ANN-Based Discernment of Septic and Inflammatory Synovial Fluid: A Novel Method Using Viscosity Data from a QCR Sensor

The synovial fluid (SF) analysis involves a series of chemical and physical studies that allow opportune diagnosing of septic, inflammatory, non-inflammatory, and other pathologies in joints. Among the variety of analyses to be performed on the synovial fluid, the study of viscosity can help distinguish between these conditions, since this property is affected in pathological cases. The problem with viscosity measurement is that it usually requires a large sample volume, or the necessary instrumentation is bulky and expensive. This study compares the viscosity of normal synovial fluid samples with samples with infectious and inflammatory pathologies and classifies them using an ANN (Artificial Neural Network). For this purpose, a low-cost, portable QCR-based sensor (10 MHz) was used to measure the viscous responses of the samples by obtaining three parameters: Δf, ΔΓ (parameters associated with the viscoelastic properties of the fluid), and viscosity calculation. These values were used to train the algorithm. Different versions of the ANN were compared, along with other models, such as SVM and random forest. Thirty-three samples of SF were analyzed. Our study suggests that the viscosity characterized by our sensor can help distinguish infectious synovial fluid, and that implementation of ANN improves the accuracy of synovial fluid classification.


Introduction
Synovial fluid (SF) is a viscous liquid located in the joints whose primary functions are twofold. The first is the joint's mechanical function, which involves lubricating the articular surface and cushioning movements. The second one is to contribute to the nutrition of the articular cartilage by acting as a nutrient transport medium. It is composed of dialysate of plasma and a high content of hyaluronic acid (HA), which is responsible for its viscosity [1,2].
The analysis of SF begins with the extraction of the sample (by arthrocentesis), which involves a joint puncture. Then, the sample is collected in tubes containing anticoagulants such as EDTA (ethylenediaminetetraacetic acid) and lithium heparin [1,3,4]. Regarding volume, the maximum amount obtained from a normal joint is between 0.1 and 3.5 mL. The knee can have up to 4 mL. The volume required depends on the analysis (and may vary between laboratories). For example, for an accurate cell count, approximately 1 mL is required; 2 to 3 mL is an adequate volume to perform the complete tests needed. If a low-volume sample is obtained, it should be sent for the analysis of crystals and culture, which are more useful for the diagnostic [1,4,5].
To determine the viscosity (η), it was usual to observe the stranding, i.e., to measure the "thread" formed by the liquid when extended. This can be done by placing the sample drop on a slide and lifting it with a spatula or using the thumb and forefinger to spread it out. The "thread" may measure between 3 and 6 cm for a healthy fluid. SF with poor viscosity will form a "thread" of less than 3 cm [1,4,6]. Being a subjective method, as it depends on the operator's skills and experience, its use has been decreasing. As an objective assessment of viscosity, it is possible to use a viscometer or rheometer; however, they usually require more sample volume than is available or are expensive and large.
HA concentration determines the SF's viscoelastic properties. Arthritic diseases are associated with the reduction of HA [4,5,7,8]. In healthy SF, the concentration is around 3.5 mg/mL, whereas in osteoarthritis (OA), the HA concentration decreases to 1.3 mg/mL, and in rheumatoid arthritis (RA) to approximately 0.84 mg/mL [9]. This reduction in HA leads to a decrease in SF viscosity [5]. Joint diseases increase the risk of septic arthritis, which requires prompt diagnosis, as it is essential to provide the treatment as soon as possible [10].
Within the ViSQCT project of the Universidad Politécnica de Madrid (UPM), we developed a prototype sensor whose operation is based on the use of QCR. Its use in characterizing the viscosity of hydrogel formation has been previously demonstrated [30]. In a previous study [11], its operation was detailed, and its usefulness in measuring samples of artificial synovial fluid was tested. The sensor's objective is to measure the viscosity of a fluid with a small sample volume and to use this information to discriminate between pathologies and thus provide a timely diagnosis.
As part of the development of the device, an Artificial Neural Network (ANN) was implemented to optimize the classification of SF samples. ANNs have made inroads in biomedical engineering thanks to their ability to find relationships between data for prediction or classification [31,32]. Some examples of their use in biomedical applications can be seen in [31][32][33][34]. Additionally, their use with QCM sensors can be seen in the works [35][36][37]. In this work, we show the application of and comparison between parameters of an ANN to classify synovial fluid as inflammatory or infectious. This was done with data obtained from measurements performed with the QCR-based sensor. As a comparison, two other classification models were trained: support vector machine (SVM) and random forest (RF). SVM models are related to multilayer ANNs, and their operation is based on establishing a boundary (margin) that separates the two classes [38]. On the other hand, RF is an ensemble learning technique that has gained popularity due to its great capacity for classification [39].
The main contributions of the paper can be summarized as follows: • It is demonstrated that the ViSQCT sensor effectively measures the viscosity change in low-volume samples of SF. • A complete methodology is proposed to differentiate between inflammatory and infectious SF. • We show that using classification models such as ANN improves the methodology by increasing classification accuracy. • We compare the performance of the methodology and the system when using SF samples stored in two types of tubes (tubes with EDTA and tubes with lithium heparin) and evaluate their influences on making an accurate differentiation.
The present work shows the use of a portable and low-cost (less than EUR 200) QCR-based sensor named "ViSQCT" (developed in-house at the UPM) which allows the characterization of the viscosity of a small volume sample (few microliters) to classify between inflammatory and septic SF. The ethics committees of both the hospital and the university approved this work.

Sensor
The sensor used has been developed as part of the ViSQCT project of the Bioinstrumentation and Nanomedicine Laboratory (LBN) of the UPM. A complete description can be found in [11]. Its basis of operation is the use of the series resonance frequency ( f s ) of the QCR. Resonance frequency obtention is achieved by exciting the crystal with a frequency sweep near the fundamental resonance frequency and obtaining the conductance curve. With this, we locate the frequency where the maximum conductance is. The frequency shift (Equation (1)) is obtained by doing this process in air (without sample) and then with the sample deposited on the crystal. The Kanazawa relationship gave the connection between the frequency shift and the density-viscosity product of the fluid in contact with the crystal (Equation (2)) [40]. The half-bandwidth at half-maximum (Γ) is also acquired from the conductance curve, and like the resonance frequency case, the shift ∆Γ is obtained. This parameter is related to the energy transferred from the crystal to the sample over time and can provide information on the viscoelastic properties of the sample [41].
where ρ q = 2.648 gcm −3 and G q = 2.947 × 10 10 Nm −2 are the specific density and the shear modulus of quartz, respectively; f 0 is the fundamental resonance frequency of the quartz; f s is the series resonance frequency of the crystal loaded; ρ L is the fluid's density; η L is the fluid's viscosity; ∆ f is the frequency shift; and finally, n is the overtone number. In this work, the fundamental frequency of the crystal was used; thus, n was 1. This work was performed using QCR with f 0 = 10 MHz, gold electrodes, 5 and 11 mm electrode dimensions, roughness < 1 nm, and mounted in HC-51 holder. The crystals were purchased from Krystaly (Hradec Králové, Czech Republic).

Experimental Set-Up
The experimental setup is illustrated in Figure 1. The QCR was placed inside the holder cell where the liquid sample was dropped. The sample volume was 50 µL, since it was to cover the crystal's surface entirely and not completely evaporate. Experiments were performed at room temperature. Each experiment was repeated three to five times. Each experiment lasted 5 min, wherein 50 measured points were obtained (1 point every 6 s). In this way, the dataset was formed. The parameters ∆ f , ∆Γ, and η obtained from ∆ f were measured. After each experiment, the crystal was cleaned using a 2% solution of sodium dodecyl sulfate, rinsed with distilled water, disinfected with 70% ethanol, and then rinsed again with distilled water. Finally, the electrode surface was dried with air.

Experimental set up 123
The experimental setup is illustrated in Figure 1; the QCR is placed inside the holder 124 cell where the liquid sample is dropped. The sample volume is 50 µl since it covers the 125 crystal's surface entirely and prevents the liquid's complete evaporation. Experiments were 126 performed at room temperature. Each experiment was repeated three to five times. Each 127 experiment lasted 5 minutes, where 50 measured points were obtained (1 point every 6 128 seconds). In this way, the data set was formed. The parameters ∆ f , ∆Γ, and η obtained from 129 ∆ f were measured. After each experiment, the crystal was cleaned using a 2% solution of 130 sodium dodecyl sulfate, rinsed with distilled water, disinfected with 70% ethanol, and then 131 rinsed again with distilled water. Finally, the electrode surface is dried with air.

133
Statistical analysis was performed with SPSS, statistical software. Means are expressed 134 as mean ± standard deviation. Mann-Whitney U was used for analytic comparison; 135 p-values less than 0.05 were considered statistically significant. The predictive abilities 136 regarding septic SF of ∆ f , ∆Γ, and η were expressed as the area under the receiver operating 137 characteristic curve (AUC-ROC), AUC values are reported with the 95% confidence interval 138 (95% CI). ]. These tools help find relationships between data and 143 also in classification and prediction. They can also improve their performance by using 144 information obtained from previous tasks. The basic model of these ANN (known as the 145 multilayer perceptron model) is shown in Figure 2. It comprises three layers: an input 146 layer, an output layer, and hidden layers (HL). This model allows information to flow in 147 one direction, from input to output, and is known as a feedforward neural network. This 148 way, data will enter the network's input nodes, then be processed in the hidden layers, and 149 finally deliver processed data to the output layer [? ].

Statistical Analysis
Statistical analysis was performed with SPSS, statistical software. Means are expressed as mean ± standard deviation. Mann-Whitney U was used for analytic comparison; p-values less than 0.05 were considered statistically significant. The predictive abilities regarding septic SF of ∆ f , ∆Γ, and η were expressed as the area under the receiver operating characteristic curve (AUC-ROC); AUC values are reported with their 95% confidence intervals (95% CI).

Artificial Neural Networks
ANNs are a case of Artificial Intelligence (AI) that, based on examples, can induce concepts. They are data processing systems whose operation is based on the networks of neurons in the brain [31,32]. These tools help find relationships between data and also in classification and prediction. They can also improve their performances by using information obtained from previous tasks. The basic model of the ANN (known as the multilayer perceptron model) is shown in Figure 2. It comprises three layers: an input layer, an output layer, and hidden layers (HL). This model allows information to flow in one direction, from input to output, and is known as a feedforward neural network. This way, data will enter the network's input nodes, then be processed in the hidden layers, and finally be delivered to the output layer [32].

Experimental set up 123
The experimental setup is illustrated in Figure 1; the QCR is placed inside the holder 124 cell where the liquid sample is dropped. The sample volume is 50 µl since it covers the 125 crystal's surface entirely and prevents the liquid's complete evaporation. Experiments were 126 performed at room temperature. Each experiment was repeated three to five times. Each 127 experiment lasted 5 minutes, where 50 measured points were obtained (1 point every 6 128 seconds). In this way, the data set was formed. The parameters ∆ f , ∆Γ, and η obtained from 129 ∆ f were measured. After each experiment, the crystal was cleaned using a 2% solution of 130 sodium dodecyl sulfate, rinsed with distilled water, disinfected with 70% ethanol, and then 131 rinsed again with distilled water. Finally, the electrode surface is dried with air.

Statistical analysis 133
Statistical analysis was performed with SPSS, statistical software. Means are expressed 134 as mean ± standard deviation. Mann-Whitney U was used for analytic comparison; 135 p-values less than 0.05 were considered statistically significant. The predictive abilities 136 regarding septic SF of ∆ f , ∆Γ, and η were expressed as the area under the receiver operating 137 characteristic curve (AUC-ROC), AUC values are reported with the 95% confidence interval 138 (95% CI). ]. These tools help find relationships between data and 143 also in classification and prediction. They can also improve their performance by using 144 information obtained from previous tasks. The basic model of these ANN (known as the 145 multilayer perceptron model) is shown in Figure 2. It comprises three layers: an input 146 layer, an output layer, and hidden layers (HL). This model allows information to flow in 147 one direction, from input to output, and is known as a feedforward neural network. This 148 way, data will enter the network's input nodes, then be processed in the hidden layers, and 149 finally deliver processed data to the output layer [? ].  The ANN was applied using the algorithm illustrated in the diagram in Figure 3. As shown in Figure 2, the input data were the parameters ∆ f , ∆Γ, and η obtained with the sensor. The output values (or labels) were some the two possible diagnoses provided by the hospital (inflammatory and infectious SF). Having a larger amount of inflammatory SF samples (imbalanced data), the algorithm was tested using the imbalanced data and then with balanced data. The balanced data were obtained by randomly oversampling the septic SF data, thereby achieving the same data for both classifications.
The dataset size was 4972 data for samples in tubes with EDTA and 5248 for samples in tubes with lithium heparin. After loading the input data, the data were randomly segmented for the training, validation, and test phases as 70, 15, and 15%. Thus, 70% of the dataset was used for training, 15% was used for validation, and the remaining 15% was isolated for testing with the trained model. This way, we had three datasets: training, validation, and test. The training dataset contained the examples used during the learning process and was used to adjust the parameters. A validation dataset was a set of examples used to adjust the hyperparameters. The test dataset was a separate dataset from the training dataset used to test the model after training. After the data splitting, the data were optimized by scaling them to a range of values between 0 and 1. A robust scaler was employed, which scales the information according to the quantile range, making it robust against outliers. Figure 4 shows the steps of the ANN model.
The accuracy value was obtained for each case to observe the algorithm's performance. Accuracy is obtained from the fraction of the total number of correct predictions divided by the sum of all predictions (Equation (3) The ANN was applied using the algorithm illustrated in the diagram in Figure 3. As 151 shown in Figure 2, the input data will be the parameters ∆ f , ∆Γ, and η obtained with the 152 sensor. The output values (or labels) will be the two diagnostics provided by the hospital 153 (inflammatory and infectious SF). Having a larger amount of inflammatory SF samples 154 (imbalanced data), the algorithm was tested using the imbalanced data and then with the 155 balanced data. The balanced data was obtained by randomly oversampling the septic SF 156 data, thus achieving the same data for both classifications.

157
The dataset size was 4972 data for samples in tubes with EDTA and 5248 for samples in 158 tubes with lithium heparin. After loading the input data, the data was randomly segmented 159 for the training, validation, and test phases with a weighting of 70, 15, and 15%. Thus, 160 70% of the dataset was used for training, 15% for validation, and the remaining 15% was 161 isolated for testing with the trained model. This way, we will have three datasets: training, 162 validation, and test. The training dataset contains the examples used during the learning 163 process and is used to adjust the parameters. A validation dataset is a set of examples used 164 to adjust the hyperparameters. And the test dataset is a separate dataset from the training 165 dataset used to test the model after training. After the data splitting, the data was optimized 166 by scaling it to a range of values between 0 and 1. A robust scaler was employed, which 167 scales the information according to the quantile range, making it robust against outliers. 168 Figure 4 shows the steps of the ANN model.  Finally, to compare different ANN configurations, the HL of the networks were varied 175 between 1 and 2 layers, and the number of training epochs between 100, 200, and 300. These 176 configurations are shown in Table 2. Parameters such as the optimizer, activation function, 177 biases, etc., were left constant since it is beyond the scope of this work to go into this topic 178 in more detail. A more extensive study with a more significant number of configurations is 179 possible as the field of ANN is vast; however, it is beyond the intended scope of this paper. 180 Finally, to compare different ANN configurations, the HL of the networks were varied between 1 and 2 layers, and the number of training epochs among 100, 200, and 300. These configurations are shown in Table 2. Parameters such as the optimizer, activation function, and biases were left constant, since it is beyond the scope of this work to go into this topic in more detail. A more extensive study with a more significant number of configurations is possible, as the field of ANN is vast; however, this is beyond the intended scope of this paper.   All algorithms were developed using the Keras and sci-kit learn libraries in Pyth We used a linear kernel for SVM, c = 1, loss = 'squared hinge'. While for RF, we us 2171 trees, minimum sample split = 2, maximum depth = 200, and criterion = 'gini'. T hyperparameters for the RF model were established by making a previous explorati (tuning) with a grid search. For this, a range of values was defined, and a search algorith developed a random search between those values and found the best parameters. T default setting was used for the SVM model and adding the "squared hinge loss," which common for binary classifications [? ].

Results
Concerning the parameters measured with the sensor, there were no statistica significant differences between the mean values of η and ∆ f for the case of SF contained tubes with EDTA. However, in this case, a statistically significant difference was observ for ∆Γ. When comparing both samples of SF collected in tubes with lithium heparin, th were significant differences in the mean values of ∆ f and ∆Γ, but not for η (Tables 3, When looking at the differences between the data provided by the hospital, WBC is sho to be the most consistent data, with significant differences in both cases. All algorithms were developed using the Keras and sci-kit learn libraries in Python. We used a linear kernel for SVM, c = 1, loss = "squared hinge." For RF, we used 2171 trees, minimum sample split = 2, maximum depth = 200, and criterion = "gini." The hyperparameters for the RF model were established by a previous exploration (tuning) with a grid search. For this, a range of values was defined, and a search algorithm performed a random search of those values and found the best one. The default setting was used for the SVM model while adding the "squared hinge loss," which is common for binary classifications [42]. For the SVM and RF cases, 85% of the dataset was used for training, and 15% as test set.

Results
Concerning the parameters measured with the sensor, there were no statistically significant differences between the mean values of η and ∆ f for the case of SF contained in tubes with EDTA. However, in this case, a statistically significant difference was observed for ∆Γ. When comparing both samples of SF collected in tubes with lithium heparin, there were significant differences in the mean values of ∆ f and ∆Γ, but not for η (Tables 3 and 4). When looking at the differences between the data provided by the hospital, WBC is shown to have the most consistent data-significant differences in both cases. The predictive ability of each parameter is shown in Figure 5 (ROC curve) and Tables 5 and 6, which illustrates the area value under the ROC curve (AUC), confidence interval (CI), and standard error (SE). Shown for reference are the WBC, serum procalcitonin (PCT), and SF PCT parameters obtained in a different study [10]. In Figure 5, we can see that the viscosity calculation obtained does not discriminate the infectious SF well. On the other hand, ∆ f and ∆Γ had better results on the samples contained in tubes with lithium heparin, although they did not become a test that stands out. In Figure 5, we can see that the viscosity calculation obtained does not discriminate the infectious SF well. On the other hand, ∆ f and ∆Γ had better results on the samples contained in tubes with lithium heparin, although they did not become a test that stands out. Table 6. Area under the ROC curve of the parameters as predictors for infectious fluid (S with lithium heparin) In Figure 5, the viscosity calculation obtained does not discriminate well the SF. On the other hand, ∆ f and ∆Γ had a better result in the samples contained in t lithium heparin, although they do not become a test that stands out. The obtained parameters showed a slightly better performance in samples tubes with lithium heparin; nevertheless, they are far from being decisive for cla One study [? ] shows that procalcitonin (PCT) is used as a marker to discriminate SF. The study shows that the WBC value is the most accurate at the time of disti infectious SF (AUC = 1). When evaluating the value of PCT in serum and PCT i show that PCT in serum was better (AUC = 0.82) than PCT in SF (AUC = 0.65) value is comparable with the ∆ f (AUC = 0.61) and ∆Γ (AUC = 0.65) obtained in (tubes with lithium heparin).
When observing the results, it is noticeable that the SF samples contained with lithium heparin show higher ∆ f , ∆Γ, and η values. This may be due to a the sample's viscosity generated by the type of anticoagulant in the tube. Stu that lithium heparin can lead to accumulations of white blood cells, which ma this phenomenon[? ? ].
Based on the low performance of each parameter individually to differ precisely, there is an interest in testing AI algorithms to see if they can help bett the samples. When classifying by ANN, six scenarios were analysed for each cont of the SF samples. Table 7 shows the accuracy values obtained in each case. In t the confusion matrix for each scenario is distributed as follows: TP: The real cla is inflammatory SF, and the prediction is made correctly. TN: The real classi infectious SF, and the prediction is made correctly. FP: The real classification is The obtained parameters showed slightly better performance in samples stored in tubes with lithium heparin; nevertheless, they are far from being decisive for classification. One study [10] showed that procalcitonin (PCT) is used as a marker to discriminate infectious SF. The study showed that the WBC value is the most accurate at the time of distinguishing infectious SF (AUC = 1). When evaluating the value of PCT in serum and PCT in SF, they showed that PCT in serum was better (AUC = 0.82) than PCT in SF (AUC = 0.65). This last value is comparable with the ∆ f (AUC = 0.61) and ∆Γ (AUC = 0.65) obtained in this work (tubes with lithium heparin).
When observing the results, it is noticeable that the SF samples contained in tubes with lithium heparin showed higher ∆ f , ∆Γ, and η values. This may have been due to a change in the sample's viscosity generated by the type of anticoagulant in the tube. Studies show that lithium heparin can lead to accumulations of white blood cells, which may explain this phenomenon [43,44]. The obtained parameters showed slightly better performance in samples stored in tubes with lithium heparin; nevertheless, they are far from being decisive for classification. One study [10] showed that procalcitonin (PCT) is used as a marker to discriminate infectious SF. The study showed that the WBC value is the most accurate at the time of distinguishing infectious SF (AUC = 1). When evaluating the value of PCT in serum and PCT in SF, they showed that PCT in serum was better (AUC = 0.82) than PCT in SF (AUC = 0.65). This last value is comparable with the ∆ f (AUC = 0.61) and ∆Γ (AUC = 0.65) obtained in this work (tubes with lithium heparin).
When observing the results, it is noticeable that the SF samples contained in tubes with lithium heparin showed higher ∆ f , ∆Γ, and η values. This may have been due to a change in the sample's viscosity generated by the type of anticoagulant in the tube. Studies show that lithium heparin can lead to accumulations of white blood cells, which may explain this phenomenon [43,44].
Based on the low performance of each parameter individually in differentiating SF precisely, there was interest in testing AI algorithms to see if they can help better classify the samples. When classifying by ANN, six scenarios were analyzed for each container case of the SF samples. Table 7 shows the accuracy values obtained in each case. In this article, the confusion matrix for each scenario is distributed as follows: TP: the real classification was inflammatory SF, and the prediction was made correctly. TN: the real classification was infectious SF, and the prediction was made correctly. FP: the real classification was infectious SF, and the prediction was made incorrectly. FN: the real classification was inflammatory SF, and the prediction was made incorrectly. This can be best seen in Figure 6. 30, 2022 submitted to Journal Not Specified 9 of 12 SF, and the prediction is made incorrectly. FN: The real classification is inflammatory SF, 224 and the prediction is made incorrectly. This can be best seen in Figure 6. When viewing the accuracy values obtained in Table 7, it is clear that samples con-226 tained in lithium heparin tubes perform better in classifying both cases. Considering the 227 imbalanced data as input elements to the ANN, increasing the number of epochs also 228 increases the accuracy. By increasing the number of hidden layers, the accuracy converges 229  When viewing the accuracy values obtained in Table 7, it is clear that samples contained in lithium heparin tubes performed better in the classification for both cases. Considering the imbalanced data as input elements to the ANN, increasing the number of epochs also increased the accuracy. By increasing the number of hidden layers, the accuracy converged faster to values close to 100%. The worst accuracy using ANN was for samples on EDTA tubes, using 1 and 2 HL and 100 epochs with a value of 85%; this improved to reaching 91% with 2 HL and 300 epochs. For the samples in lithium heparin tubes, all accuracy values were between 97% and 99% within both input dataset. Data balancing improved accuracy slightly for samples contained in EDTA tubes; for samples stored in lithium heparin tubes, there was no significant improvement. When using random forest models, the high accuracy obtained for the unbalanced data was remarkable, being the best for the case of SF in EDTA tubes. Again, data balancing slightly improved the accuracy. The SVM models were found to have low accuracy, having the lowest accuracy of all the models compared.    Tables 8 and 9 bring together the confusion matrices for each scenario. Note that the TP and TN parameters are shaded and follow the distribution in Figure 6. As can be seen, the FN and FP parameters for the cases with higher accuracy tended to 0.

Conclusions
This work showed that the technique used to characterize the viscous properties of SF using a QCR-based sensor could help classify and differentiate infectious SF from other nosological entities. The results are encouraging; however, a more extensive study is needed. We have shown an ANN that aids in the classification of inflammatory and infectious SF using data associated with the viscous properties of SF obtained using a QCR sensor. The extraordinary ability of AI technologies to classify data in a way that is superior to conventional techniques was demonstrated. In the comparison carried out in this work, the improvements through the use of classification models such as random forests and neural networks were noticeable. When comparing both classifications of SF using ∆ f , ∆Γ, and η individually, there were some statistically significant differences. Still, they did not perform well on their own in classification. However, high accuracy was obtained by training an ANN to differentiate between two types of SF. We achieved higher precision values for samples stored in tubes with lithium heparin. With the results obtained, developing a sensor using QCR for SF classification is promising. However, it is necessary to continue increasing the amount of information obtained with the sensor by measuring more samples and extending its application to other types of biological fluids. The proposed technique presents a novel method for the classification of human fluids. The advantages are: (i) the use of a low sample volume (50 µL), (ii) the low cost of the device, and (iii) portability. This makes it accessible to any laboratory and should promote interest in further development. As future work, the dataset could be further augmented, and a comparison between different classification models can be performed. Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.