Skip to main content
NCBI home page
ACS Omega logoLink to ACS Omega
. 2026 Sep 29;11(40):61020–61031. doi: 10.1021/acsomega.6c05231

Concentration Determination of Aspirin Deposit Patterns Using Quantitative Image Analysis and Artificial Intelligence

Sanam Pudasaini †, Danielle O Eniola †, Oliver Steinbock ‡, Beni B Dangi †,*
PMCID: PMC13670302  PMID: 42889880

Abstract

The rapid detection of substandard or falsified pharmaceuticals remains a major challenge, particularly in regions with limited analytical infrastructure. In this study, a low-cost imaging-based method combined with machine learning was investigated for quantitative estimation of aspirin (acetylsalicylic acid) concentration through analysis of dried sessile droplet patterns. Aspirin solutions at concentrations ranging from 6% to 100% (saturation) were deposited on glass slides and allowed to evaporate under ambient conditions. The resulting crystallization patterns were imaged using a high-resolution camera, and 46 geometric and textural features were extracted from each image. Principal component analysis (PCA) revealed concentration-dependent clustering of deposit morphologies, indicating that structural differences in the dried patterns correlate with aspirin concentration. Regression models including Random Forest and XGBoost were trained using the extracted image features, while deep learning models (ResNet-50 and EfficientNet-B3) were applied for direct image-based regression. Amongst the models, EfficientNet-B3 achieved the highest predictive accuracy with an R2 = 0.99, while feature-based models also demonstrated strong predictive performance with R2 = 0.98. This approach was further evaluated using a commercial aspirin tablet sample, and results were compared with concentrations obtained through more conventional UV–visible spectroscopy using the standard addition method. This work demonstrates that droplet deposition imaging combined with artificial intelligence provides a promising, rapid, and non-destructive strategy for pharmaceutical concentration analysis and quality assessment.

graphic file with name ao6c05231_0008.webp

graphic file with name ao6c05231_0007.webp

1. Introduction

Substandard and falsified pharmaceutical products represent an ongoing global public health concern, particularly in low- and middle-income countries where regulatory capacity, analytical infrastructure, and routine quality surveillance are limited. The circulation of poor-quality medicines can lead to subpar therapeutic dosing, treatment failure, increased morbidity, and diminished public confidence in healthcare systems. These challenges highlight the need for analytical approaches that are both reliable and accessible for pharmaceutical quality assessment beyond centralized laboratory settings. Established techniques for drug quantification, including high-performance liquid chromatography (HPLC) and UV–visible spectroscopy, provide high accuracy and precision but are often associated with substantial cost, operational complexity, and solvent consumption. − Moreover, their reliance on specialized instrumentation and trained personnel restricts widespread implementation, particularly in resource-constrained regions. Consequently, there is growing interest in alternative methodologies that reduce analytical burden while preserving meaningful discriminatory precision. − Among these, analysis of deposition patterns formed during evaporation of sessile droplets has emerged as a promising non-destructive approach. , The morphology of dried droplets reflects underlying physicochemical processes influenced by solute concentration and composition, allowing chemically relevant information to be extracted through quantitative image analysis and multivariate statistical methods. −

The crystallization behavior of salts, polymers etc., has been widely investigated using various techniques including light microscopy, infrared thermography, , differential scanning calorimetry (DSC), X-ray crystallography, atomic force microscopy, and Fourier transform infrared spectroscopy (FTIR), in order to regulate its crystallinity and crystallization rate. These non-destructive methods of investigation have been widely adopted in recent years. However, unlike active infrared thermography, which involves temperature and emissivity variations that can alter the material’s physical state, an imaging technique allows for the detailed observation of crystal growth and structure without disturbing the sample. , Using imaging-based approaches enables continuous monitoring of the crystallization process, providing valuable insights into supersaturation buildup, burst-like nucleation events, rapid front propagation, creeping behavior, and the reversible transitions driven by humidity changes. , Unlike techniques that require physical contact, chemical modification, or extensive sample preparation, imaging-based analysis enables direct observation of the crystallization patterns after droplet evaporation without further disturbing the formed deposit. This non-contact approach preserves the morphology of the dried crystalline deposit, making it well suited for quantitative characterization of crystallization behavior.

In our previous work, we analyzed a diverse set of deposit patterns formed by 42 different salts. Through the use of a large-scale image dataset, we identified distinct families of crystallization patterns and achieved accurate pattern classifications using artificial intelligence. In subsequent studies, we conducted experiments under reduced pressure and microgravity to deepen our understanding of salt crystallization under non-equilibrium conditions. These investigations highlighted the influence of surface tension and the Marangoni effect on crystal formation. Additionally, we developed a robotic system capable of rapidly and precisely dispensing salt solutions, generating up to 2,500 high-resolution images of dried deposits per day. Finally, the image dataset was further tested using GPT-4o and GPT-4o-mini to evaluate their ability to identify specific salts. Recently, we investigated the influence of relative humidity on morphological and textural features of dried salt deposits using high-resolution imaging and machine learning/artificial intelligence (MI/AI) analyses. ,

Our previous studies established that quantitative image analysis combined with machine learning can reliably distinguish crystallization patterns arising from numerous inorganic salt systems over a range of concentrations and environmental conditions. Building upon this validated framework, the objective of the present study is not to re-establish the general applicability of the methodology, but rather to investigate whether the same approach can be successfully extended to a pharmaceutically relevant compound and provide quantitative concentration estimates for aspirin. In this work, a low-cost imaging-based method is investigated for estimating drug concentration through analysis of dried sessile droplet patterns formed from saturated aspirin solutions. Droplets prepared from defined concentration standards and an unknown commercial tablet sample were evaporated under controlled conditions and imaged at high resolution. Quantitative geometric and textural features were extracted from each droplet image and evaluated using principal component analysis (PCA) to identify concentration-dependent trends. Regression and machine learning models were subsequently employed to assess the predictive capability of the extracted image features. To provide comparison with a conventional analytical technique, a standard addition method was implemented and analyzed using UV-visible spectroscopy. Collectively, this study demonstrates the potential of sessile droplet imaging combined with statistical and machine learning analysis as an accessible and non-destructive tool for pharmaceutical quality assessment, particularly in settings where traditional analytical infrastructure is limited.

2. Materials and Methods

2.1. Chemicals and Reagents

Aspirin analytical standard was obtained from Sigma-Aldrich (CAS-No: 50-78-2), and a commercial aspirin tablet (Genuine Aspirin, 325 mg) was used as the sample tablet. Isopropanol (IPA) was purchased from Spectrum Chemical MFG Corp. and used as the sole solvent for preparation of saturated solutions, standards, sample tablet solutions, and standard addition experiments for UV-visible spectroscopy. All chemicals used were of analytical grade and were used as obtained without further purification.

2.2. Sample Preparation and Operation

A standard stock solution of aspirin was prepared by gradually dissolving 650 mg aspirin powder in 15 mL of pure isopropanol (IPA), under continuous stirring with a magnetic stirrer. To enhance dissolution, the mixtures were heated during stirring and the process was continued until an excess amount remained undissolved. After cooling, the solutions were filtered to remove residual solid particles. Working standard solutions ranging from 100% to 6% were subsequently prepared through serial dilution of the stock solution using IPA where 100% solution being saturated point. Using a micro- pipette (Eppendorf, model: I46914M) 3 μL droplets were dispensed onto microscope slides (VWR, Cat. No: 48300-026) and left to dry in ambient condition. After drying process, deposit patterns were captured using 24.3 MP full frame mirrorless camera (Nikon Z5) with a high-quality macro lens (Nikkor Z MC 105 mm f/2.8). To avoid shadows and maintain uniform illumination, a 3D-printed hood with 6500K white LED lights (phopollo (Amazon)) was used.

Commercial 325 mg aspirin tablets were weighed, finely ground using a mortar and pestle, and a portion of the powder equivalent to 650 mg of aspirin (two tablets) was transferred to a beaker. The sample was dissolved in 15.00 mL of isopropyl alcohol (IPA), producing a stock solution of 43.33 mg/mL, and the mixture was heated and stirred to ensure complete dissolution. The solution was subsequently filtered to remove insoluble excipients. For quantitative determination, the standard addition method was applied to minimize potential matrix effects from tablet components. Aliquots of 0.25 mL of the sample solution were transferred into separate vials and spiked with increasing volumes (0.00–1.00 mL) of a standard aspirin solution (43.33 mg/mL), followed by dilution to a final volume of 1.50 mL with IPA to obtain total concentrations ranging from 7.22 to 36.11 mg/mL. We further diluted each solution by 200 times for UV–visible measurements. Measurements were performed using a calibrated Lambda 25 UV/vis Spectrophotometer with IPA background correction. Each solution was scanned between 200-350 nm and absorbance was recorded using 1.00 cm quartz cuvettes. A calibration plot of absorbance versus added standard concentration was constructed, and the original aspirin concentration in the tablet sample was determined from the x-intercept of the linear regression line.

2.3. Image Processing and Regression

For each concentration of aspirin standard, 200 images were acquired and subsequently subjected to analysis. The original images (6048 × 4024 pixels) were cropped to a 3000 × 3000 pixel region containing the complete droplet while removing only the surrounding background. No portion of the droplet was excluded during cropping. A feature-based regression approach was implemented to estimate concentration using classical machine learning models. A total of 46 quantitative features were extracted from each image using MATLAB, and stored in separate text files corresponding to different concentration levels. This feature set was adopted because it was previously demonstrated to outperform a reduced 16-feature implementation in capturing concentration-dependent morphological variations in dried crystalline deposits. The selected descriptors encompass complementary geometric, intensity-based, and textural characteristics, providing a comprehensive quantitative representation of crystal morphology. The dataset was then divided into input features and the concentration labels. To improve data quality, invalid values such as positive or negative infinity were replaced with missing values, and Z-score-based outlier detection was applied using the Z-score method. For this, samples with feature values exceeding three standard deviations from the mean were removed, ensuring that extreme outliers did not bias the regression models. After filtering, all features were standardized using standardization (z-score normalization), which scales features to a zero mean and unit variance. This step improves model stability and ensures that features with larger numeric ranges do not dominate the learning process. Two tree-based ensemble regression models were then trained and compared: Random Forest Regressor and XGBoost Regressor. Random Forest builds multiple independent decision trees using bagging and averages their predictions to reduce variance and overfitting. In contrast, XGBoost uses gradient boosting, where trees are built sequentially to correct errors made by previous trees, often resulting in higher predictive accuracy. Model performance was evaluated using mean squared error (MSE), mean absolute error (MAE), and the coefficient of determination (R2), enabling a quantitative comparison between the two regression approaches. Table S1 shows the parameters used during the training and testing purpose for all models used in this study.

We also developed a deep learning-based regression framework to estimate salt concentration directly from image data. The models were implemented using the PyTorch library, and two convolutional neural network architectures, i.e., ResNet-50 and EfficientNet-B3, were employed to compare their performance for concentration predictions. , The dataset consisted of labeled images representing different salt concentration levels, organized into folders where the folder names encoded the corresponding concentration values. The dataset was divided into training, validation, and testing subsets in the ratio 70:20:10. The complete droplet image was used as the input for all analyses. The original images included in Figure b and Figure S4 contain scale bars representing the physical dimensions of the imaged area. For deep learning analysis, the cropped images were uniformly resized to 224 × 224 pixels for ResNet-50 and 300 × 300 pixels for EfficientNet-B3 to satisfy the input requirements of the pretrained models. No additional cropping, spatial selection, or image transformation was applied. The resizing operation preserved the overall crystallization pattern while reducing the image dimensions to meet the input requirements of the pretrained networks. Representative resized images are provided in Figure S2. Therefore, the extracted features and model predictions represent the overall droplet morphology, including variations in crystal distribution and orientation across different regions. All images were normalized using the standard mean and standard deviation values derived from the ImageNet dataset to ensure compatibility with pretrained weights. Both networks leveraged transfer learning, where pretrained weights were used to initialize the models, and the final classification layer was replaced with a fully connected layer containing a single neuron to perform regression and output continuous concentration values. Although the ImageNet dataset contains natural images that differ from crystallization patterns, ImageNet normalization was adopted because the networks were initialized with ImageNet-pretrained weights. This preprocessing maintains consistency with the original training configuration and allows effective transfer learning, while subsequent fine-tuning enables adaptation of the models to the crystallization image domain. ResNet-50 relies on residual connections that allow intricately deep networks to train effectively by mitigating vanishing gradient problems, whereas EfficientNet-B3 employs compound scaling to balance network depth, width, and resolution, often achieving higher accuracy with improved computational efficiency. During training, the Mean Squared Error (MSE) loss function was used, and model parameters were optimized using the Adam optimizer with a learning rate of 1×10–4. Training was conducted for up to 30 epochs with a batch size of 8, and gradient clipping was applied to stabilize training. To reduce overfitting, an early stopping strategy was implemented, terminating training if validation loss did not improve for seven consecutive epochs. The best-performing model based on validation loss was saved for evaluation. Model performance was assessed on the test dataset using regression metrics including the MSE, Mean Absolute Error (MAE), and the coefficient of determination (R2), enabling a quantitative comparison between the ResNet-50 and EfficientNet-B3 architectures.

1.

1

Open in a new tab

(a) Schematic of the system. (b) Deposit patterns of standard aspirin for six different concentrations: 6%, 12.5%, 25%, 50%, 75%, and 100%, where 100% corresponds to a saturated solution. Insets show 25 times zoomed area of bounding box for the corresponding patterns. Scale bar: 5 mm.

To evaluate the applicability of the trained models to a pharmaceutical sample, images from the commercial aspirin tablet were analyzed independently. These images were not included in the training, validation, or testing datasets used for model development.

3. Results and Discussion

3.1. Morphological and Textural Properties of Aspirin

Figure a shows the schematics of the experimental system. Briefly, a droplet of aspirin solution was placed on a glass slide and left it to dry under ambient conditions. Parts of the system are listed in the figure. The dried pattern was then imaged and stored in the computer for further analysis. Figure b shows the representative morphological pattern of aspirin droplets under different concentrations. At a low concentration (6%), the deposit exhibits a heterogeneous and relatively disordered morphology. The central region appears randomly textured with irregular, fine crystalline aggregates, while a distinct peripheral ring is visible, consistent with a capillary flow-driven “coffee-ring” effect. , Crystal domains are small and poorly defined, indicating limited solute availability and rapid nucleation during solvent evaporation. At 12.5% and 25%, the morphology transitions toward more organized radial structures. The deposits become more uniform across the droplet footprint, and fine, leaf-like or dendritic crystalline features emerge, particularly near the periphery. The ring structure remains discernible but is less dominant compared to the lowest concentration. These intermediate concentrations suggest a balance between outward capillary flow and bulk crystallization within the droplet interior. At 50%, the crystalline pattern becomes markedly denser and increasingly homogeneous. Radially oriented crystalline domains extend from the center toward the edge, forming a highly interconnected network. The increase in solute concentration enhances crystal growth and promotes further continuous coverage of the substrate, reducing empty or weakly textured regions. At 75%, pronounced faceted crystalline structures dominate the deposit. Large, well-defined crystalline plates and dendritic motifs are observed throughout the droplet area. The peripheral rim becomes thicker and more structured, indicating substantial solute accumulation at the contact line during evaporation. The overall morphology reflects enhanced crystal growth kinetics due to higher supersaturation levels. At 100% (saturated solution), the deposit exhibits the highest degree of structural organization and surface coverage. The entire droplet footprint is densely packed with faceted crystalline domains, and a prominent, highly textured peripheral rim is observed. Crystal size and thickness are maximized under these conditions, consistent with rapid nucleation and growth driven by extreme supersaturation during solvent loss. In addition, crystallization-based image analysis is sensitive to environmental factors such as temperature, relative humidity, and substrate properties. In this study, all experiments were conducted under controlled temperature and humidity conditions to minimize these effects and improve reproducibility. Previous investigations have shown that humidity variations can influence crystal morphology and pattern formation. Therefore, future implementation under variable environmental conditions will require additional calibration datasets, environmental monitoring, or incorporation of environmental parameters into the prediction framework to ensure reliable performance.

The observed concentration-dependent morphological evolution can be understood from the interplay between solvent evaporation, solute transport, supersaturation development, nucleation, and crystal growth. During evaporation, isopropanol is preferentially lost from the droplet surface, causing the local aspirin concentration to increase until supersaturation is reached. At lower initial concentrations, supersaturation develops more gradually, resulting in fewer nucleation sites and slower crystal growth. Under these conditions, outward capillary flow transports dissolved aspirin toward the pinned contact line, producing relatively sparse crystalline deposits with a pronounced peripheral ring. As the initial aspirin concentration increases, supersaturation is achieved earlier during evaporation, promoting more frequent nucleation events and faster crystal growth throughout the droplet interior. Consequently, crystal domains become denser, more interconnected, and increasingly faceted, leading to the highly organized morphologies observed for concentrated solutions.

3.2. Image Processing and Principal Component Analysis

Deposition patterns produced by dried aspirin sessile droplets exhibit subtle, highly intricate patterns as concentration varies, making direct visual differentiation challenging. To assess pattern consistency and extract concentration-related structural information, an initial quantitative analysis was conducted using Principal Component Analysis (PCA). As discussed in the Methods section, each droplet image was characterized using 46 feature-dependent parameters describing their textural properties. Figure S1 shows the workflow for extracting 46 parameters from aspirin deposits of 100% concentration. Each sample was thus embedded within a 46-dimensional space. After standardization of the feature set by z-score normalization, PCA was applied across the full range of aspirin concentrations studied.

To further interpret this high-dimensional feature space, PCA was used to condense the original 46 descriptors into a smaller set of orthogonal variables that capture the dominant sources of variation in the droplet patterns (Figure a). The first three principal components together account for 64.1% of the total variance (Figure b) and define a reduced three-dimensional representation of the data, in which each dried droplet is represented as a single point based on its extracted features. The resulting distributions form concentration-dependent groupings that occupy distinct regions of the reduced feature space. At higher concentrations, these groups appear more tightly clustered, whereas lower concentrations exhibit broader dispersion, a trend most clearly expressed in the two-dimensional projection defined by the first two principal components.

2.

2

Open in a new tab

Representation of approximately 1,000 aspirin deposit patterns, with marker colors indicating the concentration standard aspirin in IPA. For each deposit, 46 geometric and textural metrics were extracted, and the resulting dataset was transformed using principal component analysis (PCA). (a) The plot shows the coordinates of each deposit pattern in the space defined by the three dominant PCA components, with the variance explained by each component indicated in the axis labels. Additionally, all data points are projected onto the PCA1-PCA2 plane at the base of the plot. (b) Explained cumulative variance against number of principal components.

3.3. Regression using Multidimensional Machine Learning

Following the separation of aspirin droplet patterns in PCA1-3 space, we examined image metrics for concentrations ranging from 6% to 100%, with 100% representing saturation. Outliers exceeding three standard deviations from the mean for each concentration were removed from the z-scored dataset, and the remaining samples were divided into 70% for training, 10% for validation and 20% for testing. To evaluate the predictive potential of the image features, multidimensional nonlinear regression models were applied using Random Forest (Figure a) and XGBoost (Figure c), with concentration as the dependent variable and 46 image metrics as predictors. Predicted concentrations for the test set closely matched the true values, with Random Forest achieving an MSE of 15.25% and R2 of 0.98 and XGBoost an MSE of 12.87% and R2 of 0.98 These statistical outcomes are summarized in Table .

3.

3

Open in a new tab

A multidimensional nonlinear regression model was employed to evaluate aspirin deposition patterns. The 46 image-based metrics, describing the geometric and textural properties of the deposits, were treated as predictor variables, and the true concentrations were used as the response variable. (a) Predicted concentration is presented as mean ± standard deviation for each true concentration using Random Forest regression. (b) Top 10 important features as identified by the regression model (Random Forest). (c) Predicted concentration is presented as mean ± standard deviation for each true concentration using XGBoost regression model. (d) Top 10 important features as identified by the regression model (XGBoost).

1. Comparison of Metrics between Random Forest, XGBoost, ResNet-50 and EfficientNet-B3 on Standard Dataset for Aspirin.

Model Input parameters/Image Size MSE MAE R 2
Random Forest 46 15.25 1.75 0.98
XGBoost 46 12.87 1.38 0.98
ResNet-50 224 × 224 13.54 2.72 0.98
EfficientNet-B3 300 × 300 3.49 1.36 0.99
Open in a new tab

Both models identified the same top ten features as most influential, though their ranking differed. Notably, as shown in Figure b and Figure d, stdRaw was the second-most important feature in Random Forest, whereas intensitySkewness held that position in XGBoost. The relative contributions of individual metrics varied between models, highlighting differences in how each algorithm interprets the high-dimensional feature space. A limitation of this study is that only six predefined concentration levels were evaluated. Although these concentrations covered a wide range, intermediate concentration values were not experimentally included. Regression modeling was used to overcome this limitation by establishing a continuous relationship between image features and concentration, enabling prediction of concentrations between the tested levels. Future studies should include additional randomly distributed concentration levels to further validate and improve the robustness of the prediction model.

3.4. Direct Image-Based Regression

The predictive performance of the deep learning models was evaluated using regression analysis on the test dataset. A total of 1200 images of aspirin standard solutions at different concentration levels were used, and the dataset was divided using a 70-30 split for training and testing. The comparison between the two architectures, ResNet-50 and EfficientNet-B3, is summarized in Table , while the regression plots for the predicted versus true concentrations are presented in Figure a and Figure b, respectively. The ResNet-50 model demonstrated strong predictive capability, achieving an R 2 value of 0.98 with a Mean Squared Error (MSE) of 13.54 and a Mean Absolute Error (MAE) of 2.72 (Table ). As shown in Figure a, the predicted concentrations follow a linear relationship with the true concentrations; however, slight deviations from the ideal regression line are observed, particularly at higher concentration levels. The fitted regression equation y = 0.94x + 0.28 indicates a tendency for the model to slightly underestimate concentrations as the true value increases.

4.

4

Open in a new tab

Regression plots obtained for aspirin standard on test set for different concentrations using (a) ResNet-50. (b) EfficientNet-B3. The total number of images in test dataset is 200. Dataset was split to 70-30% split ratio with 10% used for validation.

In comparison, EfficientNet-B3 achieved higher predictive accuracy with an R 2 value of 0.99, MSE of 3.49, and MAE of 1.36 (Table ). The regression plot in Figure b shows that the predicted values closely align with the ideal line across the full concentration range. The fitted regression equation y = 1.02x – 0.36 suggests minimal systematic bias and improved agreement between predicted and actual concentrations. Residual distribution of obtained for aspirin standard on test set for different concentrations is shown in Figure S3.

The quantitative comparison presented in Table indicates that EfficientNet-B3 substantially reduces prediction errors relative to ResNet-50. The MSE decreased from 13.54 to 3.49 and the MAE was reduced from 2.72 to 1.36, while R2 value increased from 0.98 to 0.99. These results demonstrate that EfficientNet-B3 provides more accurate and consistent predictions for aspirin concentration from image data compared to ResNet-50. The observed differences in prediction error can be attributed to architectural characteristics of the two models. ResNet-50 relies on residual connections that enable effective training of deep networks by alleviating the vanishing gradient problem. While this structure is effective for general feature extraction, its fixed scaling of network depth and resolution may limit its ability to capture subtle visual variations in the images that correspond to small concentration differences. In contrast, EfficientNet-B3 employs a compound scaling strategy that simultaneously adjusts network depth, width, and input image resolution. The use of a higher input resolution (300 × 300) allows the model to preserve finer spatial details in the images, while balanced scaling improves feature representation. As a result, EfficientNet-B3 can better capture small visual differences associated with concentration changes, leading to lower prediction errors. It is to note that the larger prediction error observed at concentrations close to 100% is mainly attributed to the reduced distinguishability of image patterns at very high concentrations. At lower and intermediate concentrations, the crystals are more sparsely distributed, resulting in clearer variations in crystal morphology, size, and spatial distribution that can be effectively captured by the extracted image features. However, at very high concentrations, the crystals become overcrowded and overlap with each other, leading to a more uniform and saturated appearance. This reduces the visibility of individual crystal features and decreases the discriminative capability of the extracted morphological and textural features, which contributes to the increased prediction error.

Both architectures demonstrate the feasibility of applying convolutional neural networks for direct image-based regression of aspirin concentration. However, EfficientNet-B3 shows superior predictive performance, likely due to its optimized scaling strategy and ability to capture fine-grained visual features relevant to concentration estimation. These findings highlight the importance of model architecture in determining regression accuracy for image-based quantitative analysis.

3.5. Pharmaceutical Sample Concentration Determinations

To minimize the influence of matrix effects inherent in commercial aspirin tablets on crystal deposition patterns, the standard addition solutions prepared above were employed for both quantitative UV–vis analysis and crystal morphology studies. By using aliquots of the sample solution spiked with known amounts of standard, we ensured that the solute environment closely matched that of the standard, thereby reducing variability arising from excipients or other tablet constituents. These solutions not only provided reliable absorbance measurements for the three spectral evaluation approaches, but also served as the basis for controlled droplet deposition experiments, enabling consistent formation of crystal patterns suitable for subsequent machine learning analysis. Figure S4 presents deposit patterns of aspirin samples for six different concentrations: 10.94 %, 21.87 %, 32.83 %, 54.71 % and 73.77 %. These values represent total concentration of aspirin, including from standard and sample.

Quantitative analysis was performed using three approaches: (i) integration of the area under the curve across the full spectral range (200–360 nm), (ii) measurement of maximum absorbance within 220–230 nm, and (iii) measurement of maximum absorbance within 260–280 nm.

Figure a presents full spectral range for aspirin sample at different concentrations of added standards with 0 being solution containing only the sample. The standard addition plots demonstrated good linearity using all three spectral evaluation approaches. Integration of the area under the curve (AUC) for 200–360 nm produced the regression equation AUC = 1.717707·C + 6.834932 with excellent linearity (R2 = 0.99) as presented in Figure b. The concentration determined from the x-intercept yielded a final aspirin concentration of 49.04–63.66–81.20 mg/mL (lower bound-midpoint-upper bound, 95% confidence interval) in sample solution. When quantification was performed using the maximum absorbance within 220-230 nm (Figure c), the regression equation was Max Abs = 0.040883·C + 0.104768 (R2 = 0.98), giving a calculated concentration of 11.04–41.00–89.44 mg/mL (95% confidence interval). Similarly, analysis based on the maximum absorbance within 260–280 nm resulted in the equation Max Abs = 0.005529·C + 0.013073 (R2 = 0.98), corresponding to a final concentration of 13.12–37.84–73.92 mg/mL (95% confidence interval) (Figure d). Among the three approaches, the full-spectrum AUC method exhibited the highest coefficient of determination and comparatively narrower confidence limits. The quantitative validation of commercial tablet samples is limited by the uncertainty associated with the UV–vis reference measurements. Consequently, the reported prediction performance should be interpreted within the uncertainty of the reference method. Future studies employing higher-precision analytical techniques, such as HPLC, would provide a more rigorous benchmark for evaluating the proposed image-based approach.

5.

5

Open in a new tab

UV–visible spectra of aspirin obtained using the standard addition method. (a) UV–visible spectra for range 200-360 nm at different concentrations. (b) Measured area under curve for the range 200–360 nm. (c) Maximum absorbance value recorded within the range 220–230 nm for different concentrations. (d) Maximum absorbance value recorded within the range 260–280 nm for different concentrations. Concentrations shown in all figures are for added standard aspirin in the units of μg/mL.

The UV spectrum of aspirin showed three main absorption regions (∼204 nm, 220–230 nm, and 260–280 nm). The intense band around 204 nm was not used for quantitative analysis because it corresponds to a strong π → π* transition of the aromatic ring, which is highly non-specific and prone to interference from solvents (e.g., IPA), impurities, and baseline noise; additionally, absorbance at this wavelength is often too high to remain within the linear range of the Beer–Lambert law. , The band in the 220–230 nm region also arises primarily from aromatic π → π* transitions influenced by substituents and, although measurable, it lacks selectivity because both aspirin and potential degradation products absorb significantly in this region, making it unsuitable for specific aspirin quantification. In contrast, the absorption band in the 260–280 nm range is more appropriate for aspirin analysis because it is associated with transitions influenced by the ester carbonyl bond conjugated with the aromatic ring of the acetylsalicylic acid, providing better selectivity and improved linearity for quantitative measurements. , The negligible absorbance observed at 280 nm and above indicates the absence of a significant peak attributable to salicylic acid, which typically shows a strong absorption near this region; this confirms that no appreciable hydrolysis of aspirin occurred under the experimental conditions.

As shown in Figure , the concentration of aspirin in pharmaceutical tablets was predicted using four regression models: Random Forest, XGBoost, ResNet-50, and EfficientNet-B3, with a dataset of 192 images spanning five concentration levels. For Random Forest and XGBoost, 46 extracted image parameters were used, while ResNet-50 and EfficientNet-B3 employed raw images of sizes 224 × 224 and 300 × 300, respectively. As shown in Table , XGBoost achieved the lowest mean squared error (19.92) and mean absolute error (2.75) with an R2 of 0.90, slightly outperforming Random Forest (MSE = 22.02, MAE = 3.09, R2 = 0.89). Deep learning models exhibited comparable R2 values (0.88–0.89) but slightly higher errors, likely due to the relatively small test dataset size. In conclusion, feature extraction-based models demonstrated higher stability and slightly better predictive accuracy compared to direct image-based deep learning approaches.

6.

6

Open in a new tab

Multidimensional regression models and deep learning models were employed to evaluate pharmaceutical aspirin deposition patterns. Predicted concentration vs true concentration plot using (a) Random Forest regression. (b) XGBoost regression model. (c) ResNet-50 (d) EfficientNet-B3. Total number of images for all concentrations is 192. Dataset was split to 70–30% split ratio where 30% data was used for validation and testing.

2. Performance Metrics of Random Forest, XGBoost, ResNet-50, and EfficientNet-B3 Models for Predicting Aspirin Concentration Using Extracted Image Features (Random Forest, XGBoost) or Raw Images (ResNet-50, EfficientNet-B3) from Standard Addition Experiments.

Model Input parameters/Image Size MSE MAE R 2
Random Forest 46 22.02 3.09 0.89
XGBoost 46 19.92 2.75 0.90
ResNet-50 224 × 224 24.38 3.53 0.89
EfficientNet-B3 300 × 300 25.65 4.23 0.88
Open in a new tab

Predicted concentrations for each model are summarized in Table (also refer to Figure S5). Random Forest and XGBoost closely matched actual concentrations, although small overestimations were observed at lower concentration levels (Figure a and b), whereas ResNet-50 tended to underestimate mid-range concentrations (Figure c). EfficientNet-B3 predictions were generally accurate but showed larger variability at lower concentrations (Figure d). EfficientNet-B3 generally provided predictions closer to the reference concentrations compared with Random Forest and XGBoost, suggesting improved accuracy with respect to prediction bias. However, some increased variability was observed in its predictions, indicating that accuracy and precision should be considered separately. The higher variance may result from the greater sensitivity of deep learning models to morphological variations within crystallization patterns. Therefore, while EfficientNet-B3 achieved strong predictive performance, further optimization with larger and more diverse datasets may improve prediction consistency. These results suggest that classical machine learning models are more reliable for small datasets, while deep learning models may require more data or augmentation to achieve similar precision. The application of the standard addition method effectively minimized matrix effects, allowing the models to capture aspirin-specific features accurately. Collectively, these findings demonstrate that both feature-based machine learning and deep learning can be used for rapid, image-based quantification of pharmaceutical compounds, with XGBoost providing the most balanced performance across all concentration levels. To enable direct comparison with the UV/Vis standard addition results, the predicted concentrations from all ML models were transformed into the normalized μg/mL scale, using the sample baseline as zero. This approach allows visual alignment of AI predictions with experimental values, while avoiding direct extrapolation from the ML models beyond the calibrated range (Figure S6).

3. Comparison of Actual and Predicted Aspirin Concentrations (%) for All Four Models, Including Standard Deviations, Highlighting Model Accuracy Across Different Concentration Levels. Data Presents Average Value of Predictions along with Standard Deviation (Approximately n = 40 Each).

  Predicted Concentration (%)
Actual Concentration (%) Random Forest XGBoost ResNet-50 EfficientNet-B3
10.94 12.99 ± 1.20 12.32 ±1.14 12.01 ± 0.33 10.37 ± 1.94
21.88 28.26 ± 2.14 27.30 ± 2.20 25.56 ± 2.11 24.88 ± 0.51
32.83 35.91 ± 0.72 35.36 ± 0.93 28.67 ± 2.79 33.50 ± 3.99
43.77 41.73 ± 0.73 41.00 ± 0.88 37.99 ± 0.35 42.29 ± 2.31
54.71 54.08 ± 0.24 54.73 ± 0.13 55.69 ± 0.53 55.20 ± 1.98
Open in a new tab

4. Conclusion

This study demonstrates that dried sessile droplet patterns of aspirin solutions contain quantifiable morphological information that can be effectively utilized for concentration prediction through image analysis and artificial intelligence. As aspirin concentration increased from dilute to saturated conditions, the resulting deposition patterns exhibited systematic changes in morphology, ranging from heterogeneous and loosely organized structures at low concentrations to dense, well-defined crystalline networks at higher concentrations. These concentration-dependent structural variations provided measurable geometric and textural features that can be analyzed quantitatively. Principal component analysis of the extracted 46 image features confirmed that droplet patterns cluster according to concentration, indicating that the morphological descriptors capture meaningful physicochemical differences associated with crystallization during solvent evaporation. Machine learning and deep learning models were subsequently applied to evaluate the predictive capability of these image-derived features. Feature-based ensemble models, including Random Forest and XGBoost, demonstrated strong predictive performance, achieving coefficients of determination close to 0.98 when applied to the standard dataset. In parallel, deep learning architectures trained directly on the image data also produced accurate predictions. Among these models, EfficientNet-B3 showed the highest performance, achieving an R2 of 0.99 with significantly reduced prediction errors compared to ResNet-50. The improved accuracy of EfficientNet-B3 is attributed to its compound scaling strategy and higher input image resolution, which allow the model to capture subtle spatial features associated with concentration-dependent crystallization patterns. The methodology was further evaluated using a commercial aspirin tablet sample, and predicted concentrations were compared with values obtained using a conventional UV-visible spectrophotometric standard addition method. The normalized ML predictions closely align with the UV–vis results, demonstrating that image-based regression can reliably estimate aspirin concentrations in pharmaceutical tablets, even in the presence of matrix effects. Although a full-frame mirrorless camera (Nikon Z5) with a macro lens was used in this study to ensure high-quality image acquisition, these components were available in our laboratory and are not essential for implementation of the proposed method. The imaging system can be constructed using lower-cost cameras, such as a Raspberry Pi HQ camera combined with a suitable macro lens, with an estimated imaging hardware cost below USD 100. Although the effective image resolution of current research (approximately 9 MP) is comparable to that of several low-cost imaging systems, including the Raspberry Pi HQ camera, the effect of camera sensor characteristics and optical performance on prediction accuracy has not been evaluated in this study. Therefore, the suitability of lower-cost imaging systems should be validated in future work. Therefore, the low-cost advantage of this approach refers to the potential affordability and accessibility of the imaging platform compared with conventional analytical instruments. In addition, while sample availability is generally not a major limitation for pharmaceutical tablets, the non-destructive characteristic of image-based analysis allows samples to be retained for further testing, reference analysis, or long-term evaluation. The method also avoids destructive sample preparation procedures, enabling rapid and straightforward screening. The results demonstrate that imaging-based regression models can better predict the concentration values observed using traditional analytical techniques, while offering a simpler and a potentially lower-cost alternative. Although the predictive accuracy for pharmaceutical samples was slightly lower than that obtained for standard chemical solutions, the results remain promising given the limited dataset and the presence of excipients in the pharmaceutical samples. Future work should focus on expanding the experimental dataset to include a larger number of droplet images across broader concentration ranges and varying environmental conditions. In addition, the development of synthetic datasets using data augmentation or physics-informed simulation techniques may help improve model robustness and generalization performance, particularly for deep learning architectures that benefit from large training datasets. The commercial aspirin tablet served as a representative proof-of-concept demonstrating that the proposed methodology can be extended from previously studied inorganic salt systems to pharmaceutical samples. While the present work focuses on aspirin as a model compound, the underlying imaging and machine learning framework has previously been validated across diverse crystallizing systems. Future studies will investigate additional pharmaceutical formulations and active ingredients to further broaden the scope of the approach. In addition, the present study used UV–vis spectroscopy as the reference analytical method for concentration determination. Although UV–vis provides a rapid and accessible approach for quantitative analysis, it may introduce higher measurement uncertainty compared with more advanced reference techniques such as HPLC. Future investigations will include validation against HPLC-derived concentrations to further improve the accuracy of reference measurements and strengthen the evaluation of the proposed image-based prediction framework. This work highlights the potential of combining sessile droplet imaging with machine learning for rapid, non-destructive pharmaceutical analysis. With further expansion of datasets, improved image acquisition protocols, and integration with automated imaging platforms, this approach may evolve into a practical tool for rapid and decentralized screening of substandard or falsified pharmaceuticals, particularly in resource-limited settings.

Supplementary Material

ao6c05231_si_001.pdf (844.5KB, pdf)

Acknowledgments

This work was supported by NASA through NASA grant 80NSSC23M0050. We thank Prof. Bishnu Regmi for assistance with the UV–vis spectrophotometry.

The data that support the plots within this paper and other findings of this study are available from the corresponding author upon reasonable request.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c05231.

  • MATLAB workflow for quantitative image analysis of aspirin deposition patterns; extraction of 46 image-based parameters from aspirin deposits; machine-learning model parameters for Random Forest and XGBoost regression; image preprocessing procedures for ResNet-50 and EfficientNet-B3; residual distributions of ResNet-50 and EfficientNet-B3 predictions for aspirin standards; representative aspirin deposit patterns at different concentrations; predicted versus true concentrations for pharmaceutical aspirin deposits using Random Forest, XGBoost, ResNet-50, and EfficientNet-B3; and comparison of predicted aspirin concentrations with the concentration domains obtained from UV–vis spectroscopy. (PDF)

S.P. and B.B.D. designed the work, D.E. performed the experimental work, S.P. analyzed the data, O.S. and B.B.D. obtained funding, and all authors contributed in writing and editing of the manuscript.

The authors declare no competing financial interest.

References

  1. Caudron J. M., Ford N., Henkens M., Macé C., Kiddle-Monroe R., Pinel J.. Substandard medicines in resource-poor settings: a problem that can no longer be ignored. Tropical Medicine & International Health. 2008;13(8):1062–1072. doi: 10.1111/j.1365-3156.2008.02106.x. [DOI] [PubMed] [Google Scholar]
  2. Ozawa S., Evans D. R., Bessias S., Haynie D. G., Yemeke T. T., Laing S. K., Herrington J. E.. Prevalence and estimated economic burden of substandard and falsified medicines in low-and middle-income countries: a systematic review and meta-analysis. JAMA network open. 2018;1(4):e181662–e181662. doi: 10.1001/jamanetworkopen.2018.1662. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Al-Farhan B. S., Eldin G. M.. Chemometric-assisted UV–vis spectrophotometric method for the simultaneous quantification of triple anthelmintics in veterinary pharmaceuticals. Sci. Rep. 2025;15(1):44663. doi: 10.1038/s41598-025-31003-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Kovacs S., Hawes S. E., Maley S. N., Mosites E., Wong L., Stergachis A.. Technologies for detecting falsified and substandard drugs in low and middle-income countries. PloS one. 2014;9(3):e90601. doi: 10.1371/journal.pone.0090601. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Zhan X., Zhao Z., Yuan X., Wang Q., Li D., Xie H., Li X., Zhou M., Duan Y.. Microwave-Induced Plasma Desorption/Ionization Source for Ambient Mass Spectrometry. 2013;85:4512. doi: 10.1021/ac400296v. [DOI] [PubMed] [Google Scholar]
  6. Flick T. G., Leib R. D., Williams E. R.. Direct standard-free quantitation of tamiflu and other pharmaceutical tablets using clustering agents with electrospray ionization mass spectrometry. Analytical chemistry. 2010;82(4):1179–1182. doi: 10.1021/ac902277d. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Larson R. G.. Transport and deposition patterns in drying sessile droplets. AIChE Journal. 2014;60(5):1538–1571. doi: 10.1002/aic.14338. [DOI] [Google Scholar]
  8. Bhattarai, R. K. ; Pudasaini, S. ; Barstis, T. ; Giri, B. . Smartphone-Assisted Thin-Layer Chromatography for Rapid Quality Screening of Metformin. Adv. Pharm. Pharm. Sci. 2025, 2025 (1). 10.1155/adpp/3306550. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Bhattarai R. K., Pudasaini S., Sah M., Neupane B. B., Giri B.. Handmade paper as a paper analytical device for determining the quality of an antidiabetic drug. ACS omega. 2022;7(16):14074–14081. doi: 10.1021/acsomega.2c00633. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Brutin D., Starov V.. Recent advances in droplet wetting and evaporation. Chem. Soc. Rev. 2018;47(2):558–585. doi: 10.1039/C6CS00902F. [DOI] [PubMed] [Google Scholar]
  11. Wang F., He H., Wang X., Li Z., Gallego-Perez D., Guan J., Lee L. J.. Micropatterned Thermoresponsive Surfaces by Polymerization of Monomer Crystals: Modulating Cellular Morphology and Cell–Substrate Interactions. Analytical chemistry. 2012;84(21):9439–9445. doi: 10.1021/ac302267z. [DOI] [PubMed] [Google Scholar]
  12. Farkas D., Madarász L., Nagy Z. K., Antal I., Kállai-Szabó N.. Image analysis: A versatile tool in the manufacturing and quality control of pharmaceutical dosage forms. Pharmaceutics. 2021;13(5):685. doi: 10.3390/pharmaceutics13050685. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Pudasaini S., S V A., Steinbock O., Dangi B. B.. Morphological Signatures of Salt Crystals under Controlled Humidity Using Advanced Image Analysis. J. Phys. Chem. B. 2025;129(39):10184–10192. doi: 10.1021/acs.jpcb.5c04762. [DOI] [PubMed] [Google Scholar]
  14. SV A., Batista B. C., Dangi B. B., Steinbock O.. Water hardness from photos of dried drop stains: Reporter salts enable low-concentration detection. Environ. Sci. Technol. 2025;59(26):13286–13293. doi: 10.1021/acs.est.5c05770. [DOI] [PubMed] [Google Scholar]
  15. Rodriguez-Navarro C., Doehne E., Sebastian E.. How does sodium sulfate crystallize? Implications for the decay and testing of building materials. Cement and concrete research. 2000;30(10):1527–1534. doi: 10.1016/S0008-8846(00)00381-1. [DOI] [Google Scholar]
  16. Avdelidis N., Moropoulou A., Theoulakis P.. Detection of water deposits and movement in porous materials by infrared imaging. Infrared Physics & Technology. 2003;44(3):183–190. doi: 10.1016/S1350-4495(02)00212-8. [DOI] [Google Scholar]
  17. Vázquez P., Thomachot-Schneider C., Mouhoubi K., Fronteau G., Gommeaux M., Benavente D., Barbin V., Bodnar J.-L.. Infrared thermography monitoring of the NaCl crystallisation process. Infrared Physics & Technology. 2015;71:198–207. doi: 10.1016/j.infrared.2015.03.013. [DOI] [Google Scholar]
  18. Bianchi O., Oliveira R., Fiorio R., Martins J. D. N., Zattera A., Canto L.. Assessment of Avrami, Ozawa and Avrami–Ozawa equations for determination of EVA crosslinking kinetics from DSC measurements. Polym. Test. 2008;27(6):722–729. doi: 10.1016/j.polymertesting.2008.05.003. [DOI] [Google Scholar]
  19. Dewanckele J., De Kock T., Boone M., Cnudde V., Brabant L., Boone M., Fronteau G., Van Hoorebeke L., Jacobs P.. 4D imaging and quantification of pore structure modifications inside natural building stones by means of high resolution X-ray CT. Science of the total environment. 2012;416:436–448. doi: 10.1016/j.scitotenv.2011.11.018. [DOI] [PubMed] [Google Scholar]
  20. Radenović N., van Enckevort W.. Pyramids on {1 0 0} NaCl after formamide etching. Journal of crystal growth. 2002;234(2-3):589–598. doi: 10.1016/S0022-0248(01)01740-7. [DOI] [Google Scholar]
  21. Chen Z., Hay J. N., Jenkins M.. FTIR spectroscopic analysis of poly (ethylene terephthalate) on crystallization. European Polymer Journal. 2012;48(9):1586–1610. doi: 10.1016/j.eurpolymj.2012.06.006. [DOI] [Google Scholar]
  22. Hany C., Pradere C., Toutain J., Batsale J.-C.. A millifluidic calorimeter with infrared thermography for the measurement of chemical reaction enthalpy and kinetics. Quantitative InfraRed Thermography Journal. 2008;5(2):211–229. doi: 10.3166/qirt.5.211-229. [DOI] [Google Scholar]
  23. Batista B. C., Tekle S. D., Yan J., Dangi B. B., Steinbock O.. Chemical composition from photos: Dried solution drops reveal a morphogenetic tree. Proc. Natl. Acad. Sci. U. S. A. 2024;121(27):e2405963121. doi: 10.1073/pnas.2405963121. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Hadidi R., Pinckney V. D., Shaw S. A., Steinbock O., Dangi B. B.. Understanding the Salt Crystallizations from Droplets under Various Gravity and Pressure Environments: Display of the Marangoni Effect? J. Phys. Chem. B. 2025;129(11):3028–3040. doi: 10.1021/acs.jpcb.4c06963. [DOI] [PubMed] [Google Scholar]
  25. Batista B. C., S V A., Yan J., Dangi B. B., Steinbock O.. High-throughput robotic collection, imaging, and machine learning analysis of salt patterns: composition and concentration from dried droplet photos. Digital Discovery. 2025;4(4):1030–1041. doi: 10.1039/D4DD00333K. [DOI] [Google Scholar]
  26. Dangi, D. B. ; Dangi, B. B. ; Steinbock, O. . Evaluation of GPT-4o and GPT-4o-Mini’s Vision Capabilities for Compositional Analysis from Dried Solution Drops. ACS Omega 2025. 10.1021/acsomega.5c01150 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Pudasaini S., Steinbock O., Dangi B. B.. Effect of Humidity Cycling on Salt Hydration and Dehydration: Quantitative Image Analysis and Kinetic Interpretation. Cryst. Growth Des. 2026;26:2754. doi: 10.1021/acs.cgd.5c01774. [DOI] [Google Scholar]
  28. Repository, G. WaterHardness, https://www.github.com/osteinbock/WaterHardness (accessed March 10, 2026).
  29. Breiman L.. Random forests. Machine learning. 2001;45(1):5–32. doi: 10.1023/A:1010933404324. [DOI] [Google Scholar]
  30. Chen T., Guestrin C.. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 2016:785–794. [Google Scholar]
  31. Paszke, A. ; Gross, S. ; Massa, F. ; Lerer, A. ; Bradbury, J. ; Chanan, G. ; Killeen, T. ; Lin, Z. ; Gimelshein, N. ; Antiga, L. . Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 2019, 32. [Google Scholar]
  32. Tan, M. ; Le, Q. . Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning; PMLR, 2019; pp 6105–6114. [Google Scholar]
  33. He K., Zhang X., Ren S., Sun J.. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition. 2016:770–778. [Google Scholar]
  34. Weon B. M., Je J. H.. Capillary force repels coffee-ring effect. Physical Review EStatistical, Nonlinear, and Soft Matter Physics. 2010;82(1):015305. doi: 10.1103/PhysRevE.82.015305. [DOI] [PubMed] [Google Scholar]
  35. Deegan R. D., Bakajin O., Dupont T. F., Huber G., Nagel S. R., Witten T. A.. Capillary flow as the cause of ring stains from dried liquid drops. Nature. 1997;389(6653):827–829. doi: 10.1038/39827. [DOI] [Google Scholar]
  36. Hotelling H.. Analysis of a complex of statistical variables into principal components. Journal of educational psychology. 1933;24(6):417. doi: 10.1037/h0071325. [DOI] [Google Scholar]
  37. Kumar V., Prabha C., Sharma P., Mittal N., Askar S. S., Abouhawwash M.. Unified deep learning models for enhanced lung cancer prediction with ResNet-50–101 and EfficientNet-B3 using DICOM images. BMC medical imaging. 2024;24(1):63. doi: 10.1186/s12880-024-01241-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Rossmann J., Gurke R., Renner L. D., Oertel R., Kirch W.. Evaluation of the matrix effect of different sample matrices for 33 pharmaceuticals by post-column infusion. Journal of Chromatography B. 2015;1000:84–94. doi: 10.1016/j.jchromb.2015.06.019. [DOI] [PubMed] [Google Scholar]
  39. Friedel, R. A. ; Orchin, M. . Ultraviolet spectra of aromatic compounds; Wiley: New York, 1951. [Google Scholar]
  40. Perkampus, H.-H. UV–vis Spectroscopy and its Applications; Springer Science & Business Media, 2013. [Google Scholar]
  41. Elmasry M. S., Serag A., Hassan W. S., El-Mammli M. Y., Badrawy M.. Spectrophotometric determination of aspirin and omeprazole in the presence of salicylic acid as a degradation product: a comparative evaluation of different univariate/multivariate post processing algorithms. Journal of AOAC International. 2022;105(1):309–316. doi: 10.1093/jaoacint/qsab105. [DOI] [PubMed] [Google Scholar]
  42. Dacić, M. ; Uzunović, A. ; Kunić, A. ; Pilipović, S. ; Šapčanin, A. . UV–vis determination of acetylsalicylic acid in aspirin tablets using different solvents and conditions. International Conference on Medical and Biological Engineering; Springer: 2019; pp 563–567. [Google Scholar]
  43. Albakaa, A. R. M. ; Ahmed, M. a. ; Mohammed, B. t. ; Jabbar, z. a. . Development method for determination of aspirin as sodium salicylate by UV–vis spectroscopy. In IOP Conference Series: Materials Science and Engineering; IOP Publishing: 2019; Vol. 571, p 012104. [Google Scholar]
  44. Edwards L.. The hydrolysis of aspirin. A determination of the thermodynamic dissociation constant and a study of the reaction kinetics by ultra-violet spectrophotometry. Trans. Faraday Soc. 1950;46:723–735. doi: 10.1039/tf9504600723. [DOI] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

ao6c05231_si_001.pdf (844.5KB, pdf)

Data Availability Statement

The data that support the plots within this paper and other findings of this study are available from the corresponding author upon reasonable request.

Articles from ACS Omega are provided here courtesy of American Chemical Society

RESOURCES