A surface-based calibration approach to enable dynamic and accurate quantification of colorimetric assay systems

Colorimetry is widely used in assay systems for its low-cost, ease-of-use, rapidity, moderate storage requirements and intuitively visible effects. However, the application is limited due to its relatively low sensitivity. Conventional colorimetric calibration methods often use a fixed incubation time that can limit the detection range, system robustness and sensitivity. In this paper, we used color saturation to measure the accumulation of product (correlation coefficient R 1⁄4 0.9872), and we created a novel “calibration mesh” method based on an expanded sigmoid function to enhance sensitivity. The novel calibration mesh method can be adapted for a wide variety of assay systems to improve robustness and detection range, and provide a dynamic and faster output.


Introduction
Colorimetric analyses have been widely used in various types of assays, including tests for pregnancy, 1 bacteria, 2 toxicity, 3 pH, 4 soil, 5,6 industry waste water 7 and many common biomarkers in blood 8 and sweat. 9,10 The development of low-cost point-of-care devices, such as those that use simple paper pads, 17 could help address health disparity issues related to inadequate access to care. 16 Colorimetry offers a low-cost, accessible measurement technique that does not require complex imaging modalities or burdensome data storage. However, compared to uorescence or other modern detection methods, the biggest disadvantage of colorimetry is low sensitivity and high noise. Most colorimetric data analyses are based on light absorption, 11 gray-scale intensity, 12 red, green, blue (RGB) color space measurements, 8,9 or analyses based on the hue channel of the hue, saturation, value (HSV) color space. 13 Analyses based on the saturation channel of HSV color space can outperform other channels when a colored product accumulates in an assay 14 and is less dependent on ambient light conditions. 15 A linear or non-linear t of measured values (e.g., intensity, absorption, impedance) to known concentrations is oen used to establish a conventional calibration curve in an assay system. Readings based on this curve oen require a xed incubation time, which limits both the detection time and the detection range. Low concentration samples may require more incubation time to be distinguished from blank samples; high concentration samples may exceed the upper limit of detection of the reader aer a xed incubation time. Also, using a xed incubation time provides only a single readout per sample. Technical replicates are oen utilized to overcome single measurement inaccuracies, but this requires additional sample and reagents.
In this study, we developed a colorimetric paper pad sensing unit and analysis technique to quantify the concentration of the enzyme invertase. Our technique utilizes an innovative "calibration mesh" method that uses a temporal analysis of colorimetric saturation to replace the conventional xed-time calibration method. To demonstrate the robustness of this method, we compared the performance of the "calibration mesh" to the conventional calibration approach.

(a) Reagents and solutions
Sucrose was purchased from VWR. Glucose oxidase (G2133-50KU, 242 776 U g À1 ), peroxidase from horseradish (P8125-25KU, 50-150 U mg À1 ), and invertase (I4504-250MG, $300 U mg À1 ) were purchased from Sigma. D-Trehalose, 99% anhydrous was from Acros Organics (NJ). Sodium acetate was purchased from Fisher Scientic. Phosphate buffered saline (28 372) was purchased from Thermo Scientic. Sodium iodide (409286-1G) was purchased from Aldrich. Acetic acid was purchased from Sigma-Aldrich. Cellulose ber sample pad strips (CFSP203000) were purchased from Millipore. Deionized water (18.0 MU) was used in all procedures. Details of buffer preparation are available in ESI. † (b) Sensing paper pad procedure Paper pad fabrication. The cellulose ber paper was cut into round paper pads using a quarter size hole puncher (Walmart), and the edge was trimmed to remove burrs. The pad was then sealed in a plastic storage bag for future use.
Reaction cocktail solution. This solution was made immediately before use and was modied from a glucose sensing method. 9 Briey, the nal solution consisted of 1.2 mg (291 units) glucose oxidase, 0.12 mg (6-18 units) HRP, 105.2 mg trehalose, 92 mg sodium iodide, 171 mg sucrose and 1 mL 0.05 M cold sodium acetate buffer, pH ¼ 5.55. 18 The vial was covered with foil to protect it from light. More details are available in the ESI. † Substrate treated paper pad fabrication. We used a 100 mL pipette tip holder as the paper pad reaction holding plane. Twenty-four paper pads were placed on top of the holes of the holder in 8 rows and 3 columns. Each paper pad covered a pipette tip hole, promoting faster, uniform drying of the pads. 100 mL of the cocktail solution was pipetted onto each paper pad and allowed to dry at room temperature overnight.
Colorimetric mechanism. Invertase was used to hydrolyze the conversion of sucrose to glucose (eqn (1)). 19 Next, glucose was catalyzed by glucose oxidase to produce hydrogen peroxide (eqn (2)). 20 The subsequent peroxidase catalyzed reaction between iodide (colorless) and hydrogen peroxide produced iodine (brown color) for colorimetric analysis (eqn (3)). 21 A standard stock solution was prepared using 2 mg invertase dissolved into 1 mL PBS. This stock was gradually diluted with PBS into standard concentrations of 10, 5, 2.5, 1, 0.5, 0.25, 0.1 and blank 0 mg mL À1 (PBS only). Finally, 360 mL of each standard solution was added into a column of a 96-well plate.

(c) Experiment and camera setup
The whole experiment was carried out under a stable and uniform white illumination area. All steps were as follows: Step 1: place 24 substrate-treated paper pads on a cleaned pipette tip holder as described above (8 rows and 3 columns). Fix the camera (details in ESI †) 20 cm above the paper pads.
Step 2: aer focusing the image, turn off all camera automatic adjustment and balance functions, and acquire images at 5, 20, 30, 40, and 72 min.
Step 3: use an 8-channel pipette to add 100 mL invertase solution to the center of each paper pad in the rst column avoiding touching or moving the paper pads.
Step 4: repeat step 3 for the second and third columns. The time interval between each addition is 20 s.

(d) Image processing
Aer capturing the images of paper pads, we converted the captured image to a gray scale image, an R, G, B color space image, and an H, S, V color space image (Fig. 1).
To calculate the average saturation value of HSV color space for each paper pad, the pad area was selected using a circular mask, and the saturation values of regions outside of the paper pads were set to 0. The average saturation value of all pixels in a paper pad was calculated as follows: where s k is the saturation value of each pixel, N is the number of pixels in a paper pad, s is the average of saturation value of all pixels in a paper pad. The background was selected as the average saturation value of each paper pad at the 5 min time point. Unless otherwise specied, all saturation data are given aer subtracting the background.

(e) Correlation coefficients calculation
Similar to the saturation channel processing described above, we calculated both raw data and data with the background subtracted for the gray image, and all channels of R, G, B, H, and V. We then calculated the correlation coefficient between the target concentration and magnitude of each of these channels (Table 1).

(f) Calibration method
The data obtained from the standard concentrations of 0.1, 0.5, 1, 2.5, and 10 mg mL À1 were used as training dataset to establish conventional calibration curve or calibration mesh. The dataset obtained from the standard concentrations of 0.25 and 5 mg mL À1 were used as testing data to evaluate the performance of the algorithm and different methods.
1. Conventional calibration curve. We established a conventional calibration curve by linearly tting the training dataset obtained at 30 min. Three technical replicates were acquired for each concentration (0.1, 0.5, 1, 2.5, and 10 mg mL À1 ) to yield 15 data points in total.
2. Calibration mesh. We established the calibration mesh by using training dataset at 5 min, 20 min, 30 min, 40 min, and 72 min with three technical replicates for each concentration (0.1, 0.5, 1, 2.5, and 10 mg mL À1 ). This yielded 75 total data points. The time ðt 75Â1 Þ and the concentration ðc 75Â1 Þ data were set as inputs. The saturation data were set as outputsðỹ 75Â1 Þ. Two different models were proposed for establishing the calibration mesh: 2.1. Linear regression model. In this model, we generated the prediction function as a polynomial function: The t and c here were normalized by dividing each time and concentration by the maximum values of time and concentration, respectively.
First, we dened an input matrix: Then, the prediction function can be rewritten as: where q ! ¼ ½q 0 ; q 1 ; .; q 9 T .
Next, we dened the cost function as: where m ¼ 75 The goal was to solve the coefficient q ! to make the predicted output saturation value hð q ! ;xÞ ! as close as possible to the training saturation valuesỹ 75Â1 by nding the minimum of Jð q ! Þ:Since the dataset was not large, we used a normal equation approach to solve for theta: Note: if X T X is not inversable, we apply a pseudo inverse. A regularization was applied to prevent over tting. Then, where l was choosen as 0.01. D is a 10 by 10 matrix, which is similar to the identity matrix except that the rst element is 0. ; z ¼ kð1Þ kð2Þ þ e Àkð3ÞÂðzðt;cÞÀkð4ÞÞ þ kð7Þ; wherek is a 7 Â 1 vector of [k(1), k(2), ., k (7)] The sðk; zÞ function retains a sigmoid shape but is shied and stretched by the coefficientsðkÞ. To extend the sigmoidal function (s) to a surface, we dene z(t,c) to capture the relationship between time and concentration.
Next, we dened the cost function: The training goal was to nd an optimal solution ofk to minimize the cost of JðkÞ. Since the training dataset was discrete, we used the fminsearch function within Matlab.
For both the linear regression model and the sigmoid model, once the optimal solution for q ! ork was obtained, we could build the calibration mesh within the given range of time (0-74 min) and concentration (0-10 mg mL À1 ).

(g) Performance evaluation
Evaluation of calibration mesh with training dataset. We rst plotted the average training dataset on the calibration mesh (Fig. 3), and qualitatively evaluated the performance of the mesh, where the average training saturation data was calculated by taking the average of saturation data of 3 different paper pads at a single capture time. Quantitively, by using the optimal solution ð q ! andkÞ within the respective cost function, we compared the total errors or difference between the predicted values and the output values (Table 2). Compare conventional calibration curve to calibration mesh. To evaluate the performance of traditional conventional calibration curve methods, we used the saturation data for the 0.25 and 5 mg mL À1 samples collected at 30 min in the linear t equation to calculate the predicted concentration.
For the calibration mesh, we used the saturation data obtained from 0.25 and 5 mg mL À1 concentrations measured at   Fig. 4 for mapping the predicted concentration. Absolute relative difference (ARD) for each of these methods was then calculated as:

Results and discussion
(a) Saturation channel performance Rather than working in the orthogonal RGB space, which depends on linear combinations of red, green, and blue values to describe a color, saturation data obtained in the HSV space does not depend on the color produced in a reaction (Fig. S7 †). The accumulation of any color product yields a saturation value that can be directly related to concentration. By observing the histogram and normal distribution of pixel saturation (Fig. S1 to S3 †) and the comparison between the raw image and the data channels shown in Fig. 1, we found that the average saturation value of all pixels from a paper pad represents the color change of the paper pad well.
The correlation coefficient values in Table 1 show that the saturation channel had the best correlation with concentration among all channels. Also, the saturation data with background subtracted gave a higher correlation coefficient than the raw data. Therefore, we selected the average saturation value of the raw image minus background as our training or testing saturation data in this study.

(b) Calibration curve and calibration mesh development
A conventional calibration curve (Fig. 5) was established using a linear t based on the 30 minutes training dataset. These data yielded the following t: Next, we developed the rst calibration mesh model using a linear regression technique. Aer training, the optimal solution of q ! was [À0.0017, -0.0986, À0.0442, 0.3161, 0.8425, 0.0059, À0.1326, À0.1963, À0.3172, 0.0067]. Aer plotting the mesh (Fig. 3A), we observed that the average training dataset ("yellow dots") did not t the calibration mesh well.
From the contour plot of average saturation raw data vs. time and concentration (Fig. 2), we observed that when both time and concentration are small (le bottom corner) or large (top right corner), the distance between contour lines are relatively large, while the density of contour lines in the middle is higher. This indicates that the slope in the middle is higher than the lower le or the upper right regions of the plot. Based on the slopes observed in the contour plot data, we hypothesized that a sigmoidal function may provide a more appropriate t for our calibration mesh.
Aer training the sigmoid model, the optimal solution ofk was [0.5294 1.4278 1.2644, 5.0479, À0.6219, 0.7222, À0.0043]. The sigmoid model calibration mesh t the average training saturation data well (Fig. 3B, "yellow dots"). As shown in Table  2, the error of the sigmoid model (4.7337 Â 10 À5 ) was smaller than the error obtained using the linear regression model  Fig. 4 Flow chart for using calibration mesh. t i : the i th input time for capture of paper pads. S i : the i th saturation value (raw data). S BG : the background saturation value. SR i : data. CalMesh: calibration mesh. ARD: absolute relative difference. ***If there were N duplication inputs, we will calculate the average value of c i as: (1.7067 Â 10 À4 ). Therefore, we concluded that the sigmoid model outperformed the linear regression model in this study. The linear regression model, however, has some advantages that should be considered. First, as a common method, the linear regression model could be used for general situations, especially when the detection range of interest demonstrates a linear relationship between saturation and both concentration and time. Second, the linear regression model does not require optimization to account for variations in mesh shape. Third, training by solving the normal equation is more cost-effective than "fminsearch". In this study, the training time (averaged over 10 runs) for the linear regression model was 0.0002 s, while the average training time of the sigmoid model was 0.0323 s. Therefore, the linear regression model may be more appropriate for platforms in which computational effort is a major concern.
(c) Calibration mesh utilization For the conventional calibration method, we can map a saturation value at a xed time (e.g., 30 min) on the calibration curve to determine the associated concentration. Similarly, for the "calibration mesh", we can map a saturation value (s) at a time (t) on the "calibration mesh" and nd the associated concentration (c). The added benet of the calibration mesh method is the ability to capture images at multiple time points and increase the accuracy of the concentration measurement (Fig. 4).
(d) Performance of the conventional calibration curve vs. calibration mesh models As shown in Table 3, although the conventional calibration curve method did well in identifying concentration for 5 mg mL À1 (ARD ¼ 0.48%), it was unable to accurately identify the 0.25 mg mL À1 condition (ARD ¼ 410.22%). The calibration mesh using the sigmoid model accurately calculated both concentrations (ARDs were 0.67% and 0%, respectively).

(e) Merits of calibration mesh
In this study, the conventional calibration curve was constructed based on a xed time (30 min). However, from the calibration mesh (Fig. 3), we can see that a higher concentration sample has a faster chemical reaction and a rapid increase in saturation value, which requires a shorter time to be distinguished from the blank. On the other hand, as shown in the results presented above, the conventional calibration curve may be inaccurate for low concentration prediction. The calibration mesh can overcome this limitation by extending the reaction time so that the lower concentration samples would eventually exhibit and increase in saturation to improve the signal-to-noise ratio. Using this method, an investigator does not need to determine the incubation time a priori. Further, if multiple samples with large concentration differences are being assayed simultaneously (or  perhaps several biomarkers with drastically different concentrations in a multiplex assay), the calibration mesh technique can provide reliable quantication without the need to perform separate incubation times for each sample or biomarker. The designed algorithm for utilizing the calibration mesh can give a robust and accurate output by setting conversion criteria for the predicted concentration over time. The calibration mesh can expand the detection range and provide an earlier output.

Conclusions
We have demonstrated that color saturation has a higher correlation to accumulated product in a paper pad assay system. We have also shown that the calibration mesh method has an improved dynamic range and accuracy compared to a conventional calibration curve. The utilize of calibration mesh is simple and can be adapted for a wide variety of assay systems.
There are several things we can do to improve the performance of the calibration mesh method in future iterations. First, increasing the standard concentrations and automating image capture with higher frequency could provide better model inputs and achieve higher resolution calibration mesh in both concentration and temporal axis. Second, instead of using saturation only, we could explore using a combination of gray, R, G, B, hue, saturation, and value data to obtain a combined output, which may further enhance the correlation between the imaged data and the concentration. Third, since temperature changes can affect chemical reactions, we can also integrate temperature as a variable in the training model. Fourth, for the calibration mesh utilization algorithm, we can also introduce a factor that changes the weight of a single measurement based on the ARD at particular time points in the assay. Fih, we can utilize the slope of calibration mesh to eliminate potential background error. Utilization and iterative improvements of the calibration mesh technique could yield a robust algorithm that will enhance the accuracy and sensitivity of colorimetric assays.

Conflicts of interest
There are no conicts to declare.