Deep-Learning Assisted Cactus-inspired Osmosis-enrichment for Biosafety-isolative Wearable Sweat Metabolism Assessment

Deep-Learning Assisted Cactus-inspired Osmosis-enrichment for Biosafety-isolative Wearable Sweat Metabolism Assessment | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 26 June 2025 V1 Latest version Share on Deep-Learning Assisted Cactus-inspired Osmosis-enrichment for Biosafety-isolative Wearable Sweat Metabolism Assessment Authors : Yuwen Yan , Miaorong Lin , Yueli Luo , Jihan Qu , Yonghuan Chen , Zhihao Zhang , Congxin Xia , Jianxin Meng , and Fengyu Li 0000-0003-2481-6111 [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.175092536.60750459/v1 Published Biosensors Version of record Peer review timeline 240 views 145 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Wearable sweat collection sensors have attracted wide attention due to their advantages on portability and continuous monitoring. However, the skin direct-contacting of sensor patches usually triggers the skin allergy or adverse reactions by the dyes, biomarkers or sensing chemical substances. Here, we proposed a biomimetic Janus membrane, inspired by the structure of cactus, that could unidirectionally transport and converge sweat to the assigned detection point. Via unidirectional transferring from hydrophobic layer of Janus membrane to the hydrophilic layer, the sweat droplets were enriched to the assigned detection point of the conical hydrophilic pattern by the Laplace Pressure. The bionic osmosis-enrichment sensing patch effectively inhibits direct-contact of indicators to skin, eliminating potential epidermal contamination. This achieved the effect of in-situ perspiration collection under the premise of biosafety isolation. We apply Deep-Learning (DL)-assisted fluorescence sensor to efficiently and accurately detect the concentration of biomarkers in sweat. The convolutional neural network (CNN) model could easily and accurately classify and quantitatively analyze the concentrations of amino acids, Ca 2+ and Cl - , with 100% classification accuracy. The method shown good reliability in collecting and analyzing sweat, and provided a simple index for clinical health monitoring, disease intervention prevention and clinical diagnosis. Cite this paper: Chin. J. Chem. 2021 , 39 , XXX—XXX. DOI: 10.1002/cjoc.202100XXX Deep-Learning Assisted Cactus-inspired Osmosis-enrichment for Biosafety-isolative Wearable Sweat Metabolism Assessment Yuwen Yan, 1 Miaorong Lin, 1 Yueli Luo, 3 Jihan Qu, 1 Yonghuan Chen, 1 Zhihao Zhang, 2 Congxin Xia, 3 Jianxin Meng* ,1 , and Fengyu Li* ,1,2 1. College of Chemistry and Materials Science, Guangdong Provincial Key Laboratory of Speed Capability Research, Su Bingtian Center for Speed Research and Training, Jinan University, Guangzhou 510632, China 2. College of Chemistry, Zhengzhou University, Zhengzhou 450001, China 3. School of Physics, Henan Normal University, Xinxiang 453007, China Email: [email protected] & [email protected] Cactus-inspired | Osmosis-enrichment | Deep learning | Biosafety isolation | Wearable sweat detection Wearable sweat collection sensors have attracted wide attention due to their advantages on portability and continuous monitoring. However, the skin direct-contacting of sensor patches usually triggers the skin allergy or adverse reactions by the dyes, biomarkers or sensing chemical substances. Here, we proposed a biomimetic Janus membrane, inspired by the structure of cactus, that could unidirectionally transport and converge sweat to the assigned detection point. Via unidirectional transferring from hydrophobic layer of Janus membrane to the hydrophilic layer, the sweat droplets were enriched to the assigned detection point of the conical hydrophilic pattern by the Laplace Pressure. The bionic osmosis-enrichment sensing patch effectively inhibits direct-contact of indicators to skin, eliminating potential epidermal contamination. This achieved the effect of in-situ perspiration collection under the premise of biosafety isolation. We apply Deep-Learning (DL)-assisted fluorescence sensor to efficiently and accurately detect the concentration of biomarkers in sweat. The convolutional neural network (CNN) model could easily and accurately classify and quantitatively analyze the concentrations of amino acids, Ca 2+ and Cl - , with 100% classification accuracy. The method shown good reliability in collecting and analyzing sweat, and provided a simple index for clinical health monitoring, disease intervention prevention and clinical diagnosis. Background and Originality Content Personal health monitoring plays more and more crucial role in daily healthcare and chronic disease management. [1] Various methods have been explored to acquire vital signs from the human body. [2-4] As a representative biological fluid, sweat contains an abundance of biomarkers. Sweat analysis provides a simple, convenient, and non-invasive approach to obtain the chemical constituents and biomarker information for daily health monitoring. [5] The researchers developed methods for collecting and analyzing topical sweat in vitro, including sweat absorbing patches, textiles, and microfluidic channels. [6-8] However, the skin direct-contacting of sensor patches usually triggers the skin allergy or adverse reactions by the dyes, biomarkers or sensing chemical substances. [9] Sample collection and biosafety isolation are the critical issue for wearable sweat detections. Therefore, it calls an efficient approach to rapidly collect sweat on skin with the noncontacted wearable or adherable sensing devices or patches. Janus membranes possess asymmetrical morphological structures or chemical compositions on either side, making them widely applicable in fields such as membrane distillation, oil-water separation, and seawater desalination. [10-12] The asymmetric wettability can realize the unidirectional transport of droplets from the hydrophobic layer to the hydrophilic layer. It can effectively isolate the contact between the skin and chemical indicators. In addition, although the sweat glands are distributed in all parts of the body, the sweat secretion rates are low on the arms or legs. [13] In certain scenarios, the volume of sweat may not meet monitoring requirements. The spines of the cactus have a conical structure, and the droplets move spontaneously towards the bottom under the action of Laplace pressure. [14] Inspired by the special structure of cactus, the hydrophilic layer with conical structure can be prepared on the hydrophobic substrate to obtain patterned Janus membrane. It can solve the problem of limited volume of sweat and directionality during sweat collection. Common methods for sweat analysis include colorimetry, [15,16] fluorescence spectroscopy, [18,19] electrochemistry, [20,21] and surface-enhanced Raman spectroscopy (SERS). [22-24] However, the accuracy of these strategies depends heavily on the performance of the instrument and the technical proficiency of the operator. Inherent disadvantages, such as expensive equipment, time consuming, complex sample pretreatment processes, and the need for trained technicians, limit their widespread use. Visual fluorescence sensors have attracted much attention in sweat sensing because of their excellent selectivity and sensitivity. However, data processing and analysis of fluorescent images inevitably require increased processing time and professional personnel. In addition, image datasets with nonlinear, high and complex data are difficult to analyze using traditional statistical methods. With the rapid development of artificial intelligence and the increasing demand for data analysis in the field of chemical research, more and more deep learning technologies have been applied to the field of chemistry, including drug design, [25,26] materials science, [27] analytical chemistry [28,29] and so on. Convolutional neural network (CNN) is a branch of deep learning algorithm, which has significant advantages in processing high-dimensional data and solving nonlinear problems. It is widely used in image processing and recognition tasks. [30-32] In fluorescence image processing, the CNN model can quickly analyze the whole detection environment of the image and build a multi-parameter nonlinear model. However, due to the complexity of the parameter network of the CNN model, the internal workings of its neural network are considered to be a kind of ”black box”. [33] It is difficult to analyze the characteristics that networks focus on or rely on when they work. To solve this problem, the class activation mapping (CAM) algorithm, which can generate attention maps, is increasingly being applied to explain the reasoning behind CNN-based decision making and feature extraction. [34] It intuitively reveals the feature dependence of DL model in processing chemical sensing information. In this study, we designed a cone-patterned Janus membrane for sweat collection, combining the functional ad-vantages of fluorescence sensor and artificial intelligence algorithms for sweat collection and biomarker analysis. Inspired by the structure of cactus, a cone hydrophilic pattern was made on the hydrophobic layer TPU of electro-spinning. After unidirectional transfer of droplets from the hydrophobic layer of Janus membrane to the hydrophilic layer, the droplets were enriched to the centre region of the conical hydrophilic pattern by the Laplacian force, as shown in Figure 1a. After collecting sweat, the fluorescence hydrogel sensor obtains information about the types and concentrations of amino acids, Ca 2+ , and Cl - . Combined with artificial intelligence algorithm to process and analyze image data, amino acids, Ca 2+ , and Cl - concentrations in sweat can be easily classified and quantified with 100% classification accuracy. The proposed Janus membrane and fluorescence sensor can reliably collect and analyze sweat. It provides a new approach for the design and preparation of sensors with rapid sweat analysis capabilities, providing detection strategies for personalized health monitoring and early detection of developing health conditions. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Figure 1 Schematic diagram of Janus membrane for spontaneous in situ sweat collection and DL-assisted programmable fluorescent microsensor patch for biomarker analysis. (a) Sweat is unidirectionally transported from the hydrophobic layer of the Janus membrane to the hydrophilic layer, then from the hydrophilic layer’s conical tips to the designated location, where the hydrogel patch swells calibrated to the collected sweat volume and performs fluorescence analysis detection of amino acids, Ca 2+ , and Cl+ in the sweat. (b) Inspired by the conical spines of cacti and hydrophilic hairs, we designed conical hydrophilic patterns on a hydrophobic surface to manipulate droplets for continuous and effective unidirectional transport to designated locations. (Ⅰ: Hydrophobic area,: Ⅱ: Hydrophilic area) (c) Data collection through an intelligent photographic device. (d) The response information is input into an explainable CNN. (e) Classification and quantitative results of different biomarkers in the sweat are obtained. Results and Discussion Patterned Janus membranes for in situ sweat collection and detection and DL-assisted fluorescence sensor chips for sweat analysis. Sweat contains information related to human health and metabolism. The analysis of its components and concentrations can provide a new approach to non-invasive detection for health monitoring and instant diagnosis. However, the uneven distribution of sweat on the human body and its volatility pose significant challenges for in situ sweat collection and detection. Inspired by the conical structure of cacti, this paper 2D plans the cactus’s conical structure and designs a unique Janus membrane. This membrane integrates binary interface wettability (hydrophilic layer-hydrophobic layer) and geometric asymmetric pattern hydrophilic layer structure, cleverly combining the dual liquid collecting mechanism of wettability gradient and Laplace pressure gradient to achieve efficient directional sweat collection (Figure 1a). Sweat can be unidirectionally transported from the hydrophobic layer to the hydrophilic layer under the action of the wettability gradient, and then transported from the planarized conical tip to the central area under the action of both gradients, providing conditions for in situ sweat detection (Figure 1b). For the analysis of sweat components and concentrations, referencing previous work, we designed a programmable fluorescent hydrogel patch, prepared by crosslinking PVA and PAA. To satisfy the simultaneous detection of different components, various fluorescent indicators were selected to obtain fluorescent sensing information for the concentration analysis of sweat biomarkers. Here, we use 2-mercaptoethanol-o-phthalaldehyde to detect total amino acids, Calcium Green for calcium ions, and Quinine Sulfate for chloride ions. More details about the indicators and data collection equipment are provided in the supporting materials. Deep learning algorithms can perform a series of comprehensive nonlinear operations and have advantages in analyzing nonlinear and multidimensional data. After inputting the collected fluorescent image information into an explainable CNN (Figure 1d), classification and quantitative results of each component in the sweat can be obtained (Figure 1e). Additionally, we used the AI interpreter CAM to generate attention maps and explain the reasons behind the CNN-based decisions in sweat biomarker analysis. The evaluation results will serve as feedback to prove the rationality of the patch design and practice process, guide the evolution of programmable fluorescence methods, reduce algorithm waste, simplify operations, efficiently collect data, and pro-vide flexible and reliable biomarker detection strategies. Janus membrane structure and morphology characterization. The Janus membrane consists of three layers, as illustrated in Figure 2a. The first layer is a non-woven fabric serving as the support layer, primarily composed of polypropylene, with its morphology and fiber diameter shown in Figure S4. The second hydrophobic layer, obtained by electrospinning TPU for 50 minutes, has a contact angle of 130.6° ± 1.3° and exhibits fibrous morphology with a fiber diameter of 1.71 ± 0.49 μm. The density of fibers in the hydrophobic layer is related to the duration of electrospinning (Figure S3). The third layer, a patterned multi-conical hydrophilic layer, is formed after the deposition of PDA/PEI on the TPU fibers, displaying a porous structure with a contact angle of 36.3° ± 2.2°. The variation of the contact angle with deposition time and PEI concentration is shown in Figure S5a. In the FT-IR spectrum (Figure 2b), the characteristic peaks of polypropylene at 2917.8 cm -1 and 2870.5 cm -1 correspond to the stretching vibrations of -CH 3 , while 1374.1 cm -1 and 1452.6 cm -1 are attributed to the bending vibrations of -CH 3 and -CH 2 -. For TPU, the characteristic peaks at 3325.3 cm -1 and 2943.2 cm -1 are mainly due to the stretching and bending vibrations of -NH- and -CH- in the polyurethane, with the vibrational peaks at 1724.6 cm -1 and 1525.3 cm -1 belonging to the -NH-COO- group in the polyurethane. After the deposition of PDA/PEI on the TPU layer, the absorption peaks at 1618.6 cm -1 and 1507.3 cm -1 correspond to the C=C resonance vibration from aromatic Figure 2. Morphology and structural characterization of the Janus membrane for spontaneous in situ sweat collection. (a) SEM images of the hydrophilic and hydrophobic regions of the Janus membrane and their corresponding static contact angles. (b) FT-IR spectra of the three layers of the Janus membrane. (c) XPS spectra and N1s fine spectra of the hydrophobic and hydrophilic layers, (d) TPU and (e)PDA/PEI, respectively. rings and the bending vibration of -N-H-, indicating the presence of PDA, while 3286.3 cm -1 corresponds to the stretching vibration of -NH-. X-ray photoelectron spectroscopy (XPS) was used to verify the chemical composition of the prepared samples (Figure 2c). Both the TPU and PDA/PEI layers contain C, N, and O elements, with an increase in the peak intensity of N1s in the PDA/PEI layer, where the nitrogen content increased from 1.32 % to 6.17 %. In Figure 2e, the fine spectrum of N1s in PDA/PEI shows peaks at 399.59 eV and 401.48 eV, corresponding to C-N and C=N bonds, respectively. Compared to the 400.0 eV peak attributed to C-N in Figure 2d, this indicates that PEI undergoes Michael addition or Schiff base reactions between amine and catechol to form C=N bonds. Spontaneous droplet sampling. The unique wettability of the Janus membrane results in unidirectional water transport behavior, as shown in Figure S5, where water droplets penetrate from the hydrophobic layer to the hydrophilic layer. However, when the Janus membrane is flipped, the water droplet is blocked and only spreads on the hydrophilic layer without penetrating the membrane. The mechanism of unidirectional water transport of the Janus membrane is illustrated in Figure 3a. When a water droplet is on the hydrophobic layer, it initially maintains a Wenzel-Cassie state and is influenced by two opposing forces, namely HF (hydrophobic force) and HP (hydrostatic pressure). HF relates to the breakthrough pressure of the hydrophobic layer, preventing water penetration. For a specific membrane, HF has a constant value. Conversely, HP is proportional to the height of the water, enabling it to pass through vertical porous channels and penetrate the membrane. Once HP exceeds HF, the water droplet fully contacts the rough surface and transitions to a Wenzel state. In practice, the thicker the hydrophobic layer, the greater the HP required for complete penetration. Once water penetrates the hydrophobic layer into the hydrophilic layer, both CF (capillary force) and HP together facilitate the diffusion and penetration of water until it completely traverses the Janus membrane. Conversely, when a water droplet falls on the hydrophilic layer, the CF provided by the PDA/PEI hydrophilic layer causes the water to spread on the surface. When the water reaches the interface between the hydrophilic and hydrophobic layers, the HF provided by the hydrophobic layer blocks further penetration. An increase in water quantity on the hydrophilic layer only increases the diffusion area but does not elevate the water height; thus, the HP remains lower than the HF, preventing water from penetrating the hydrophobic layer. When a droplet reaches the hydrophilic layer, like the conical spines and hydrophilic hairs of cacti, the gradient wettability and conical hydrophilic patterns generate Laplace pressure, driving the unidirectional transport of the droplet. Figure 3b shows the continuous WCA (water contact angle) of TPU and Janus membranes. The TPU fiber membrane exhibits good hydrophobicity, with little change in the water droplet’s shape over time. The water droplet on the TPU layer of the Janus membrane penetrates to the hydrophilic layer within 6 seconds and quickly spreads on the PDA/PEI layer (Figure S6), demonstrating excellent unidirectional permeability from the hydrophobic side to the hydrophilic side. To further investigate the influence of water on both sides of the Janus membrane, the hydrostatic pressure (HP) required to pass water through the porous channels of the membrane was measured. With a fixed area of the Janus fabric membrane, water was slowly dripped from a tubular container, and the height of the water column at the start of penetration was recorded, measuring the HP values from the hydrophobic TPU side to the hydrophilic PDA/PEI side and vice versa. Figure 3c shows the relationship between TPU electrospinning time and HP. The HP from the hydrophobic TPU side to the hydrophilic layer is significantly lower than that in the reverse direction, indicating that water droplets more easily transmit from the hydrophobic side to the hydrophilic side. Figure 3. Unidirectional transport and enrichment of sweat. (a) Schematic diagram of the unidirectional transport mechanism of the Janus membrane (HP: Hydrostatic Pressure, HF: Hydrophobic Force, CF: Capillary Force). (b) Continuous WCA of the Janus membrane. (c) The relationship between HP on both sides of the Janus membrane and the electrospinning time used for the TPU fiber membrane. (d) Schematic diagram of unidirectional droplet transportation caused by the Laplace pressure gradient. (e) Droplet movement speed on hydrophilic layers with different conical pattern angles. (f) Droplet collection capacity in grams for Janus membranes with different numbers of cones on the hydrophilic layer. As the TPU electrospinning time increases, the corresponding HP increases due to the longer hydrophobic channels that the water must traverse, making water transport more difficult. By comparing the rate of increase in HP, an electrospinning time of 50 minutes was selected for the hydrophobic layer. Inspired by the cactus conical structure, a two-dimensional conical hydrophilic layer was prepared with an asymmetric structure that gradually enlarges from the tip to the base of the cone. When a droplet is transported to the conical pattern hydrophilic layer, it undergoes asymmetric deformation confined within the conical channels, forming droplets with different curvature radii. This is manifested by the difference in the radius of curvature (r 1 and r 2 ) at the three-phase contact line (TCL) between the front and back of the droplet (Figure 2d). The curvature radius near the tip is smaller, while the curvature radius near the base is relatively larger. Due to r 1 < r 2 , an imbalanced surface tension generates Laplace pressure (ΔP) from the side with a smaller curvature to the side with a larger curvature: \begin{equation} \Delta P=\gamma\left(\frac{1}{r_{1}}-\frac{1}{r_{2}}\right)\nonumber \\ \end{equation} where γ is the surface tension of the droplet. In this case, the droplet spontaneously moves toward the base of the conical pattern. The movement speed of the droplet is related to the angle α at the tip of the conical pattern. The larger the angle, the smaller the Laplace pressure difference, and hence the slower the droplet moves. Figure 2e compares the movement speeds for conical pattern tips with angles of 10.2°, 14.6°, 20°, and 22°. When the conical pattern angle is 10.2°, the movement speed can reach 20 mm/s. By fixing the conical angle, a Janus membrane with different conical patterns was prepared (Figure 2f). Compared to a Janus membrane with circular hydrophilic patterns, the Janus membrane with octagonal cone patterns had a 3.85-fold increase in droplet enrichment capability. The chemical sensing mechanism of sweat biomarkers. Due to the complexity of human sweat composition and the variability in sweat volume, quantitative detection of biomarkers in sweat has always been a challenge. To overcome this difficulty, we adopted a method that combines programmable fluorescent chips with explainable DL algorithms for biomarker detection. In this study, we selected amino acids (1×10 -5 ~ 9×10 -3 mol/L), Ca 2+ (1×10 -4 ~ 9×10 -2 mol/L), and Cl - (1×10 -4 ~ 9×10 -2 mol/L) as research subjects. The adequacy of total amino acid content is crucial for maintaining protein synthesis, cell repair, and overall physiological functions; Ca 2+ plays a key role in maintaining bone strength, nerve conduction, muscle contraction, and blood coagulation, and its deficiency may lead to osteoporosis and other health issues; Cl - mainly participates in the regulation of body fluid osmotic pressure and acid-base balance, which is critical for maintaining normal blood pressure and cardiac function, and its imbalance may trigger electrolyte disorders. These indicators are essential for assessing individual health status and guiding scientific training. The chemical sensing mechanism of the sweat marker fluorescence hydrogel patch and the corresponding spectra are shown in Figure 4. When 2-mercaptoethanol acts as a reducing agent, o-phthalaldehyde reacts with amino acid compounds through cyclization and condensation reactions to form strong fluorescent substances. At an excitation wavelength of 347 nm, the fluorescence intensity is directly proportional to the concentration of amino acids, as shown in Figure 4b. The measurement mechanism for Ca 2+ relies on the increased fluorescence upon binding of calcium yellow green with calcium ions to detect concentration (Figure 4c). For Cl - detection, quinine sulfate is chosen as a fluorescent Figure 4. Chemical sensing mechanism of the fluorescent hydrogel patch and corresponding spectra. (a) The chemical fluorescence response mechanism of indicators to sweat biomarkers. (b–d) Corresponding spectra for the response of selected indicators to different sweat biomarkers. (e) Optical images and fluorescence intensity plots for amino acids (1×10 -5 ~9×10 -3 mol/L), (f) Ca 2+ (1×10 -4 ~9×10 -2 mol/L), and (g) Cl - (1×10 -4 ~9×10 -2 mol/L). indicator, which exhibits intense blue emission fluorescence that quenched in the presence of chloride ions (Figure 4d). Figures 4e-g show the optical images and corresponding fluorescence spectral intensities of these sweat biomarkers at different concentration levels across various concentration gradients. The fluorescence intensity of the three biomarkers is only linearly related to concentration within a certain range. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf DL-assisted detection of sweat biomarkers using fluorescence chips. A total of 4500 fluorescence hydrogel patch photographic images were collected for 15 concentrations of amino acids (1×10 -5 ~ 9×10 -3 mol/L), Ca 2+ (1×10 -4 ~ 9×10 -2 mol/L), and Cl - (1×10 -4 ~ 9×10 -2 mol/L), forming an analytical dataset. The dataset was randomly divided into three parts: training, validation, and test sets at a ratio of 8:1:1. The training set was used for model training, the validation set for adjusting hyperparameters and preliminarily evaluating model capabilities, and the test set for assessing the final model’s generalization ability. In machine learning, linear discriminant analysis (LDA) results for classifying sweat biomarkers are shown in Figure 5a-c; it cannot completely distinguish between different concentrations of biomarkers. There is significant cluster overlap, with amino acid classification accuracy of 60.7-62.8 %, Ca 2+ classification accuracy of 33.3-46.3 %, and Cl - classification accuracy of 25.9-27.9 %. ML models fail to achieve sensitive multivariate fluorescence analysis. Subsequently, advanced analysis was conducted using DL algorithms. The ResNet-18 network, known for its CNN residual learning framework, can optimize the network and achieve accuracy by progressively increasing depth. We designed a reasonable CNN structure based on the ResNet-18 network and optimized the model’s hyperparameters, including the number of epochs, batch size, and learning rate (see Supplementary Table S1). CNN effectively extracts features from fluorescence response data and performs effective classification through confusion matrices (Figure 5d-f). As shown in Table S7, the selected three ions/molecules achieved 100% accuracy across all three subsets, demonstrating that the CNN model’s classification accuracy surpasses that of the LDA model. We investigated two DL or seven ML models, including CNN, artificial neural networks (ANN), extreme gradient boosting (XGBoost), decision trees (DT), K-nearest neighbors (KNN), logistic regression (LR), naive Bayes (NB), random forests (RF), and support vector machines (SVM) (model structures and parameters are detailed in Tables S2 and S3). While ANN, XGBoost, KNN, RF, and DT models showed good fitting capability in the amino acid classification task and nonlinear description, their classification accuracies for Ca 2+ and Cl - validation and test sets decreased to varying degrees. Only the CNN model displayed 100% accuracy across all three subsets (Figure S10 and Table S7). This is due to the following reasons: 1) Deep learning models often have complex structures that allow them to handle more intricate nonlinear relationships, leading to better performance in classification tasks. Additionally, by tuning the model’s hyperparameters, further optimization of the model’s performance can be achieved. 2) Deep learning models can automatically extract high-level feature representations from raw data. These features are often richer and more effective than those manually designed by traditional machine learning methods. Figures 5a-c present optical images of these sweat biomarkers at different concentration levels along with their corresponding fluorescence spectral responses; Figures 5d-f introduce a jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Figure 5. Classification and quantitative analysis of various sweat biomarkers using ML and DL models. (a) Three-dimensional LDA score plots for amino acids (1×10 -5 ~9×10 -3 mol/L), (b) Ca 2+ (1×10 -4 ~9×10 -2 mol/L), and (c) Cl - (1×10 -4 ~9×10 -2 mol/L). Confusion matrices for CNN predictions of (d) amino acids, (e) Ca 2+ , and (f) Cl - . (g-i) Match between actual and predicted concentrations in the test set for CNN-based quantitative analysis. (j-l) Accuracy assessment comparing the prediction results from CNN-assisted fluorescent patches with those from fluorescence spectroscopy measurements. convolutional neural network (CNN) quantification model we developed (for specific architecture, see Supplementary Table S4). This model aims to precisely predict the specific concentration values of the studied biomarkers. With this technology, our DL-assisted programmable fluorescence patch can accurately assess the concentrations of amino acids, Ca 2+ , and Cl- in actual samples and predicted samples within the specified range. To verify the accuracy of the established model, we employed various quantitative performance metrics for evaluation (as shown in Tables S8-10), including the coefficient of determination (R²), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). Scatter plot analysis of the relationship between predicted results and true values on the test set revealed that the model performed excellently, with R² values exceeding 0.999. Additionally, the linear fit slope between predicted and actual values was very close to the ideal state of 1, indicating a high degree of consistency between them. Moreover, when comparing CNN with several other common deep learning/machine learning algorithms such as artificial neural networks (ANN), extreme gradient boosting trees (XGBoost), decision trees (DT), K-nearest neighbors (KNN), linear regression (LR), random forests (RF), and support vector machines (SVM). Although some methods like ANN, XGBoost, and RF achieved good performance in amino acid detection modeling, they exhibited varying degrees of overfitting issues during subsequent modeling of the other two ions, failing to demonstrate good generalization ability. In contrast, CNN not only significantly reduced the possibility of overfitting but also exhibited superior generalization performance (more details can be found in Tables S5-6 and S8-10). Individual differences can affect sweat biomarkers. Leveraging the excellent quantitative analysis capability of the DL-assisted fluorescent patch, we further investigated the concentration levels of sweat biomarkers in six subjects under exercise conditions. The deep learning predictions for different subjects’ biomarker concentrations are shown in Figure S11a-c. The total concentration of amino acids in human sweat is about 10 -3 mol/L, Ca 2+ concentration is about 10 -4 to 10 -3 mol/L, and Cl - concentration is about 10 -2 mol/L. These results are consistent with the reported normal physiological range for humans. To validate the accuracy of the DL-assisted programmable patch, we compared it with laboratory fluorescence spectroscopy (Figures 5j-l). We found that the test results from the DL-assisted fluorescent patch matched the laboratory measurements with a rate of 91.4-96.0%. The above results effectively demonstrate the reliable analysis and practical applicability of the DL-assisted programmable fluorescence patch in quickly and accurately evaluating sweat. The interpretability of CNN for fluorescent patch. During the CNN modeling process, the internal structure extracts feature information from images through layer-by-layer convolution operations, which is highly abstract, making it difficult to fully understand the internal workings of the DL model. CAM technology can uncover the internal mechanisms of the network, making the sensing process visualized. Figure S12a shows a schematic diagram of the CAM illustrating the response data classification feature extraction. Each neuron is a linear combination of neurons from the previous layer, obtained through a series of nonlinear functions. As shown in the CAM of Figure S12b, the heatmap displays the attention mechanism during the CNN classification operation. The redder the color, the higher the network’s attention; the bluer the color, the lower the attention. Based on the CAM maps of amino acids, Ca 2+ , and Cl - , we can derive the following mechanism conclusions: (1) The CNN primarily focuses on regions where fluorescent intensity changes in the fluorescent sensing patch and utilizes this information without additional human intervention (end-to-end). (2) The CAM activity (red areas) matches with the chemical sensing design on the fluorescent sensing patch. The designed fluorescent hydrogel sensing patch provides the corresponding features needed for DL analysis, and the interpretability of the DL model can be validated by testing the performance of the network model to eliminate potential errors or risks. By testing the performance of the network model, CAM can be used to facilitate software design, extract relevant features, and optimize theory. It provides convenient indicators for health management, disease prevention, and even new scientific discoveries. Conclusions In summary, we developed a wearable cactus-inspired patterning Janus membrane and explainable DL-assisted fluorescence hydrogel sensing analysis platform to enable spontaneous sweat collection and efficient and accurate detection of biomarkers. The patterned Janus membrane was obtained by fabricating a conical patterned hydrophilic layer on the electrospun TPU. Combined with hydrostatic pressure and Laplacian pressure, a dual process of unidirectional transport of sweat from hydrophobic layer to hydrophilic layer and unidirectional movement and enrichment on conical hydrophilic layer is realized. Based on the advantages of explainable DL algorithms, we developed a CNN-assisted PAA-PVA hydrogel fluorescence sensor with fluorescence indicator fixation. A data set of 4,500 fluorescence images were collected as an analytical dataset, and two DL algorithms and seven machine learning (ML) algorithms were evaluated. The CNN model can easily accurately classify and quantify the concentration of amino acids, Ca 2+ and Cl - in sweat, achieving 100% accurate classification and 99.9% quantitative prediction with a wide range. DL-assisted fluorescence sensor detection of actual sweat and laboratory detection methods matched 91.4-96.0%, providing a simple and convenient index for rapid detection, disease intervention and clinical diagnosis. The cactus-inspired sweat osmosis-enrichment patch can achieve the effect of in-situ perspiration collection under the premise of biosafety isolation for long-term wearable sweat metabolism monitoring. Experimental Fabrication of the Patterned Janus Membrane. The hydrophobic layer TPU (15%) was spun on the non-woven fabric by electrospinning technology. The spinning solution was prepared by dissolving 1.8 g TPU in 5 mL DMF and 5 mL THF. Then the mixed spinning solution was transferred to a syringe (10 mL). In the experiment, a high voltage power supply was used to apply a voltage of 15 kV. The diameter of the syringe needle was 0.8 mm and the injection rate was 0.08 mL/h. The TPU was received on a rotary receiver fixed with non-woven fabric, and the receiver was 10 cm away from the syringe. The hydrophobic TPU was obtained after spinning for a certain time. Subsequently, molds with different numbers of cones were 3D printed. A thin layer of polydimethylsiloxane was applied to the molds and then heated in an oven at 70°C for 3 hours to obtain hydrophobic molds. The molds with varying hydrophobic patterns were fixed on the TPU hydrophobic layer. A certain volume of dopamine hydrochloride and polyethyleneimine (DA/PEI) solution was added to the molds. After 12 hours, patterned Janus membranes with circular, double-cone, quadruple-cone, and octuple-cone hydrophilic patterns were obtained (Figure S1). jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Preparation of Programmable Fluorescent Hydrogel Chips. Fluorescent indicators (including 2-mercaptoethanol-o-phthalaldehyde, Calcium Green, Quinine Sulfate) were prepared and solidified into PVA-PVA fluorescent hydrogel patches for detecting biomarkers in sweat. A certain amount of fluorescent indicators, 1.000 g PAA, and 4.000 g PVA were added to 100 mL of deionized water. The solution was stirred and dissolved at 100°C in an oil bath pot and then left to stand for 12 hours to remove bubbles, resulting in a fluorescent indicator-fixed PVA-PVA hydrogel solution (Figure S2). The hydrogel solution was poured into surface plates with a depth of 5 mm and dried using a dehumidifier for 12 hours to obtain fluorescent hydrogel films with a thickness of 0.10 ± 0.03 mm. Finally, a hole punch with a diameter of 5 mm was used to cut out several patches with a direct size of 5.0 ± 0.03 mm for detecting biomarkers in sweat. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Supporting Information The supporting information for this article is available on the WWW under https://doi.org/10.1002/cjoc.2021xxxxx. jabbrv-ltwa-all.ldf jabbrv-ltwa-en.ldf Acknowledgement (optional) This work is supported by the National Natural Science Foun-dation of China (22474049). References 1. Davis, N.; Heikenfeld, J.; Milla, C.; Javey, A. The Challenges and Promise of Sweat Sensing. Nat. Biotechnol . 2024 , 42, 860-871. 2. Arwani, R. T.; Tan, S. C. L.; Sundarapandi, A.; Goh, W. P. Stretchable Ionic–Electronic Bilayer Hydrogel Electronics Enable in Situ Detection of Solid-State Epidermal Biomarkers. Nat. Mater. 2024 , 23, 1115-1122. 3. Xu, Y.; De la Paz, E.; Paul, A.; Mahato, K.; Sempionatto, J. R.; Tostado, N.; Lee, M.; Hota, G.; Lin, M. In-Ear Integrated Sensor Array for the Continuous Monitoring of Brain Activity and of Lactate in Sweat. Nat. Biomed. Eng . 2023 , 7, 1307-1320. 4. Ye, C.; Wang, M.; Min, J.; Tay, R. Y.; Lukas, H.; Sempionatto, J. R.; Li, J.; Xu, C.; Gao, W. A Wearable Aptamer Nanobiosensor for Non-Invasive Female Hormone Monitoring. Nat. Nanotechnol. 2024 , 19, 330-337. 5. Min, J.; Tu, J.; Xu, C.; Lukas, H.; Shin, S.; Yang, Y.; Solomon, S. A.; Mukasa, D.; Gao, W. Skin-Interfaced Wearable Sweat Sensors for Precision Medicine. Chem. Rev. 2023 , 123, 5049-5138. 6. Zhang, K.; Zhang, J.; Wang, F.; Kong, D. Stretchable and Superwettable Colorimetric Sensing Patch for Epidermal Collection and Analysis of Sweat. ACS Sens. 2021 , 6, 2261-2269. 7. Han, Y.; Fang, X.; Li, H.; Zha, L.; Guo, J.; Zhang, X. Sweat Sensor Based on Wearable Janus Textiles for Sweat Collection and Microstructured Optical Fiber for Surface-Enhanced Raman Scattering Analysis. ACS Sens . 2023 , 8, 4774-4781. 8. Zhang, Y.; Chen, Y.; Huang, J.; Liu, Y.; Peng, J.; Chen, S.; Song, K.; Ouyang, X.; Cheng, H.; Wang, X. Skin-Interfaced Microfluidic Devices with One-Opening Chambers and Hydrophobic Valves for Sweat Collection and Analysis. Lab. Chip 2020 , 20, 2635-2645. 9. Wu, T.; Yang, P.; Xie, X.; Cao, X.; Deng, Y.; Ding, X.; Zhang, Z. Bio-Inspired Hierarchical Wearable Patch for Fast Sweat Collection. Biosens. Bioelectron. 2024 , 260, 116430. 10. Ju, J.; Huang, Y.; Liu, M.; Xie, N.; Shi, J.; Fan, Y.; Zhao, Y.; Kang, W. Construction of Electrospinning Janus Nanofiber Membranes for Efficient Solar-Driven Membrane Distillation. Sep. Purif. Technol. 2023 , 305, 122348. 11. Sun, X.; Bai, L.; Li, J.; Huang, L.; Sun, H.; Gao, X. Robust Preparation of Flexibly Super-Hydrophobic Carbon Fiber Membrane by Electrospinning for Efficient Oil-Water Separation in Harsh Environments. Carbon 2021 , 182, 11-22. 12. Liu, H.; Gu, J.; Liu, Y.; Yang, L.; Wang, L.; Yu, J.; Qin, X. Reconfiguration and Self-Healing Integrated Janus Electrospinning Nanofiber Membranes for Durable Seawater Desalination. Nano Res. 2023 , 16, 489-495. 13. Yang, P.; Wei, G.; Liu, A.; Huo, F.; Zhang, Z. A Review of Sampling, Energy Supply and Intelligent Monitoring for Long-Term Sweat Sensors. Npj Flex. Electron. 2022 , 6, 1-13. 14. Son, J.; Bae, G. Y.; Lee, S.; Lee, G.; Kim, S. W.; Kim, D.; Chung, S.; Cho, K. Cactus-Spine-Inspired Sweat-Collecting Patch for Fast and Continuous Monitoring of Sweat. Adv. Mater. 2021 , 33, 2102740. 15. Wang, L.; Xu, T.; He, X.; Zhang, X. Flexible, Self-Healable, Adhesive and Wearable Hydrogel Patch for Colorimetric Sweat Detection. J. Mater. Chem. C. 2021 , 9, 14938–14945. 16. Xiao, J.; Liu, Y.; Su, L.; Zhao, D.; Zhao, L.; Zhang, X. Microfluidic Chip-Based Wearable Colorimetric Sensor for Simple and Facile Detection of Sweat Glucose. Anal. Chem. 2019 , 91, 14803-14807. 17. Cui, Y.; Duan, W.; Jin, Y.; Wo, F.; Xi, F.; Wu, J. Ratiometric Fluorescent Nanohybrid for Noninvasive and Visual Monitoring of Sweat Glucose. ACS Sens . 2020 , 5, 2096-2105. 18. Ardalan, S.; Hosseinifard, M.; Vosough, M.; Golmohammadi, H. Towards Smart Personalized Perspiration Analysis: An IoT-Integrated Cellulose-Based Microfluidic Wearable Patch for Smartphone Fluorimetric Multi-Sensing of Sweat Biomarkers. Biosens. Bioelectron . 2020 , 168, 112450. 19. Wang, Z.; Shin, J.; Park, J.-H.; Lee, H.; Kim, D.-H.; Liu, H. Engineering Materials for Electrochemical Sweat Sensing. Adv. Funct. Mater. 2021 , 31, 2008130. 20. Ma, H.; Li, J.; Zhou, J.; Luo, Q.; Wu, W.; Mao, Z.; Ma, W. Screen-Printed Carbon Black/Recycled Sericin @Fabrics for Wearable Sensors to Monitor Sweat Loss. ACS Appl. Mater. Interfaces. 2022 , 14, 11813-11819. 21. Wang, S.; Liu, M.; Yang, X.; Lu, Q.; Xiong, Z.; Li, L.; Zheng, H.; Feng, S.; Zhang, T. An Unconventional Vertical Fluidic-Controlled Wearable Platform for Synchronously Detecting Sweat Rate and Electrolyte Concentration. Biosens. Bioelectron. 2022 , 210, 114351. 22. Wang, D.; Xu, G.; Zhang, X.; Gong, H.; Jiang, L.; Sun, G.; Li, Y.; Liu, G.; Li, Y.; Yang, S.; Liang, X. Dual-Functional Ultrathin Wearable 3D Particle-in-Cavity SF-AAO-Au SERS Sensors for Effective Sweat Glucose and Lab-on-Glove Pesticide Detection. Sens. Actuators B Chem. 2022 , 359, 131512. 23. Chung, M.; Skinner, W. H.; Robert, C.; Campbell, C. J.; Rossi, R. M.; Koutsos, V.; Radacsi, N. Fabrication of a Wearable Flexible Sweat pH Sensor Based on SERS-Active Au/TPU Electrospun Nanofibers. ACS Appl. Mater. Interfaces. 2021 , 13, 51504-51518. 24. Zhu, K.; Yang, K.; Zhang, Y.; Yang, Z.; Qian, Z.; Li, N.; Li, L.; Jiang, G.; Wang, T.; Zong, S.; Wu, L.; Wang, Z.; Cui, Y. Wearable SERS Sensor Based on Omnidirectional Plasmonic Nanovoids Array with Ultra-High Sensitivity and Stability. Small. 2022 , 18, 2201508. 25. Xu, J.; Qin, G.; Luo, F.; Wang, L.; Zhao, R.; Li, N.; Yuan, J.; Fang, X. Automated Stoichiometry Analysis of Single-Molecule Fluorescence Imaging Traces via Deep Learning. J. Am. Chem. Soc. 2019 , 141, 6976-6985. 26. Krishnan, S. R.; Bung, N.; Vangala, S. R.; Srinivasan, R.; Bulusu, G.; Roy, A. De Novo Structure-Based Drug Design Using Deep Learning. J. Chem. Inf. Model. 2022 , 62, 5100-5109. 27. Yang, X.; Wang, Y.; Byrne, R.; Schneider, G.; Yang, S. Concepts of Artificial Intelligence for Computer-Assisted Drug Discovery. Chem. Rev . 2019, 119, 10520-10594. 28. Ju, L.; Lyu, A.; Hao, H.; Shen, W.; Cui, H. Deep Learning-Assisted Three-Dimensional Fluorescence Difference Spectroscopy for Identification and Semiquantification of Illicit Drugs in Biofluids. Anal. Chem. 2019 , 91, 9343-9347. 29. Ayres, L. B.; Gomez, F. J. V.; Linton, J. R.; Silva, M. F.; Garcia, C. D. Taking the Leap between Analytical Chemistry and Artificial Intelligence: A Tutorial Review. Anal. Chim. Acta 2021 ,1161, 338403. 30. Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; Torralba, A. Learning Deep Features for Discriminative Localization. IEEE CVPR. 2016 ,2921-2929. 31. Wang, Q.; Ye, M.; Wei, M.; Lian, G.; Li, Y. Deep Convolutional Neural Network Based Closed-Loop SOC Estimation for Lithium-Ion Batteries in Hierarchical Scenarios. Energy. 2023 , 263, 125718. 32. Nguyen, H.-T.; Cheah, C. C.; Toh, K.-A. An Analytic Layer-Wise Deep Learning Framework with Applications to Robotics. Automatica. 2022 , 135, 110007. 33. Tan, X.; Liang, Y.; Ye, Y.; Liu, Z.; Meng, J.; Li, F. Explainable Deep Learning-Assisted Fluorescence Discrimination for Aminoglycoside Antibiotic Identification. Anal. Chem. 2022 , 94, 829–836. 34. Zhang, J.; Liu, Z.; Tang, Y.; Wang, S.; Meng, J.; Li, F. Explainable Deep Learning-Assisted Self-Calibrating Colorimetric Patches for In Situ Sweat Analysis. Anal. Chem. 2024 , 96, 1205-1213. (The following will be filled in by the editorial staff) Manuscript received: XXXX, 2021 Manuscript revised: XXXX, 2021 Manuscript accepted: XXXX, 2021 Accepted manuscript online: XXXX, 2021 Version of record online: XXXX, 2021 Entry for the Table of Contents In this work, we proposed a biomimetic Janus membrane, inspired by the structure of cactus, that could unidirectionally transport and converge sweat to the assigned detection point. The bionic osmosis-enrichment sensing patch effectively inhibits direct-contact of indicators to skin, eliminating potential epidermal contamination. Over 4,500 fluorescence datasets were evaluated, and the deep-learning model could quickly and accurately classify and quantitatively analyze the concentrations of amino acids, Ca 2+ and Cl - , with 100% classification accuracy. The cactus-inspired sweat osmosis-enrichment patch shows good reliability in collecting and analyzing sweat, and provided a simple index for clinical health monitoring, disease intervention prevention and clinical diagnosis. Information & Authors Information Version history V1 Version 1 26 June 2025 Peer review timeline Published Biosensors Version of Record 1 Dec 2025 Published Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords biosafety isolation cactus-inspired deep learning osmosis-enrichment wearable sweat detection Authors Affiliations Yuwen Yan Jinan University College of Chemistry and Materials Science View all articles by this author Miaorong Lin Jinan University College of Chemistry and Materials Science View all articles by this author Yueli Luo Henan University School of Physics and Electronics View all articles by this author Jihan Qu Jinan University College of Chemistry and Materials Science View all articles by this author Yonghuan Chen Jinan University College of Chemistry and Materials Science View all articles by this author Zhihao Zhang Zhengzhou University College of Chemistry View all articles by this author Congxin Xia Henan University School of Physics and Electronics View all articles by this author Jianxin Meng Jinan University College of Chemistry and Materials Science View all articles by this author Fengyu Li 0000-0003-2481-6111 [email protected] Jinan University College of Chemistry and Materials Science View all articles by this author Metrics & Citations Metrics Article Usage 240 views 145 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Yuwen Yan, Miaorong Lin, Yueli Luo, et al. Deep-Learning Assisted Cactus-inspired Osmosis-enrichment for Biosafety-isolative Wearable Sweat Metabolism Assessment. Authorea . 26 June 2025. DOI: https://doi.org/10.22541/au.175092536.60750459/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.175092536.60750459/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ffb1e0c3f36e2c5',t:'MTc3OTQ0NTUzMQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();