based grafted to carbon nanotubes: A design-of-experiments optimisation of current density and stability

Enzymatic glucose electrodes based on mediated electron transfer have potential for application as semi- implantable or implantable sensors. Enzyme electrodes consisting of adsorbed osmium-based redox polymer crosslinked with a glucose oxidising enzyme are promising systems for continuous glucose moni- toring, but suffer from signal output magnitude and long-term stability issues. The inclusion of carbon nanosupports such as multiwalled carbon nanotubes (MWCNTs) into these sensors tends to increase char- acteristics such as current density and surface coverage of enzyme or mediator. However, large quantities of nanomaterials are often necessary to see signiﬁcant effects. Grafting of the enzyme to the surface of the MWCNTs improves dispersibility of the nanosupport aiding enzyme electrode fabrication, and increases enzyme activity. Here we report on a design of experiments (DoE) approach to determine the optimum amount of each component in enzyme electrodes, using glucose oxidase grafted to carbon nanotube support, to maximise current density and stability for application to continuous use glucose biosensing. Using the DoE approach while considering current density and stability responses delivers a set of component amounts where both responses are optimised. Thus far stability has not been in- vestigated as a response to be optimised using a DoE approach. The optimised enzyme electrodes show a current density of 3.18 ± 0.30 mA cm − 2 , representing a 146% increase in current density in 50 mM phosphate-buffered saline at 37 ° C containing 5 mM glucose when compared to similar systems where enzyme and nanosupport are not grafted to each other. Using the predictive DoE model, component amounts were then modiﬁed to minimise the quantity of the nanoconjugate while showing similar electrochemical behaviour and current density to the optimised system, using 93% less of the nanoconjugate. However, the operational stability under continuous use was moderate with only ≈ 50% amperometric current retained after 12 hr use. Overcoating with a Naﬁon protective layer improved stability to 72– 75% over the same period. The coupling of adsorbed ﬁlms to the electrode surface, use of additional perm-selective membranes, and/or use of pulsed potentials to implement intermittent sampling of glu- cose levels, rather than continuous amperometry, is proposed to improve operational stability.


Introduction
Over the past few decades there has been increased interest in electrochemical mediated enzymatic glucose sensors targeted towards semi-implantable and implantable glucose biosensors [14 , 19 , 23 , 31 , 49 , 52 , 56] . Of all glucose-oxidising enzymes, glucose oxidase (GOx) has been widely utilised due to its relatively low cost as well its high bioactivity and stability [60] . Elec-trochemical GOx-based glucose biosensors require either a high overpotential, for first-generation sensors, or the use of a mediator as an intermediate in the electron transfer process, for second-generation sensors, as the active centre of the enzyme is buried within the enzyme structure hindering direct electron transfer [8] . There has been substantial research into the use of redox hydrogels based on polymer-bound osmium complexes [32 , 43 , 46 , 49] as mediators, with polypyridyl complexes of osmium co-ordinatively bound to poly(N-vinylimidazole), poly(4vinylpyridine), poly(methacrylate) or similar polymers for this purpose [2 , 11 , 42] . Osmium-based complexes and polymers show advantages over iron and ruthenium based systems owing to the ability to modulate the mediator redox potential of the central metal via co-ordinating ligands, the relative stability of the resulting complexes in the reduced/oxidised states (Os II /Os III ), and because the hydrogel characteristics of enzymatic electrodes based on redox-polymer films permit rapid mass and charge transport, thus generating substantial current signals [31 , 42 , 51] .
The integration of nanomaterials as components of enzyme electrodes aids in increasing current capture as evidenced by reports on the effect upon inclusion of gold nanoparticles, platinum nanoparticles, or carbon nanoparticles into electrodes [21 , 33 , 50 , 58 , 61] . In the case of enzymatic electrodes for glucose oxidation, it has been demonstrated that systems integrating carbon micro-, meso-and nano-structured materials show improved performance over those prepared without addition of such materials [5 , 10 , 35 , 42] . Inclusion of multi-walled carbon nanotubes (MWC-NTs) in particular shows a distinct advantage due to fast electron transfer rate for redox reactions and high electrical conductivity, rendering them an attractive option. For instance, inclusion of MWCNTs has been shown to improve glucose oxidation currents for crosslinked films of glucose-oxidising enzymes and osmiumbased redox polymers on electrodes [36 , 42 , 45 , 54] , attributed to improved retention of enzyme activity and increased surface area for retention of redox species [48] .
Another consideration is the effect of the nanosupport-enzyme relationship. Immobilisation of enzyme shows advantages when compared to free enzyme, namely increased enzyme activity and improved specificity [26 , 44 , 55] . Considering carbon-based nanosupports, immobilisation of enzymes onto the surface on these materials has resulted in increased enzyme turnover and extension of the lifetime of the enzyme, allowing prolonged use [6 , 24 , 27] . On investigating the nanosupport-enzyme relationship with respect to specific enzyme activity, Campbell et al. demonstrated GOx showed higher enzymatic activity when covalently attached to MWCNTs [18] . However, when MWCNTs are used for biomedical applications, the dispersibility of the material, which impacts the stability and toxicokinetics of the CNTs, plays an important role [1 , 37 , 57] . Furthermore, dispersibility also plays a key role when considering fabrication of enzyme electrode biosensors, as aggregation of MWCNTs in some mixtures can result in uneven dispersion of material, and lack of precision in drop-coating material onto the electrode surface. Here we report on the use of a MWCNT-GOx nanoconjugate that alleviates these issues as grafting enzyme onto the surface of the MWCNTs permits use of lower nanosupport amounts and improves nanoconjugate dispersibility [1 , 37] .
Taking into account the multiple components of an enzyme electrode, it would be beneficial to determine how each component, and the interaction of the components with one another, affects parameters such as current density and the stability of these current signals on prolonged use. A comprehensive study would allow the optimisation of the components to maximise electrode performance. For this purpose, a design of experiments (DoE) approach shows distinct advantage over conventional one-factor-at-a time (OFAT) methods. Conventional OFAT methods rely on holding one variable constant, while varying the other and hence detecting one variable or factor at a time, how the response changes. Unlike OFAT, DoE explores, maps and models the behaviour of the response (or multiple responses) within a given reaction space across multiple factors simultaneously by varying all variables at once, giving a comprehensive picture of how factors and their behaviour affect the response [15] . Moreover, OFAT is unable to detect the effect of factor interactions, a shortcoming that the DoE method overcomes. The statistical optimisation process that DoE offers has gained popularity from scientists and engineers across a variety of industries for the optimisation of a range of scientific processes [13 , 15 , 20 , 39 , 40 , 59] .
Here we use a dual response Box Behnken design of experiments surface response approach to improve the current density and/or operational stability of the novel enzyme electrodes. The enzyme electrodes are fabricated with osmium-based redox polymer, CNT-GOx nanoconjugate and PEGDGE crosslinker and are operated under pseudo-physiological conditions. The predictive model developed was used to optimise and modify the component amounts to achieve a compromise of high current density output and maximum stability, at a low amount of nanoconjugate.

Grafting of glucose oxidase onto multi-walled carbon nanotubes
GOx was grafted onto the MWCNTs using literature procedures [37 , 59] . Briefly, MWCNTs were first oxidised by stirring (200 rpm) in an acid solution consisting of H 2 SO 4 (98%, 7.5 mL) and HNO 3 (70%, 2.5 mL), at room temperature overnight. Acid-treated and oxidised MWCNTs (CNT-ox) were washed with distilled water and dried at 80 °C in a vacuum oven. CNT-ox (20 mg) was suspended in distilled water (10 mL) and then added to a mixture of MES buffer (4 mL, 500 mM, pH 6.5), sulfo-NHS aqueous solution (4 mL, 434 mM), and EDC aqueous solution (2 mL, 53 mM). After rigorous stirring at room temperature for 1 h the suspension was filtered and washed with 100 mM MES buffer (pH 6.5). Covalent attachment of the GOx onto the MWCNTs (CNT-GOx) occurred on addition of the EDC -NHS activated CNTs (2 mg) to 1 mL of GOx solution (10 mg/mL in 100 mM PBS, pH 7.0). The mixture was stirred (200 rpm) at room temperature for 1 h and then placed in the fridge at 4 °C overnight. Aliquots of this mixture were taken and stored at −20 °C until used.

Enzyme electrode preparation
Graphite rods (Graphite store, USA, 3.0 mm diameter, NC001295) were cut, insulated with heat shrinking tubing and polished at one end using fine grit paper to give graphite working electrodes with a geometric surface area of 0.0707 cm 2 . The enzyme electrodes were assembled by drop-coating appropriate volumes of each of the components: Os(bpy)PVI redox polymer aqueous solution (5 mg mL −1 ); CNT-GOx solution (2 mg CNT in 1 mL of 10 mg mL −1 GOx); and PEGDGE crosslinker solution (15 mg mL −1 ). The deposition was allowed to dry for 24 h before the electrodes are used.

Design of experiments
The volume of each of the components deposited on the electrodes is determined by the Design Expert Software (Version 9,

Electrochemical measurements
Electrochemical tests were conducted using a CH Instrument 1030a multichannel potentiostat (IJ Cambria) coupled with a thermostated electrochemical cell containing PBS at 37 °C in the presence of ambient oxygen. The prepared enzyme electrodes were used as the working electrodes and paired with a custom-built Ag/AgCl (3 M KCl) reference electrode and a platinum mesh (Goodfellow) as a counter electrode. Currents were normalised to the geometric surface area of the graphite disk electrodes to generate current density data. The stability values presented in this paper represent, unless otherwise indicated, the percentage of amperometric current density remaining at the end of a 3-hour operational period compared to that obtained 10 min after initial polarisation at 0.35 V.

Enzymatic assay of glucose oxidase
The GOx activity was determined using an o-dianisidine, horseradish peroxidase coupled, spectrophotometric assay by monitoring absorbance change (Agilent 8453 UV-visible spectrophotometer) at 460 nm [22] .

Results and discussion
Addition of multiwalled carbon nanotubes (MWCNTs) as a nanosupport has been shown to increase the current density and redox mediator surface coverage of enzyme electrodes by providing an increased surface area for immobilisation of components [40 , 47] . However, in these cases, large quantities of MWCNTs have been used which can cause issues-mainly in stability and difficulties of co-immobilisation. Use of large quantities of components increases the difficulty in controlling the drop-coat on the electrode surface, affecting the precision and reproducibility. Keeping in mind future in vivo application, these electrodes would also require rigorous testing to ensure that the materials remain immobilised in the matrix. Grafting of enzyme onto MWCNTs has a significant effect, im proving the stability of the enzyme and of the measured current density for enzyme electrodes [28 , 59] . Using this technique, it is possible to obtain a nanocomposite of enzyme and nanosupport (CNT-GOx) which shows current densities comparable to previous systems while using significantly smaller amounts. Here we undertake a design of experiments approach to determine the optimum amount of each component to be used to deliver either the highest current density or the highest operational stability.  [34 , 46] . The half-wave potential recorded is negatively shifted slightly when in the presence of a glucose compared to the redox potential in the absence of glucose. The basis for this shift is unclear at present, but may be indicative of substrate transport that occurs for a mixed case between substrate and kinetic-limited conditions [9 , 41] . At relatively slow scan rates ( < 20 mV s −1 ), peak currents vary linearly with scan rate as expected for a surface-confined redox response [7] . Peak currents vary linearly with the square root of scan rate at higher scan rates ( > 20 mV s −1 ) when semi-infinite diffusion pertains for these multilayer films [7] . The osmium surface coverage ( os ) for the redox polymer, estimated by integrating the area under the peak for CVs recorded at slow scan rates in the absence of substrate, was found to be 182 ± 12 nmol cm −2, indicative of a multi-layer formation and similar to results obtained by others for the co-immobilisation of GOx with osmium-based redox polymers [12 , 13 , 47 , 48] . The addition of glucose to the electrochemical cell resulted in sigmoidal shaped responses characteristic of an electrocatalytic (EC') process.
Amperometric measurements were carried out at 0.35 V vs. Ag/AgCl (3 M KCl), selected as a potential 150 mV more positive than that at which a steady state current is achieved using hydrodynamic amperometry in PBS solutions containing 5 mM glucose ( Figure S2). Amperometric glucose oxidation current density as a function of glucose concentration ( Fig. 2 with raw amperometry trace in Figure S3) fitted to a Michaelis-Menten model allowed the estimation of K m app values and maximum saturation current densities (j max ) of 3.72 ± 0.20 mM and 3.53 ± 0.10 mA cm −2 respectively. The obtained K m app is similar to values obtained from previous reports for GOx immobilised with Os based polymer on the electrode surface without CNTs [48] indicating good affinity towards the substrate while j max correlates well with previous reports [40 , 48] , demonstrating that the CNT-GOx based system has comparable current density to systems with higher CNT loads. It should be noted that such a low K m app can result in a biosensor with a linear range for detection of glucose levels in saliva or sweat [16] but that the linear range will need to be expanded, for example through use of additional polymer coatings [62] to allow the  detection of glucose in physiological fluids such as interstitial fluid or blood.

Design of experiments
To conduct a systematic optimisation of relative amounts of each component to be used to prepare an enzyme electrode with high response signal and maximum operational stability, we used a response surface factorial Box-Behnken Design (BBD) with the three design variables set at the levels summarised in Table 1 . The design variables of the CNT-GOx, osmium redox polymer and PEGDGE component amounts form the inputs to the model while the current density in pseudo-physiological glucose and operational stability over 3 h are the model outputs. The number of experimental runs in a Box-Behnken design is N = 2k (k-1) + C 0 [53] , where k is the number of factors, and C 0 is the number of central points, i.e., the runs where each component is at the central values shown in Table 1 . The 17 run experimental design ( Table 2 ), representing a systematic sampling of combinations possible within the design space, was used to gauge the relative importance of the enzyme electrode components and their interactions. Each of the 17 experimental runs was performed on three electrodes and the average response was used in the design. The order of experimental runs was randomised to ensure independence of the data points of components and to separate the repeated central runs so as to account for human error. By analysing results from this small sample space, predictions can be made about any point in the entire experimental space, including those points about which we have no prior knowledge, if the model is found to be significant and valid.
In terms of selecting the inputs, the low levels of CNT-GOx and Os(bpy)PVI component amounts in this design were selected to be 10 μg each as this is the minimum level requirement for the production of glucose oxidation current density based on previous reports [17 , 29 , 30 , 40 , 43] . The CNT-GOx levels were selected keeping in mind that some amount of enzyme is required in the system to achieve glucose oxidation, but large amounts will cause difficulty in dispersing the CNT-GOx in solution to achieve reproducible co-immobilisation by drop-coating. The five runs for electrodes prepared using the central (0) component level, runs 1, 4, 6, 7 and 9 in Table 2 , attained current density and stability responses of 2.93 ± 0.67, 2.91 ± 0.67, 2.58 ± 0.22, 2.62 ± 0.11 and 2.54 ± 0.57 mA cm −2 and 4 9, 55, 4 9, 59 and 50%, respectively. When all 15 electrode responses for electrodes prepared using the central (0) component level are considered together an average current density and stability response of 2.72 ± 0.44 mA cm −2 and 53 ± 4% is obtained, respectively. Replication of the central levels for the model strengthens the model and minimises the error in predictions.
After completing the experimental runs ( Table 1 ), a statistical analysis of the variance (ANOVA) was undertaken on the results in order to identify the most significant sources of variation and thereby understand the roles of the three experimental variables on each response [20] . ANOVA yields F-ratios, which forms the basis for rank-ordering main effects and understanding their relative importance. Apart from quantifying the impact of the three main effects on each response, ANOVA is also able to identify statistically significant two-and three-factor interactions. The approach in model analysis is to check if the F-values are significant, the adjusted and predicted R 2 values are within 0.2 and the adequate precision is over 4. If these criteria are met, the model is valid and makes good predictions for average responses [3 , 4] .
In the case of both current density and stability responses, the models were found to be statistically valid with significant correlation between observed and predicted responses (R 2 of 0.94 and 0.92 for current density and stability, respectively). For current density, the F-value (13.07) and p-value (0.0013) evaluated suggests that the model is statistically significant. There is therefore only a 0.13% chance that an F-value this large could occur due to noise. Furthermore, adjusted R 2 (adj-R 2 , 0.87) and predicted R 2 (Q 2 , 0.74) values are within 0.2, and the adequate precision (11.42) is higher than 4, thereby suggesting that the model chosen predicts well in the chosen space and will give good predictions for average responses.
Similarly, for stability, the F-value (9.00) and p-value (0.0042) indicate the significance of the model. There is only a 0.42% chance that an F-value this large could occur due to noise. Furthermore, adjusted R 2 (adj-R 2 , 0.82) and predicted R 2 (Q 2 , 0.65) values are within 0.2, and the adequate precision (9) is higher than 4, indicating a valid and predictive model. The larger difference in correlation coefficients here could be due to some variability in crosslinker solution preparation. The PEGDGE is easily hydrolysed by water and thus cannot be made into one stock solution to be used for all electrodes prepared and tested. In order to use it as a crosslinker, fresh solutions must be made when preparing the electrodes, and this can cause a variability that cannot be avoided, affecting R 2 values.
Each response can be presented by a quadratic equation, where y is the predicted response value (current density in mA cm −2 or percentage stability, respectively), x 1 , x 2 and x 3 are the CNT-GOx, redox polymer and enzyme amounts in μg used in the enzyme electrode preparation, b 0 is the constant coefficient (intercept), b 1 , b 2 , b 3 and b 12 , b 13 , b 23 are linear and cross product coefficients, respectively, and the quadratic coefficients are b 11 , b 22 and b 33 . The resulting response models from the 17 runs for current density (2) and stability (3) are: The signs of the coefficients of the factors in the model equations indicate their relative effects, in which a positive sign indicates that a higher response can be obtained if the values of these factors are greater than those of the centre point [3 , 4 , 15 , 39] .
Considering the model of current density ( Eq. (2) ), all three main factors were found to be significant. Additionally, the amount of Os(bpy)PVI has a synergistic factor interaction with both CNT-GOx and PEGDGE amounts. In the case of stability ( Eq. (3) ), the Os(bpy)PVI and CNT-GOx has significant impact on the response when taken in amounts above the central levels, and their interaction has a beneficial effect on stability in the range tested in this DoE. The

Model validation
Before a system is optimised based on a design of experiments, it is worthwhile to demonstrate that the model is a reasonable representation of the actual system and is reproducible with enough accuracy to satisfy analysis objectives. The statistical analysis gives a mathematical indication of model validity but a secondary physical validation can be done by performing more experimental runs on the system and comparing the predicted values with the observed values.  Table 3 . Experimental values plotted against predicted values for each set of parameters ( Fig. 5 ), results in correlation coefficients (R 2 ) of 0.912 and 0.914 for current density and stability models, respectively, indicating that the models are valid as the experimental results correlate well with statistical model predictions in all five runs. In the case of the stability validation tests, the actual stability points are systematically lower than the predicted. However, the error bars (representing the standard deviation of the experimental data points), fall on the line representing the ideal agreement ( y = x ). As stated previously, some variability is expected in stability measurements as fresh crosslinker solution must be prepared before each deposition.  Concerning the physical validity of the models, in all five validation tests the experimental results for current density and stability are within the predicted range. Therefore, these statistical models can be considered valid mathematical representations and employed as predictive tools to find optimised component amounts.

Optimisation of enzyme electrode
The software allowed for the optimisation of the system, keeping in mind the goal to achieve high current density and maximum stability of response. Multi-response design optimisation is resolved by the Design Expert software using a desirability function, a weighted geometric mean that combines each individual response optimisation function into a single objective function (Supplementary equation S1) [4] . In this case, both responses were set by the user to be maximised, i.e., to have high desirability (Fig.  S8). The DoE optimum component amounts using the model equations are 150 μg CNT-GOx, 95 μg Os(bpy)PVI redox polymer and 34.2 μg PEGDGE. The biosensor formulated with these component values is predicted to deliver a current density of 3.18 ± 0.30 mA cm −2 and a stability of 54 ± 4% in PBS containing 5 mM glucose. An actual measured current density of 3.10 ± 0.19 mA cm −2 and a stability of 51 ± 4% ( n = 3) is obtained for the enzyme electrodes prepared using the DoE determined optimum component amounts. The current density response at physiological level from this optimised system shows a 146% or 2.4 fold increase over that of a similar system prepared using acid treated CNTs as a nanosupport with no covalent linkage between enzyme and nanosupport [12] (1.3 mA cm −2 ). The CV response for the optimised system is shown in Fig. 6 and the amperometric current density versus glucose concentration response is plotted in Fig. 7 (raw amperometry trace in supplementary Figure S3), with maximum current density j max estimated as 3.98 mA cm −2 and a K m app of 5.0 ± 0.10 mM. The K m app shows good agreement with literature on previous studies based on the use of redox polymer mediated glucose oxidation by GOx enzyme electrodes [48] . As stated previously, such a low K m app will give a linear range for detection of glucose levels in saliva or sweat [16] but the linear range will need to be expanded to allow the detection of glucose in physiological fluids such as interstitial fluid or blood. The significant increase in current density over that for electrodes prepared using acid treated CNTs as a nanosupport [12] may be attributed to the effect of the covalent attachment of enzyme onto the nanosupport. Covalent attachment of the GOx onto the wall of the MWCNTs should increase the activity of the enzyme [18] . In order to verify this, enzymatic activity was measured for enzyme in solution and compared to that for systems with GOx entrapped in redox hydrogel without CNTs, with CNTs (following the system investigated by us previously [12] ) and with the system optimised by DoE. The average activity of 1443 U mg −1 was obtained for the system optimised using the DoE. The activity obtained for the free enzyme solution, the system with entrapped GOx in redox hydrogel without CNTs and the system with CNTs was 276 U mg −1 , 235 U mg −1 and 227 U mg −1 respectively which correlates well with the reported activity from Sigma-Aldrich (100-250 U mg −1 ). This confirms that there is an increase in the enzymatic activity on grafting of the GOx onto CNTs. This enhanced activity is responsible for the increased current density while using significantly smaller amounts of enzyme and nanosupport.
A minimisation constraint was applied to the CNT-GOx levels, as keeping the amount of this component low has two advantages: a lower amount of enzyme used makes the biosensor more cost-effective to produce and minimising amount of CNT makes electrode preparation easier. Higher electrode-to-electrode precision obtained using lower CNT-GOx levels is due to increased control over the drop-coating procedure, as there is better dispersibility of the components in the drop-coat mixture. The DoE optimum component amounts using the model equations, based on the previously explained constraints, are 10 μg CNT-GOx, 90 μg Os(bpy)PVI redox polymer and 105 μg PEGDGE. The biosensor formulated with these values is predicted to deliver a current density of 2.43 ± 0.30 mA cm −2 and a stability of 52 ± 5% in PBS containing 5 mM glucose. An actual measured current density of 2.08 ± 0.33 mA cm −2 and a stability of 60 ± 3% ( n = 3) was obtained for the enzyme electrodes prepared using the selected component amounts. In terms of current density at physiological glucose levels in the presence of oxygen, the system with a minimisation constraint represents a 7.7-fold increase on the response for enzyme electrodes optimised through a OFAT method of optimisation of response (0.27 mA cm −2 ) using the same components except where the redox polymer Os(dmobpy)PVI was used [47] and a 1.7 fold increase over a system prepared using identical components, but where enzyme was not covalently linked to nanosupport [12] (1.3 mA cm −2 ). When comparing to results observed previously [40] for enzyme electrodes prepared by co-immobilisation of MWCNTs, GOx, osmium redox complex and carboxymethylated dextran (1.2 mA cm −2 ) in the absence of oxygen, a 1.7 fold increase in current density response is observed. Kumar and Leech [40] measured current density response in the absence of oxygen, a substance known to compete with the mediator and thus decrease current density. Our tests are performed in the presence of oxygen to more accurately mimic current density response under physiological conditions.
The stability of glucose-oxidising enzyme electrodes has not been investigated as a response in a DoE prior to this, to our knowledge. Our results show that the average stability obtained for both systems optimised by a design of experiments approach is around 50% after 3 hr continuous amperometry under physiological conditions in PBS. Thus, while the use of a DoE approach showed favourable results in optimising current density using the innovative approach of grafting GOx to CNTs, the same cannot be said when considering the stability. This rapid decay could be attributed to either enzyme turnover or concentration depletion due to the low level (5 mM) of glucose present during long measurements in a static electrochemical cell under an applied potential sufficient to ensure continuous glucose oxidation. The use of other techniques such as coupling of the layers to the electrode or the coating with protective polymer films could be useful to enhance the operational stability. For example, use of Nafion overcoating has been proven to increase the stability of the electrode while protecting it from interferents such as uric acid and ascorbic acid [12] . Overcoating of enzyme electrodes using a 0.5 wt% Nafion solution, following the protocol previously used [12] , showed a marked improvement in the stability for both the electrodes prepared using the optimised component amounts and electrodes prepared using the amounts selected with the minimisation constraint applied ( Fig. 8 ). Another possibility is electrochemical crosslinking through co-deposition. This has been investigated by the Schuhmann group showing that in the co-deposition of poly(benzoxazine) and Os-based redox polymer, optimising the ratio of the two polymers showed an improvement in stability [11] .

Conclusions
Immobilisation of glucose oxidase and MWCNTs for preparation of glucose biosensors was replaced by use of a covalently-bound nanoconjugate of GOx and MWCNTs (CNT-GOx), permitting lower amounts of nanoconjugate to be used in the preparation of sensors compared to amounts of MWCNT used previously. A DoE optimisation approach, while considering two responses (current density and stability) allowed the discovery of a set of component amounts where both responses could be maximised. Thus far stability has not been investigated as a response to be optimised using a DoE approach. The successful use of this predictive model in the optimisation of enzyme electrode components indicates its potential for further applications in this field. Furthermore, the current densities obtained for the optimised systems, even when a minimisation constraint is applied, are significantly higher than those obtained in a system where enzyme and nanosupport are not covalently bound. This indicates that the manipulation of the enzymenanosupport relationship can enhance enzyme activity and thus, current density, allowing the use of significantly lower amounts of active components. Stability obtained for the systems was moderate at only ≈ 50% after 12 hr continuous use, but use of a Nafion protective layer improved stability to 72-75%. Approaches, not tested in this work, that could further improve stability include coupling of adsorbed films to the electrode surface through covalent bonds, use of additional perm-selective membranes, and/or use of pulsed potentials to implement intermittent sampling of glucose levels, rather that continuous amperometry.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.