Optimizing urine protein detection accuracy using the K-nearest neighbors algorithm and advanced image segmentation techniques

This research introduces a method for protein detection in urine that leverages a digital camera sensor to extract color information. The study incorporates a urine test strip to categorize protein concentration in urine. The process involves sequential stages of sample preparation, image processing based on urine color, evaluation using the KNN model, and subsequent classification of protein levels. The study begins with sample preparation utilizing a standard protein solution. Protein analysis within the solution is executed through a urine test strip, used as a medium for color segmentation. Color segmentation includes three main steps: (1) noise removal to clean the image, (2) RGB feature extraction to analyze specific colors, and (3) classification using the KNN algorithm to predict protein concentration. Image data captured by the camera sensor are then classified based on the resultant RGB values. RGB color segmentation plays a crucial role in enhancing the accuracy of protein-level detection in urine samples by enabling precise differentiation of color variations on urine test strips. The RGB model quantifies RGB intensities, allowing for objective and consistent analysis, unlike human visual assessment, which is prone to errors. By extracting RGB features, variations in protein concentration can be mapped to specific color intensities, improving classification accuracy. The subsequent phase involves an evaluation utilizing the KNN classification method. The computational aspect of data processing is performed on a computer, enabling efficient analysis and classification. This innovative approach streamlines the detection process and provides a robust foundation for accurate protein-level classification in urine samples.

In data testing, the choice of method for measuring proteinuria is based on clinical needs and the availability of resources. Methods such as turbidimetry, nephelometry, and radioimmunoassay have their advantages. For example, turbidimetry offers a rapid process with high accuracy, while nephelometry is more sensitive to detecting low protein concentrations. Radioimmunoassay, although expensive and requiring a specialized laboratory, remains the gold standard for quantitative analysis. Protein samples were obtained from six standard protein solution samples. Making standard protein solutions is conducted by entering protein solutions into six categories using a solution of 0–11.60 g mixed with 20 mL of mineral water. After measuring the dose, wait 60 s for each sample, then place it on the slide strip. A comparison of standard protein samples are given in Table 1. In data collection, the stripped output (−) is equivalent to a protein content of 0 g/L, the stripped output (+−) is equivalent to a protein content of 0.15 g/L, the stripped output (+) is equivalent to protein content 0.3 g/L, strip output (++) equivalent to protein content 1 g/L, strip output (+++) equivalent to protein content 3 g/L, and strip output (++++) equivalent to protein content protein 20 g/L. This strip output can be interpreted as the symbol (−), which indicates that no protein content was detected, while the symbols (+), (++), (+++), and (++++) indicate that protein was detected. Difficulties such as inconsistent dipping times or uneven solution distribution are overcome by standardizing the procedure. The protein solution is mixed homogeneously, and the dipping duration is set for 5 s. Preparing standard protein solutions in categories (−) to (++++) is crucial for ensuring reliable data collection and classification. Each category represents a specific protein concentration level used as a reference for detection. This standardization allows the system to compare RGB values obtained from urine test strips with predefined references, thereby enhancing the classification reliability and the validity of the KNN model.

Protein detection in this research uses a digital camera sensor type ELP camera as the primary sensor. The ELP digital camera is used in this system to ensure consistency in image data capture. This camera can reduce the effects of shadows and external lighting, thus providing stable and accurate RGB data results. Combining the ELP digital camera sensor with real-time computer processing enables stable and accurate image acquisition for protein-level analysis. The ELP camera minimizes shadowing effects and external lighting variations that could affect urine strip color interpretation. Additionally, real-time processing ensures fast and precise detection, making this approach more reliable than visual evaluation or manual techniques. Image data were also taken using a urine test strip tool. The color segmentation method is used to analyze protein levels using urine test strips. The camera image’s initial reading is obtained from RGB colors. Real-time data processing is carried out with a computer program, as shown in Figure 1. The first procedure is to take an example of the artificial protein solution. Then, the urine test strip is dipped in a protein solution that has been collected in a measuring cup. The immersion process must be carried out quickly, and as soon as possible, the urine test strip is taken and inserted into the prototype using a strip slider. The urine strip takes approximately 30 s to display accurate color results when dipped in the protein solution. Next, the color segmentation process begins to process the urine strip data, with the stage of removing image noise. The final stage is data evaluation using the KNN model, as shown in Figure 2.

Figure 1:

Protein detection computer program. KNN, K-nearest neighbors.

Based on Figure 2, the evaluation of protein detection tools using the KKN model involves taking digital image extract data to determine the resulting RGB features. After that, the training and test data will be processed automatically to determine the k value in the KNN model. The KNN algorithm is one of the most basic and straightforward classification methods. It should be one of the first choices for classification studies when more prior knowledge is needed, regarding data distribution. The KNN classifier was developed to perform discriminant analysis when parametric probability density estimates are unknown or difficult to determine. KNN methods search the entire set of training data samples to classify input testing samples. The final stage of evaluating the KKN algorithm is to produce accuracy, precision, recall, and F1-score values.

Figure 2:

Evaluation of the KNN model. KNN, K-nearest neighbors.

The KNN algorithm is used to classify data based on the shortest distance to the object data. The KNN algorithm calculates each point on each class’s test and training data. In principle, data collection is carried out from the closest distance to the farthest distance, and the system will choose the most relevant distance between the test data and the k number of training data. The KKN algorithm has three concepts for calculating the distance between data: Manhattan distance, Euclidean distance, and Minkowski distance. The Euclidean distance metric was selected due to its computational simplicity. However, additional tests using Manhattan and Minkowski distances were conducted to compare classification performance. Although Euclidean distance performed optimally in our dataset, Minkowski distance showed slight improvements in handling outliers. Mathematically, it can be seen as Eq. (1).

The KNN model is evaluated using several metrics such as accuracy, precision, recall, and F1 score. Accuracy measures the extent to which the KNN model can classify data correctly compared to the total amount of data. This is the most commonly used metric for classification evaluation and correctly measures the percentage of classified data. Precision measures the extent to which positive predictions a KNN model makes are correct. It calculates the correctly classified positive data ratio compared to total optimistic predictions. The mathematical equation for evaluating the KKN model is presented in Eq. (2).

The second evaluation, Recall, is the ratio of true positive (TP) cases to total cases, the number of TP and false negative (FN) points, which is referred to as sensitivity, as given in Eq. (2). Next, the F1 score is given by Eq. (3).

Recall measures the extent to which the KNN model can detect TP data. It calculates the ratio of correctly classified positive data to the total TP data. F1-score is the harmonic average of precision and recall. This provides an overall picture of the balance between accuracy and recall in cases where the positive and negative classes are unbalanced.

The formula explanation above is the basic concept in confusion matrix classification, where TP is the number of positive cases correctly classified as positive by the model. True negative (TN) is the number of negative instances correctly classified as unfavorable by the model. False positive (FP) is a negative case incorrectly classified as positive by the model. FN is a positive case incorrectly classified as unfavorable by the model.

Figure 3 shows this study’s innovative protein detection prototype and features a meticulously designed 3D-printed structure. The prototype incorporates critical components such as an ELP-type camera serving as the primary sensor for capturing image data from samples, a urine test strip with 10-variable specifications as the reaction medium, a designated container housing protein powder for a standard protein solution, a 7-inch LCD touchscreen functioning as the system control device, and Jetson Nano acts as an AI developer platform that enables real-time image analysis with low computing power. Jetson Nano was chosen due to its ability to process images in real-time with low power consumption. This allows seamless integration with ELP cameras and LCDs. This multifunctional prototype aims to enhance efficiency and accuracy in protein detection. The urine test strip analysis process is designed to read the color results within 20–120 s after immersion. After 120 s, environmental factors such as oxidation and evaporation can affect the color of the strip, resulting in inaccurate results. The study analyzes how color changes over the 20–120 s period affect accuracy. Results indicate that readings taken beyond 90 s start to deviate due to oxidation effects. Therefore, an optimal reading window of 30–90 s is recommended for maximum accuracy.

Figure 3:

(A) Image of the prototype seen from the outside, (B) prototype components seen from the inside, and (C) shape of the prototype that is ready to be used.

Therefore, the analysis time must be adhered to maintain the reliability of the results. Beyond this window, the strip’s analysis may yield suboptimal or invalid results. The strip provides qualitative outcomes, distinguishing between positive and negative samples. Semiquantitative values are indicated by symbols such as (+), (++), (+++), and (++++), while quantitative values correspond to specific color levels discussed in the subsequent section. Notably, the ELP USB webcam type employed in this prototype stands as a standard industrial camera, ensuring reliable and consistent performance in capturing essential image data for further processing. The fusion of 3D printing technology with advanced sensor components demonstrates a promising leap forward in protein detection methodologies.

Based on Figure 3, at the top, there is a mount for placing an LED that functions as lighting inside the casing. LED illumination plays a crucial role in ensuring consistent color detection. We conducted experiments with varying LED intensities and angles to assess their impact on RGB color stability. The results indicate that a uniform light source at a fixed angle of 45° minimizes shadow effects and enhances classification accuracy. At the bottom, there is a rail to place the urine strip slider. Mounts on the right and left sides of the camera are connected to the camera’s sides. At the bottom, there is a hole designed to fit the camera lens size. This part is where the urine test strip is inserted into the camera casing. This slider is designed to match the rail inside the casing. This is necessary to keep the strip’s position stable, ensuring a consistent reading from the camera. The mechanics are made using white PLA material. The choice of white PLA material for the prototype casing is based on its ability to minimize light reflection and maintain consistent urine strip readings. By reducing variations in lighting conditions, the white casing ensures stable color segmentation and enhances the reliability of RGB-based analysis. The choice of white color was made considering the light reflection inside the casing. Light reflection can be minimized by using a white material, ensuring a more stable color emission from the urine strip. The integration of RGB camera-based color segmentation in this prototype can monitor their health conditions without having to visit a health facility. Thus, the development of AI-based technology in the medical field is very much needed to increase patient involvement in the management of chronic diseases such as diabetes and hypertension, as it provides easy access to early diagnosis.

Data collection is the process of collecting data from digital camera sensors in reading urine strips. Urine strips taken were used as training data. This study used 99 urine protein image data to train and test the KNN model. This dataset consists of six categories based on the level of protein concentration including: negative (−) for 6 image data, plus-minus (+−) for 24 image data, positive 1 (+) for 10 image data, positive 2 (++) for 22 image data, positive 3 (+++) for 30 image data, and positive 4 (++++) for 7 image data. Determining the size of the dataset significantly affects the performance of the KNN algorithm, especially on medical image-based data. Adding a dataset not only improves accuracy but also strengthens the model’s generalization ability in dealing with new data, especially in clinical settings with more complex variations.

Training data are used to train machines in developing models. Meanwhile, test data are used to test the results of research carried out by machines and are used to compare model performance. The training and test data in this study are shown in Table 2. Next, the system will read the protein sample data image, entering the slider tray. Data images are extracted based on the distribution of RGB features, as shown in Figure 4.

Figure 4:

Distribution of 30 test data.

Selecting the right model is very important in AI-based diagnosis. In this study, the KNN model was chosen because of its ability to handle non-linear data, flexibility in parameter settings, and classification efficiency. The results show that KNN with K = 3 provides the highest accuracy. KNN model simplifies the classification and regression process in medical diagnostics through an intuitive distance-based algorithm. The flexibility and efficiency of KNN make it an ideal tool for developing digital image-based diagnostic applications, as demonstrated in this study. The KNN algorithm is one of the machine learning algorithms that has been used to analyze metabolites in urine with high-resolution mass spectrometry. This method provides better predictive ability and time efficiency compared to conventional approaches. The basic idea behind KNN is that similar objects tend to be closer to each other in the feature space. The working system of the KNN model consists of training and prediction. During the training phase, the KNN algorithm does not learn the data but merely remembers the training data and labels. Predictive operations occur when new data are to be predicted. KNN will search for the K nearest training points (nearest neighbors) based on distance in the feature space. Then, the KNN model operates by majority voting from the set of data labels to predict the new data. Machine learning algorithms, such as KNN, have significant advantages over conventional diagnostic methods. These advantages include the ability to efficiently process non-linear data, speed of data classification, and high accuracy even with limited datasets.

Some essential points in the KNN algorithm include the K parameter, distance measurement, feature normalization, and data computation. The K parameter denotes the number of nearest neighbors taken in the prediction process. A K value that is too small might lead to overfitting, while an overly large K value might lead to underfitting. Distance measurement calculates the distance between points in the feature space. The choice of this metric can influence KNN results; in this study, the Euclidean distance measurement was used. Feature normalization is vital as KNN is sensitive to feature scaling; hence, normalization or standardization of features might be needed before training the model. Data computation means that KNN requires distance calculation for every pair of data, which can be computationally intensive with large datasets.

In this study, once introduced into the urine test strip, protein solution samples were tested within the created prototype system. The camera sensor then initialized the data. A reaction urine strip requires 30 s before the camera capturing process begins. Image data are then classified based on RGB. The resulting RGB values are evaluated using the KNN model, displaying predicted protein levels. The evaluation results, which include accuracy, precision, recall, and F1-score, are determined using the confusion matrix. The KNN algorithm in this study used the Euclidean distance method for classification and regression. The first step in the KNN model is to determine the K value, which is the variable number of nearest neighbors taken for the classification process. Setting the K value is a balancing act, as different K values can lead to overfitting or underfitting. Lower K values have higher variance, while larger K values can lead to biased classification or lower variance. In this study, the classification parameter, number of neighbors (K), was set as K = 3, K = 10, and K = 20. The performance of the KNN model based on these K values (3, 10, and 20) is shown in Table 3.

Based on Table 3, the KNN model classification system for protein detection can be used to predict data. Overall, it is used to classify protein data into the categories negative (−), trace (+−), positive 1 (+), positive 2 (++), positive 3 (+++), and positive 4 (++++). Furthermore, the KNN model classification can be realized in the form of a detailed confusion matrix for multiclassification, as shown in Figure 5 for the value of K = 3, Figure 6 for the value of K = 10, and Figure 7 for the value of K = 20. The blue color in the confusion matrix shows the amount of data in the resulting matrix column. The confusion matrix shows that the darker the blue color, the more test data there is in the matrix column, and the more correct data it shows. This confusion matrix makes it easier to read the comparative distribution of expected and predicted data. A confusion matrix is used to determine the suitability of truth and test data.

Figure 5:

Results of confusion matrix values at K = 3.

Based on Figure 5, the x-axis represents the prediction results from the KNN model, and the y-axis represents the classification results using the KNN model. The distribution consists of 30 data points across all classes, which include the negative (−) class, trace (+−) class, positive 1 (+) class, positive 2 (++) class, positive 3 (+++) class, and positive 4 (++++) class. The number of neighbors is set to 3 (K = 3), which means the average of the three nearest neighbors to the test data is calculated. The K value here indicates that the higher the K value, the more neighboring data points are considered in the calculation. Based on the study results, K = 3 showed excellent evaluation results with an accuracy rate of 96.7%, precision of 97.00%, recall of 96.7%, and an F1-score of 96.2%. In the test with K = 3, one error or misclassified data point was in the upper right.

The combination of the KNN algorithm with digital image processing improves the accuracy of protein detection in urine samples. This study shows that KNN, especially with a value of K = 3, provides the best results with an accuracy of 96.7%. The advantages of KNN are its ability to handle non-linear data and provide fast predictions based on the closest distance, making it an ideal tool for clinical applications. Subsequent tests were conducted for K = 10 and K = 20. The test results show that at K = 10, there were five misclassified data points: two data points were (−) but predicted as (++++) and two data points were (+++) but predicted as (++). At K = 20, there were even more misclassifications: two negative data points were predicted as (++++), three trace (+−) data points were predicted as (++++), and two (+++) data points were predicted as (++). The selection of the K value has a significant impact on the balance between overfitting and underfitting. A small K value (e.g., K = 3) provides higher sensitivity but risks overfitting, making the model highly responsive to noise. A larger K value (e.g., K = 20) smooths classification but may underfit by ignoring finer variations in protein concentration. This study found that K = 3 offered the best accuracy, indicating an optimal balance between sensitivity and generalization. This indicates that as the number of neighbors (K value) increases, the accuracy decreases. From the research results, K = 10 had an accuracy rate of 86.7%, precision of 75.8%, recall of 86.7%, and an F1-score of 80.7%, while K = 20 had lower results compared to K = 3 and K = 10 with an accuracy of 76.7%, precision of 60.9%, recall of 76.7%, and an F1-score of 67.3%.

The colors in Figures 5–7, as indicated by the confusion matrix, represent the amount of data in each column. The darker the blue, the more data are present in the corresponding matrix column. This is to make it easier for readers to compare the distribution of comparisons of expected data and predicted results. Comparing machine learning algorithms with other conventional methods provides an opportunity to improve the accuracy and efficiency of analysis, especially in the analysis of urine biomarkers in metabolic conditions. Machine learning provides a powerful tool for managing chronic diseases through urine biomarker analysis. For example, the detection of micro metabolites in urine using the KNN algorithm can help in the early diagnosis of diabetes and hypertension, thereby increasing the effectiveness of medical interventions.

Figure 6:

Results of confusion matrix values at K = 10.

Figure 7:

Results of confusion matrix values at K = 20.

Many researchers have developed proteinuria detection technology in various situations. The obstacles encountered in protein detection are also increasingly complex. Protein detection was initially developed using chemical/dye-based enzyme methods. Turbidimetry, nephelometry, and radioimmunoassay methods have unique advantages in measuring albuminuria. Turbidimetry is suitable for rapid clinical settings, nephelometry provides high sensitivity for low concentrations, while radioimmunoassay offers the most accurate quantitative results, although it requires a specialized laboratory. However, this tool requires professional medical personnel and can only be used in a laboratory setting. Researchers have turned to machine learning technology to process large amounts of data efficiently and accurately, resulting in better predictions and decisions. This research created a protein detection prototype equipped with an ELP camera-type digital color sensor. Next, the camera sensor will initialize the data on the urine strip, which has reacted with the urine sample. The initialization results in the form of image data are then classified based on RGB. The resulting RGB values will be evaluated using the KNN model algorithm, which displays the predicted results of protein levels. Evaluation results in the form of accuracy, precision, recall, and F1-score data will be carried out using the confusion matrix. The results of this research are related to those of previous research, as shown in Table 4.

Based on Table 4, the use of machine learning technology is essential to overcome the problem of determining color in urine. Unpredictable color changes in urine are due to contamination with other substances. Apart from that, analysis using the eyes is less effective because the effects of shadows or ambient light influence it. The difference in this research lies in the use of color retrieval technology, which uses an ELP sensor installed in a series in the protein detection prototype. The ELP camera sensor in the prototype can reduce or minimize shadow effects when capturing colors. Color segmentation for protein detection has been shown to improve diagnostic accuracy. This technology uses a smart camera to capture the color of chemical reactions on a urine strip and analyze its RGB values, using a machine learning algorithm, such as CNN. A study by Thakur et al. [41] showed 88% accuracy in detecting albumin using this method, making it a practical and affordable solution for healthcare settings.