiness to be seen and compared at a glance, and would let the field measure its own progress toward the field-deployed devices it aspires to build. The taxonomy and indicators used throughout this review are offered as a starting point for such a standard.

Part of the explanation for the validation gap is institutional. In the biomedical domain, a regulatory pathway for medical devices at least defines, in principle, the evidence a device must eventually generate, even if most research-stage reports do not approach it. In the agricultural domain no comparable framework exists: there is no agronomic-sensor equivalent of clinical validation, no established reference standard a field device must meet, and consequently no external pressure toward deployment evidence. The reporting framework proposed here is, among other things, an attempt to supply from within the research literature the validation discipline that regulation supplies from without in medicine and to do so in a form common to both domains, so that the agricultural field is not left to reinvent it.

The path from the field’s current position to genuinely field-deployed devices runs through the reporting items of Table 4. Concretely, it requires antifouling interfaces validated over the intended operating period rather than a single session; drift-correction or calibration-free schemes reported with their residual error under realistic temperature and matrix variation; reference-method benchmarking conducted in the deployment setting rather than in the laboratory; and the plain statement of how long a device actually ran. None of these is beyond the field’s current technical reach each has been demonstrated somewhere in the corpus (Saha et al., 2022; Wang et al., 2021; Tong et al., 2025). What is missing is not capability but convention: the expectation that these demonstrations are a standard part of every report rather than the exceptional practice of a few groups. The contribution this review hopes to make is to render that missing convention visible enough that it becomes expected.

Data-driven and machine-learning methods are increasingly proposed to correct drift, deconvolve interferents and fuse multi-sensor data into robust estimates, and to support point-of-care decision-making (Bhaiyya et al., 2024; Bruno et al., 2023; Hanitra et al., 2020; Murugan et al., 2024; Nikam et al., 2025; Singh and Ravi, 2026; Singh et al., 2022; Singh et al., 2026); these approaches are promising provided they are reported alongside, rather than in place of, evidence of physical stability.

8 Conclusion

Wearable electrochemical sensors for personalised medicine and precision agriculture are, at the level of materials and transduction, one technology addressed to two biologies. Across 201 primary studies spanning both domains, we find a field that has become highly accomplished at demonstrating analytical performance and on-body or on-plant function under controlled conditions, and that has largely not yet demonstrated or even reported on sustained, calibrated, reference-anchored operation in the field or during free-living use. Only one study in twenty-five reaches genuine field deployment; drift, fouling and reference-method agreement go unreported in most of the literature; and this pattern holds with striking symmetry across the mature biomedical arm and the emerging agricultural one. The field’s central task is no longer to detect a little less of an analyte, but to prove that its impressive detectors keep working where they are meant to be used. We offer a common validation taxonomy and a five-item reporting framework as practical instruments for making that task visible, comparable, and we hope the organising priority of the field’s next phase.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fsens.2026.1951973/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Celia Toyos-Rodriguez, Institute of Carbon Science and Technology (INCAR-CSIC), Spain

Reviewed by: Keith Stine, University of Missouri–St. Louis, United States

Miaomiao Zhu, Nanjing Forestry University, China