Paper‐Based Wearable Patches for Real‐Time, Quantitative Lactate Monitoring

Wearable sensors are establishing themselves as options for real‐time continuous health monitoring in health care and wellness. In particular, the use of flexible interfaces that conform to the skin have attracted considerable interest for the extraction of meaningful pathophysiological information through continuous and painless sampling and analysis of biofluids. In contrast, conventional techniques for biomarkers analysis are difficult to adapt to real‐time portable monitoring due to their invasive sampling protocols, biosample preparation and reagent stabilization. Here a shelf‐stable, non‐invasive, paper‐based colorimetric wearable lactate sensor is reported. This sensor exploits the ability of silk to control the concentration, print, and functionally preserve labile transducing biomolecules in the format of a shelf‐stable digital patch for optical readout. This novel approach overcomes major challenges associated with the commercialization of colorimetric wearable sensors (e.g., enzyme thermal instability, narrow sensing range, low sensitivity, and qualitative response) by showing a combination of unprecedented stability (i.e., up to 2 years in refrigerated conditions), wide sensing range, and high sensitivity. Additionally, real‐time quantitative signal readouts are achieved using machine learning‐driven image analysis enabling physiological status evaluation with a simple smartphone camera.


Introduction
Non-invasive continuous health monitoring is crucial to obtain physiological and pathological information to assess human wellbeing. [1]In this context, wearable sensors provide great utility providing painless sampling of biofluids and real-time analysis of relevant biomarkers, with emerging applications in healthcare and sport medicine.Lactate oxidase (LOx), horseradish peroxidase (HRP) and dyes are added to a silk fibroin solution to form a chromogenic enzymatic ink, which is then drop-cast on paper.Circular lactate-sensing interfaces are laser-cut and applied on Tegaderm™ films to obtain wearable patches whose color shifts from yellow to dark red by increasing the concentration of lactate in sweat.B) Schematic representation of the LOx/HRP cascade reaction.LOx oxidizes lactate to produce pyruvate and hydrogen peroxide, which is used by HRP to oxidize the chromogenic substrates generating a visible color change.C) Photographs of the lactate-sensing patches applied on skin.The insets show the circular lactate-sensing interfaces before (top) and after (bottom) the colorimetric response.
][30][31] Sweat monitoring can reveal important information regarding patients' and athletes' physiological status. [32]The distribution of sweat glands across the skin enables its sampling over the entire body making it an excellent biofluid both for localized and distributed sensing. [32]n counterpoint to its sampling ease, sweat is a complex biofluid affected by environmental and physiological interferences that make its analysis challenging.
As the end product of glycolysis, lactate in sweat is a target of interest for non-invasive monitoring.[47] While its accumulation causes soreness that can deter from further physical activity, [9] its production is deemed essential to improve endurance. [1,48]Consequently, lactate monitoring can improve athletes' performances while also preventing injuries caused by overtraining. [49]To date, conventional lactate detection techniques rely on blood samples and are not suitable for real-time portable monitoring due to their invasive sampling protocols (e.g., venipuncture or finger prick). [50,51]he wearable sensors presented in this work are light, flexible, conformable to the skin and can be worn on different parts of the body for long periods of time providing continuous distributed sampling of sweat without causing discomfort.

Colorimetric Wearable Lactate-Sensing Patches
The silk-based chromogenic enzymatic inks were formulated by incorporating lactate oxidase (LOx) and horseradish peroxidase (HRP) with the chromogenic substrates in a regenerated silk fibroin aqueous solution that enabled the stabilization of the labile molecules.The biosensing composite ink was infiltrated in small (i.e., 3 mm diameter) filter paper discs (Figure 1a) which are then arranged in pre-specified geometries onto a Tegaderm wound dressing film.Additional non-reactive inks containing reference dyes are added to the construct to produce the final wearable patches.These are flexible, conformable and can be worn for several hours without causing discomfort.The use of a semipermeable wound dressing film allows for natural breathing of the skin [52] -being permeable to water vapor, oxygen, and carbon dioxide -while ensuring contact with the absorbing bioresponsive discs for sweat collection (Figure 1c).
Upon contact with sweat, the lactate-responsive discs change color from yellow to dark red as a function of lactate concentration, while the nonreactive reference circles provide fiducial markers to correct for lighting artifacts and boundary conditions for the machine-learning driven image processing stage.The colorimetric response of the composite inks follows the LOx/HRP cascade reaction (Figure 1b) where LOx oxidizes lactate to produce pyruvate and hydrogen peroxide, which is then used by HRP to oxidize the chromogenic substrates generating an immediately visible color change.

Colorimetric Response Evaluation
Despite the effectiveness of colorimetric sensing techniques for rapid detection of analytes, the reproducibility of these systems is compromised by the potential presence of color gradients in the sensing areas.The lack of color uniformity compromises the readout reliability especially if coupled with image recognition models to obtain quantitative results. [53]The color gradient, namely coffee-ring effect, in microfluidic paper-based sensors is attributed to the sample solution transporting the chromogenic substrates and the enzymes while diffusing from the center of the sensing areas to the edges, resulting in a heterogeneous coloration. [53,54][56] The effect of the deposition of a base layer of chitosan on the performances of the silk-based ink was investigated using Whatman grade 1 filter paper (Figure 2a).First, the sensing interfaces were calibrated using solutions that mimic sweat composition with lactate concentrations in the 0 -90 mM mм range to assess the effect of chitosan on the colorimetric response.The sensitivity and sensing range were calculated by evaluating the slope of the calibration curve (i.e., Green channel (G) as a function of the logarithm of lactate concentration (Log C )).The presence of the base layer of chitosan yielded an increased sensing range (with chitosan, 0 -90 mм; without chitosan, 0 -60 mм) without compromising the sensitivity (with chitosan, −84.11 ± 3.31 G Log C −1 ; without chitosan, −87.29 ± 3.77 G Log C −1 in the 0 -60 mм range; avg ± s.e., n = 5) and improved the reproducibility of the colorimetric response at high lactate concentration (coefficient of variation in the 50 -90 mм range: with chitosan, 4.62 ± 0.68%; without chitosan, 11.34 ± 1.97%; avg ± s.e.) proving its efficacy in improving the colorimetric response.The combination of such high sensitivity and wide linear sensing range is of high utility for the colorimetric detection of lactate, which is commonly reported to have an upper detection limit of 25 mм [9][10][11]14,15] (i.e., lower than the concentration required for the analysis of undiluted sweat). [57] Te colorimetric sensor presented in this study has a wide linear sensing range (i.e., 0 -90 mм) suitable for both training monitoring and disease diagnostic.Specifically, for sport medicine applications lactate concentrations in undiluted sweat can reach 60 mм and further increase during training [46,47,57,58] while for clinical applications, a upper detection limit of 50 mм is required.[7] The specificity of the sensor was demonstrated by evaluating the absence of colorimetric response after the exposure to interfering substances (i.e., urea, sodium, potassium, ammonium, calcium, magnesium ions)(Figure S1, Supporting Information).The sensing performances of Ahlstrom grade 55 filter paper and Whatman grade 4 filter paper were also evaluated (Figures S2 and S3, Supporting Information).Whatman grade 1 filter paper showed the best performances among the three types of paper. The ifference in colorimetric response is attributed to the pore size of the paper: Whatman grade 1 has the smallest pore size (i.e., 11 μm), compared to Ahlstrom grade 55 and Whatman grade 4 (i.e., 15 and 20-25 μm respectively), which decreases the transport of the chromogenic substrates and the enzymes to the edges of the sensing areas increasing both the color homogeneity and reproducibility.

Shelf-Life Evaluation
The commercialization of the enzymatic sensors is set back by the progressive decrease in stability of the enzymes and dried proteins during storage at room temperature.Despite representing their main drawback, this problem has not been addressed by previous studies and stability tests are often lacking.Here, the demonstrated ability of regenerated silk-fibroin to stabilize labile biomolecules [16,17,26,27,[18][19][20][21][22][23][24][25] is exploited to improve the ther-mal stability of the sensors.In fact, encapsulating enzymes in a silk fibroin matrix delays the degradation of their native structure by reducing their molecular mobility and providing protection against environmental factors such as temperature and pH changes through a buffering action. [16,27]To prove the prolonged shelf-life of the sensing interfaces both with and without the base layer of chitosan, the sensors were subjected to accelerated degradation tests by storing them at 60 °C (i.e., above the degradation temperature of the enzymes) for 8, 24, and 120 h.The ability of the inks to preserve enzymatic activity was assessed when exposed to lactate variations (i.e., 0 -90 mм).Storage for 8 h at 60 °C did not affect the sensing range of the sensing interfaces with chitosan, which showed a linear response in the 0 -90 mм range.The sensing range without chitosan was reduced to 0 -30 mм.Chitosan did not affect the sensitivity in the 0 -30 mм range (with chitosan, −67.73 ± 5.86 G Log C −1 ; without chitosan, −72.21 ± 3.75 G Log C −1 ; avg ± s.e., n = 3).When increasing the storage time to 24 and 120 h at 60 °C, the presence of chitosan caused a slower reduction of the sensing range (with chitosan: 24 h, 0 -60 mм; 120 h, 0 -30 mм; without chitosan: 24 h, 0 -10 mм, 120 h, 0 -10 mм), which further confirms the combined ability of silk and chitosan to improve the colorimetric response.The key role of silk fibroin in the stabilization of the enzymes was demonstrated by comparing the sensing performances of silk-based and water-based inks (both in the presence of a base layer of chitosan) when subjected to accelerated degradation tests (i.e., storing them at 60 °C, above the degradation temperature of the enzymes for 8, 24, and 120 h).The colorimetric response of the sensing interfaces was evaluated after 8 h of storage (Figure 2c).  2d).Additionally, the colorimetric response of the silk-based interfaces is stronger (i.e., dark red) than for the water-based ones (i.e., light pink) (Figure 2c, insets).The sensing performances of Ahlstrom grade 55 filter paper and Whatman grade 4 filter paper were also evaluated after accelerated degradation tests (Figures S2 and S3, Supporting Information).Also in this case, Whatman grade 1 filter paper showed the best performances among the three studied papers.,15] The long-term stability of the wearable sensors was evaluated after 18, 21, and 24 months of storage at 4°C (Figure 2e) and 2.5 months of storage at 60 °C (Figure 2f).The silk-based chromogenic enzymatic mixture was able to preserve 66% of its activity after 2.5 months at 60 °C, as opposed to water-based control whose activity decreased to 2%.Additionally, when stored at 4 °C, the silk-based chromogenic enzymatic mixture completely retained its activity for up to 24 months at 4 °C, while in the water-based control the activity was reduced to 17%.These results further demonstrate the ability of silk fibroin to stabilize the chromogenic enzymatic mixture allowing the fabrication of shelf-stable sensors whose colorimetric response is not undermined after years of storage.

Colorimetric Wearable pH-Sensing Patches
The versatility of this approach and the possibility to develop multi-sensing patches, was demonstrated by developing silkbased chromogenic inks incorporating pH-responsive molecules (i.e., nitrazine yellow (NY), bromocresol green (BG), and phenol red (PR)).The inks presented in this study show high sensitivity (i.e., BG, −39.8 ± 1.3; NY, −76.1 ± 1.4; PR, −40.9 ± 1.9; avg ± s.e., n = 3) and reversibility suitable for the detection of pH variations in real-time.The intensity of the colorimetric response in the RGB color space for each ink was evaluated (Figure 3).The three different inks have three complementary sensing ranges (i.e., BG, pH range 3 -7; NY, pH range 5.5 -7.5; PR, pH range 6.5 -8.5) and, by combining them on the same sensing patch, it is possible to read the pH value in the physiologically relevant sweat pH range (i.e, pH 3 -8.5 pH).][61] The application of both the lactate and pH silk-based sensing interface on the same patch yields shelf-stable multianalyte sensors for comprehensive health monitoring.

Machine Learning-Driven Readout
To allow easy quantitative readouts of the colorimetric response using a smartphone camera, 1961 images of the sensors at 6 different concentrations of lactate (i.e., 0, 1, 5, 10, 30, 50 mм) were obtained and used to develop a machine learning model.The images were acquired after drop-casting the simulated sweat solution with a known lactate concentration on the sensing interfaces of the sensors (Figure 4a).To obtain a universal model able to recognize images under non-ideal light exposure, the images were acquired under different light conditions and then randomly divided in two groups: 1373 labeled images were used for the model training (Figure 4c) and 588 were used for the evaluations of its performances (Figure 4d (top)).The images of the sensors were acquired over three different days that corresponded to three different datasets, the combination of the three was analyzed as a fourth dataset (i.e., full dataset) (Tables S1 and S2, Supporting Information).
A Support Vector Machine (SVM) model was built to read and predict the colorimetric response of the sensors as a function of lactate concentration.Plots of the RGB values extracted from each dataset show that it is possible to clearly identify six data clusters (i.e., one for each lactate concentration).The clusters have a low standard deviation (i.e., 0.92 -4.13) and their organization in a 3D space enables the identification of separatory hyperplanes suitable to build a SVM predictive model [62] (Figure 4b; Figure S4, Supporting Information).Given the considerations above, the SVM model was found to be ideal for this system based on early data exploration during which the colorimetric variation of the sensors at different lactate concentrations was analyzed for each dataset.The increase in lactate concentration produced a color change from yellow to red, which, at the image analysis stage, corresponds to a decrease of the value of the green channel in the RGB color space (Figure S5, Supporting Information).The boxplots show narrow data distribution for each lactate concentration, showing the ability of the model to take into account and correct for different light conditions which, if not accounted for, would cause low prediction accuracy for colorimetric sensors.The predictive model was built for each dataset, using the 70% of the images in each dataset for the model training.The remaining 30% of the images in the datasets was used as an input to evaluate how accurately the SVM model classifies the sensor images.To calculate the accuracy of the model, the confusion matrix -whose diagonal shows the percentage of correctly classified images -was evaluated for the full dataset (Figure 4d (bottom)) as well as for the three smaller datasets (Figure S6, Supporting Information).The model was able to reach an overall accuracy of 93%, 96%, and 92% for datasets 1, 2, and 3 respectively, which proved the excellent predictive capabilities of the model.To show the classification accuracy of the model on a big and heterogenous dataset, the model was built and trained on the full dataset (namely dataset 4) and reached an overall accuracy of 89.3%.
After obtaining high prediction accuracy using a quantitative classification of the images, an additional model was built using a qualitative classification by dividing each dataset into four categories (i.e., no lactate, low concentration, medium concentration, high concentration) (Tables S3 and S4, Supporting Information).The suitability of the SVM model for this type of classification was confirmed by the presence of four bigger data clusters (i.e., one for each lactate concentration class) in the scatterplots of the RGB values (Figure S7, Supporting Information).Similarly, the boxplots of the green channel values for each lactate concentration class showed a clearly descending trend (Figure S8, Supporting Information).The evaluation of the classification accuracy of the SVM model was performed, revealing the highest prediction accuracy (i.e., 98.6%, 100%, and 98.9% for dataset 1, 2, and 3, respectively) as shown in the confusion matrix for each dataset (Figure S9, Supporting Information).Similarly, the model built on the full dataset presented high performances with an overall accuracy of 98.8%, showing the ability to obtain accurate readouts.Finally, the feasibility of the application of machine-learning driven quantitative readouts in a real-life scenario was demonstrated by placing the wearable sensors on the skin of a volunteer during a treadmill exercise session.The image of the wearable sensors after the session was acquired and processed by the SVM model which predicted a 30 mм lactate concentration (Figure 4e).

Conclusion
This study presents silk-based colorimetric wearable sensing patches for lactate concentration and pH monitoring in sweat.These sensing patches can be comfortably worn during exercise or day-to-day activities and address the drawbacks of existing wearable sensing technologies by providing long shelf-life, wide sensing range, high reproducibility, and compactness.The use of silk-fibroin as a stabilizer of biological labile components is broadened, proving its applicability in the fabrication of shelfstable sensing patches which can be used after years of storage.These flexible silk-based sensors open the way toward real-time detection of multiple analytes for continuous health and performance monitoring.Finally, pairing these wearable sensors with machine learning models for image recognition provides rapid and quantitative readouts characterized by high accuracy, high sensitivity, and sensing range suitable both for applications in sport medicine and clinical settings.

Experimental Section
Materials: Sodium carbonate, lithium bromide, Peroxidase from horseradish Type I (HRP) (P8125), sodium 3,5-dichloro-2hydroxybenzenesulfonate (B), 4-aminoantipyrine (A), sodium lactate, chitosan, acetic acid, phenol red sodium salt, nitrazine yellow, bromocresol green sodium salt, Whatman Grade 1 filter paper, and Whatman Grade 4 filter paper were purchased from Sigma-Aldrich (USA).Acid Yellow 34 (Y) was purchased from Chem Cruz (USA).Lactate oxidase Grade III (LOx) was purchased from Toyobo (USA).Tegaderm Films (size: 4.4×4.4cm) were purchased from 3 M (USA).All chemicals were used as received and they followed trace metal standard, when possible.The chromogenic substrates (i.e., A, B, and Y) were chosen considering their toxicity levels, avoiding the use of harmful or cancerogenic compounds.Silk cocoons of Bombyx mori silkworm were purchased from Tajima Shoji (Japan).Deionized (DI) water with resistivity of 18.2 MΩ cm was obtained with a Milli-Q reagent-grade water system and used to prepare aqueous solutions.
Silk Fibroin Solution Preparation: Silk fibroin was extracted following a previously reported protocol.Briefly, finely chopped Bombyx mori silk cocoons were boiled in a solution of 0.02 M sodium carbonate to remove the sericin layer for 120 min.The fibers were washed three times for 20 min in DI water, dried overnight, and dissolved in a solution of lithium bromide 9.3 M at 60 °C for 4 h.The obtained solution was dialyzed against deionized water for 2 days, changing the deionized water 6 times at regular intervals.The final solution was centrifuged twice at a speed of 9000 rpm, at 4 °C, for 20 min and then filtered yielding 7-8 wt% silk fibroin solution.
Chromogenic Enzymatic Inks Preparation: Silk-based chromogenic enzymatic inks were made of 4 wt% silk fibroin solutions containing 0.1 M PBS as ionic background which keeps the pH constant when reacting with strongly acidic or basic sweat.The final concentration of the enzymatic reagents was 339 and 150 U mL −1 for HRP and LOx, respectively.Chromogenic substrates were then dissolved in the silk-based enzymatic mixture to yield final concentrations of: A, 0.86; B, 1.82; Y, 0.3 mg mL −1 .
Water-based chromogenic enzymatic inks were made of DI water containing 0.1 M PBS as ionic background.The final concentration of the enzymatic reagents was 339 and 150 U mL −1 for HRP and LOx, respectively.Chromogenic substrates were then dissolved in the water-based enzymatic mixture to yield final concentrations of: A, 0.86; B, 1.82; Y, 0.3 mg mL −1 .
Chromogenic pH-Sensitive Inks Preparation: Silk-based chromogenic pH-sensitive inks were made of 4 wt% silk fibroin solutions containing a pH indicator (i.e., phenol red sodium salt, nitrazine yellow, bromocresol green sodium salt) with a final concentration of 2.5 mg mL −1 .
Wearable Sensing Patches Fabrication: The three different paper substrates were cut to obtain squares (side length: 3 cm).Chromogenic enzymatic ink of 150 μL were drop-cast in the center of each square and left to dry for 1 h at room temperature.This step was repeated three times to obtain three layers of ink.The functionalized paper was laser-cut to obtain circles (diameter: 3 mm) using a Trotec Speedy 300 Laser Cutter, with 75 W CO 2 laser.The functionalized paper circles were then placed on Tegaderm Films to obtain wearable patches.
Functionalization With Chitosan: Chitosan of 0.5 w/v% was dissolved in 2 v/v% acetic acid.To obtain a base layer of chitosan, 150 μL of the solution were drop-cast in the center of each square and left to dry for 1.5 h before the deposition of the three layers of chromogenic enzymatic ink.
Simulated Sweat Solution Preparation: NaCl, KCl, Urea, NH 4 Cl, CaCl Accelerated Degradation Tests: Accelerated degradation tests were performed on lactate sensing patches fabricated with silk-based or waterbased chromogenic enzymatic ink, both with or without a base layer of chitosan.The patches were stored at 60 °C, above the enzyme degradation temperature, for 8, 24120 h and 2.5 months.After storage, the colorimetric response was analyzed to assess the ability of silk-based inks to maintain enzymatic activities when exposed to lactate variations in the 0-90 mм range.To evaluate the long-term stability, the patches were stored at 4 °C for 18, 21, and 24 months.After storage, the colorimetric response was analyzed to assess the ability of silk-based inks to maintain enzymatic activities when exposed to lactate variations in the 0-90 mм range.The retained activity was evaluated by measuring the variation in the sensors' colorimetric response at 90 mм (in the Green Channel) right after fabrication and after storage: Neural Network Training: Wearable sensing patches for neural network training were made as follows.Whatman Grade 1 filter paper was functionalized with a base layer of chitosan solution and three layers of silk-based chromogenic enzymatic ink.The paper was laser-cut to obtain circles (diameter: 3 mm) using a Trotec Speedy 300 Laser Cutter, with 75 W CO 2 laser.Four sensing paper circles were applied on Tegaderm Films.To allow easier machine learning-driven lactate concentration prediction, four reference colors (RGB values: red (255, 0, 0), green (0, 255, 0), blue (0, 0, 255), and light yellow (253, 252, 188)) were included in the form of nonsensing colored paper circles.The reference colors were printed on Whatman Grade 1 filter paper using a Laser Jet Pro MFP M127fn printer from HP (USA), 24-bit color depth and resolution of 600 dpi.The non-sensing colored paper circles (diameter: 3 mm) were laser-cut using a Trotec Speedy 300 Laser Cutter, with 75 W CO 2 laser and applied on Tegaderm Films where the sensing circles were previously applied.
Image acquisition: Simulated sweat solution of 1 μL containing a precise amount of lactate at specific concentration of lactate (i.e., 1, 5, 10, 30, 50 mм) was drop-cast on each sensing circle.After the colorimetric response, images of the sensors were acquired in different light conditions using a smartphone camera (Apple, iPhone SE 2020).The images were used to train the neural network and to evaluate its performances.The trained neural network was able to predict the lactate concentration when an image of the sensor was given as input.
Image Pre-processing: Computer Vision libraries and Machine Learning algorithms were used to allow quantitative readouts.The lightness and warmth level of the acquired images were standardized using reference images (i.e., images acquired in controlled light conditions to avoid shadows and differences in warmth level that could affect the analysis of the colorimetric response of the sensing circles).The brightness of the acquired images was adjusted to match the reference images in the HSV (Hue, Saturation, and Value of brightness) color space to obtain uniform light conditions in the datasets.Afterwards, using the CIELAB (International Commission on Illumination L*a*b*, L*-Lightness, a*-redness, b*yellowness) color space, the average b* values of the background color of the reference image were extracted and set as standard values.The b* values of all the images were shifted to match these standard values, while the L* values were set to 130 to avoid that lightness differences could affect the colorimetric analysis.Finally, a color-based image filter was used to select only pixels corresponding to the sensing circles of the patches.
The OpenCV and Pillow libraries were used to convert images through different color spaces and to create masks for color extraction. [63,64]achine Learning: The data was analyzed to determine the best model category for the colorimetric analysis of the lactate-sensing patches.After early data exploration, different Support Vector Machine (SVM) models were trained for the multi-class classification.The model building was developed using RStudio and the package "e1071" for SVM models. [65]he SVM model was trained with three different datasets (DS1, DS2, DS3) containing pictures of the lactate-sensing patches applied on the skin of three different people (DS1, n = 955; DS2, n = 594; DS3, n = 665 images) (Tables S1 and S2, Supporting Information).The datasets were studied either analyzing 6 predicted lactate concentrations (i.e., 0, 1, 5, 10, 30, 50 mм), or analyzing 4 predicted classes of lactate concentration (i.e., No lactate = 0 mм, Low concentration = 1 mм, Medium concentration 5-10 mм, and High concentration 30-50 mм), (Tables S3 and S4 S5 to Table S12 (Supporting Information).

Figure 1 .
Figure 1.A) Fabrication schematic of the paper-based lactate-sensing patches.Lactate oxidase (LOx), horseradish peroxidase (HRP) and dyes are added to a silk fibroin solution to form a chromogenic enzymatic ink, which is then drop-cast on paper.Circular lactate-sensing interfaces are laser-cut and applied on Tegaderm™ films to obtain wearable patches whose color shifts from yellow to dark red by increasing the concentration of lactate in sweat.B) Schematic representation of the LOx/HRP cascade reaction.LOx oxidizes lactate to produce pyruvate and hydrogen peroxide, which is used by HRP to oxidize the chromogenic substrates generating a visible color change.C) Photographs of the lactate-sensing patches applied on skin.The insets show the circular lactate-sensing interfaces before (top) and after (bottom) the colorimetric response.

Figure 2 .
Figure 2. A) Calibration curve for the silk-based chromogenic enzymatic ink with and without the deposition of a base layer of chitosan on Whatman grade 1 filter paper.Sensing range with chitosan: 0-90 mм; sensing range without chitosan: 0-50 mм.Colored circles show the colorimetric response recorded at different lactate concentrations.B) Sensing range and sensitivity values after 8, 24, and 120 h of storage at 60 °C for the silk-based chromogenic enzymatic ink with and without the deposition of a base layer of chitosan on Whatman grade 1 filter paper.Inset shows the colorimetric response of the silk-based chromogenic enzymatic ink with a base layer of chitosan on Whatman grade 1 filter paper after 120 h of storage at 60 °C.C) Calibration curve for the silk-based and water-based chromogenic enzymatic inks with a base layer of chitosan on Whatman grade 1 filter paper.Sensing range of silk-based ink: 0-90 mм; sensing range of water-based ink: 0-10 mм.Colored circles show the colorimetric response recorded at different lactate concentrations.D) Sensing range and sensitivity values after 8, 24, and 120 h of storage at 60 °C for the silk-based and water-based chromogenic enzymatic inks a base layer of chitosan on Whatman grade 1 filter paper.Inset shows the colorimetric response of the water-based chromogenic enzymatic ink with a base layer of chitosan on Whatman grade 1 filter paper after 120 h of storage at 60 °C.E) Retained activity after 18, 21, and 24 months of storage at 4 °C for the silk-based and water-based chromogenic enzymatic inks with a base layer of chitosan on Whatman grade 1 filter paper.Insets show the colorimetric response of the silk-based and water-based chromogenic enzymatic ink 24 months of storage at 4 °C F) Retained activity after 8, 24, 5 days and 2.5 months of storage at 60 °C for the silk-based and water-based chromogenic enzymatic inks with a base layer of chitosan on Whatman grade 1 filter paper.Insets show the colorimetric response of the silk-based and water-based chromogenic enzymatic ink 2.5 months of storage at 60 °C.

Figure 4 .
Figure 4. A) Schematic of the SVM model training and lactate prediction process.Sweat with a known lactate concentration is drop-cast on the sensor.After the response, images of the sensor are acquired in different light conditions.The images are used to train the SVM model and to evaluate its performances.The trained SVM model is able to predict the lactate concentration when an image of the sensor is given as input.B) (Top) Scatterplot of the full dataset in the three-dimensional RGB color space for the different lactate concentrations (0-50 mм).(Bottom) Scatterplot of the full dataset in the 2D Blue Channel versus Red Channel plane for the different lactate concentrations (0-50 mм).C) The SVM model was trained using 1316 images of the colorimetric response of the wearable sensor upon exposition to known lactate concentrations in the 0-50 mм range.D) (Top) The performances of the SVM model were evaluated using 564 images of the colorimetric response of the wearable sensor upon exposition to known lactate concentrations in the 0-50 mм range.(Bottom) Confusion matrix for the 564 evaluation images showing that the SVM model accurately classified the images into the 6 lactate concentration categories with an overall accuracy of 89.3%.E) Images of the colorimetric response of the wearable sensors before (left), during (center) and after (right) a treadmill exercise session.The predicted lactate concentration at the end of the session was 30 mм.

2 ,
MgCl 2 , were dissolved in DI water to yield final concentrations of 40, 3, 22, 3, 0.4 mM, and 50 μM respectively.Analysis of Colorimetric Response: To induce a colorimetric response, 1 μL of simulated sweat solution at specific concentration of lactate (i.e., 1, 5, 10, 30, 50, 60, 90 mм) was drop-cast on each sensing circle.The fluid spread on the surface of the sensing circle and mixed with the reagents to produce a colorimetric response.The colorimetric response was analyzed collecting images using a Laser Jet Pro MFP M127fn scanner from HP (USA), 24-bit color depth and resolution of 600 dpi.ImageJ was used to quantify the response as variations in the Red, Green or Blue channel intensities.The specificity was evaluated by measuring the variation in the sensors' colorimetric response in the Green Channel before and after the exposure to interfering substances (NaCl, KCl, Urea, NH 4 Cl, CaCl 2 , MgCl 2 , with concentrations of 40, 3, 22, 3, 0.4 mM, and 50 μм respectively): Green Variation (%) = G before − G after exposureG before(1) , Supporting Information).Each dataset was divided into six categories corresponding to the different lactate concentrations, leading to three different datasets.To have a larger and more heterogeneous dataset, Dataset 1, Dataset 2, and Dataset 3 were merged to obtain the full dataset (n = 1880, Dataset 4).The values of sensitivity (True Positive Rate), specificity (True Negative Rate), Positive Predicted Value (PPV) and Negative Predicted Value (NPV) were reported in Table