SERS Combined with PCA-SVM-Based Identification Study of Different Brands of Levetiracetam

Abstract

This study addresses critical concerns in epilepsy management regarding generic drug substitution and the bioequivalence risks associated with switching antiepileptic drugs (AEDs). For patients undergoing long-term treatment, precise brand identification of levetiracetam (LEV) holds significant clinical value to ensure therapeutic consistency and safety. To meet this need, we developed a qualitative brand discrimination method based on surface-enhanced Raman spectroscopy (SERS) combined with machine learning. Silver sol was used as the SERS substrate, with potassium iodide as an aggregating agent and Rhodamine 6G (R6G) as an optimized probe. After validating LEV’s characteristic Raman peaks through integrated theoretical and experimental approaches, spectra from three commercial LEV brands were analyzed using principal component analysis (PCA) and support vector machine (SVM) classification. The model achieved 100% accuracy, precision, and sensitivity in 10-fold stratified cross-validation, demonstrating robust brand-discrimination capability. Unlike conventional quantitative assays that verify API content but cannot distinguish between different manufacturers, this SERS-based approach detects formulation-specific differencessuch as variations in crystal form and preparation processthrough unique spectral fingerprints. It thus provides a rapid, label-independent tool for brand authentication, batch consistency evaluation, and on-site pharmaceutical screening, effectively complementing pharmacopeial methods. This work supports medication safety and offers a reliable strategy for drug quality control and market surveillance, particularly in preventing counterfeit drugs and ensuring formulation stability in long-term epilepsy therapy.

1. Introduction

Epilepsy is one of the most prevalent and complex disorders of neurological disorders, affecting an estimated 65 million individuals worldwide. At one-year follow-up, persistent seizures have been shown to increase the risk of physical injuries (e.g., fractures, burns, concussions), with a risk range from 20% to 40%. The primary goals of epilepsy treatment are 2-fold: (1) to reduce the frequency of seizures and (2) to minimize the adverse effects associated with antiepileptic drugs (AEDs). , Ultimately, the goal is to achieve seizure freedom while ensuring optimal tolerability. However, the majority of pharmacological treatments for epilepsy remain imprecise, primarily focusing on empirically reducing seizures rather than targeting specific disease processes. ,

Levetiracetam (LEV), a second-generation AED, has demonstrated superior tolerability and efficacy compared to other antiepileptic agents, making it one of the most commonly prescribed medications. LEV is a broad-spectrum AED that utilizes a distinctive mechanism of action. It interacts with synaptic vesicle protein 2A (SV2A), − a feature that distinguishes it from other antiepileptic drugs. Its efficacy, safety, and tolerability have been well established in both adult and pediatric populations. Nevertheless, multiple LEV tablet brands are commercially available worldwide. Although these products generally meet regulatory standards for active pharmaceutical ingredient content, differences may exist in excipient composition, crystal form, and manufacturing processes. Such brand-related variations have raised clinical concerns among healthcare professionals and patients, particularly regarding unintended brand substitution during long-term antiepileptic therapy, which may lead to fluctuations in therapeutic stability. In response to this practical need, the present study was devoted to developing a rapid and reliable qualitative method for LEV brand identification.

The primary techniques for identifying drug constituents include high-performance liquid chromatography (HPLC), titrimetric analysis, and the spectrophotometric method, among others. Despite their accuracy, these methods often require costly, large-scale instrumentation, complex sample preparation, and skilled personnel, making them expensive, time-consuming, and impractical for rapid on-site detection. , Surface-enhanced Raman spectroscopy (SERS), a form of molecular vibrational spectroscopy, offers several advantages, including portability, rapidity, sensitivity, and nondestructiveness. These advantages make it an ideal tool for detecting pharmaceutical ingredients. −

In recent years, SERS has become a widely adopted technique in pharmaceutical research. For example, deep ultraviolet resonance Raman spectroscopy has been utilized for the highly sensitive detection of antibacterial agents in aqueous media and human urine. , Moreover, cellulose hydrogel SERS platforms combined with machine learning enable sensitive label-free identification of antithyroid drugs in plasma. Likewise, machine learning coupled bimodal SERS detection achieves ultrasensitive simultaneous screening of multiple quinolone antibiotic residues in milk. Machine learning-based feature extraction from SERS data has been applied to the detection and characterization of amphetamine analogs. In studies on LEV, Raman spectroscopy has been used to investigate polymorphisms in the chiral system of LEVs and the impact of pressure on these properties. Additionally, SERS has been employed to develop novel paper-based substrates for AED detection in the treatment of meningitis or encephalitis and to assess the CNS distribution of these drugs.

Previous studies have demonstrated that the integration of machine learning significantly enhances the practical application of SERS in drug testing. This enhancement is achieved by leveraging the autonomous learning capabilities and advanced data processing of machine learning, which enables accurate and sensitive qualitative analysis. For example, a classification study of various carbonate and sulfate mineral medicines using Raman spectroscopy, in conjunction with a principal component analysis-support vector machine (PCA-SVM) algorithm, has been conducted. The application of machine learning in SERS has enabled the biochemical differentiation of yeast cells expressing Bcl-2 and Mcl-1 after treatment with the antifungal agents ketoconazole and fluconazole. Furthermore, the combination of a magnetic Fe₃O₄@COF@Ag SERS substrate with machine learning algorithms facilitated the detection of three quinolone antibiotics: ciprofloxacin, norfloxacin and levofloxacin, among others. Despite these advancements, the application of SERS coupled with PCA-SVM has yet to be utilized for differentiating between various brands of LEV tablets in both domestic and international contexts. Additionally, SERS has been applied to develop novel paper-based substrates for AED detection in meningitis or encephalitis and to assess the central nervous system (CNS) distribution of these drugs.

This study demonstrates a rapid method for identifying commercial brands of LEV tablets by combining SERS with machine learning. Using silver nanoparticles as the SERS substrate, we acquired characteristic spectra from three different LEV products. A PCA-SVM model was then developed based on these spectral features, which successfully enabled precise brand discrimination. These findings underscore the practical utility of this SERS-machine learning approach for antiepileptic drug authentication, presenting a highly promising strategy for efficient drug classification.

2. Materials and Methods

2.1. Instruments and Reagents

Raman spectra were acquired using a BWS415–785H portable Raman spectrometer (excitation wavelength: 785 nm), with a spectral range of 68–2700 cm^–1 and a resolution of <3 cm^–1. The laser power was set to 30 mW, and the integration time was 5 s. Scanning electron micrographs (SEM) of the silver sols were captured using a Hitachi SU8010 scanning electron microscope.

Silver nitrate (AgNO₃) and sodium citrate were obtained from Sinopharm Chemical Reagent Co. Ascorbic acid (AA), potassium iodide (KI), sodium chloride (NaCl), and sodium sulfate (Na₂SO₄) were sourced from Xilong Chemical Co. Levetiracetam reagent (250 mg) was obtained from Aladdin Chemical Co. All experiments were performed using ultrapure water. Filter paper (rapid quantitative type, pore size 20–25 μm) was sourced from Fushun Civil Affairs Filter Paper Factory.

Levetiracetam tablets (500 mg) from various brands were procured from the Chinese mainland market, specifically JiYiKe (Tablet 1), PuLuo (Tablet 2), and Keppra (Tablet 3).

2.2. Preparation of Silver Sol

Silver sols were prepared using a modified method based on the procedure described by Zhang et al. , Ascorbic acid (26.45 mg) and sodium citrate (220.6 mg) were weighed and added to a 250 mL conical flask. These compounds served as the reducing and stabilizing agents, respectively, for the synthesis of silver nanoparticles. Initially, 250 mL of an aqueous solution containing ascorbic acid (6.0 × 10^–4 mol/L) and sodium citrate (3.0 × 10^–3 mol/L) was heated to 30 °C with gentle stirring in a water bath. Subsequently, 2.5 mL of 0.1 mol/L AgNO₃ was rapidly introduced into the solution. After 15 min of stirring at 900 rpm, the mixture was heated to 100 °C and maintained at this temperature for 2 h, allowing the spherical nanoparticles to further uniformize. The resulting silver sol was cooled to room temperature and stored at 4 °C in the dark.

2.3. Preparation of Samples

Three commercial LEV tablet brands were used in this study. For each brand, tablets were obtained from a single production batch to avoid batch-to-batch variation and ensure consistency in raw materials, excipients, and manufacturing processes. One representative tablet was randomly chosen from the single batch of each brand, giving a total of three tablets for the whole experiment.

To ensure data independence and avoid sample redundancy, 20 conventional Raman spectra and 20 SERS spectra were collected from 20 separate and nonoverlapping positions on each tablet. This strategy reduces the influence of local heterogeneity on the tablet surface and avoids repeated sampling at the same site. Each spectrum was therefore treated as an independent sampling unit.

This resulted in 60 conventional Raman spectra and 60 SERS spectra in total, amounting to 120 spectra for subsequent analysis. This sampling approach ensured data reproducibility.

The spectral data sets used in this study possess reliable independence at three levels. First, interbrand independence is guaranteed by three different manufacturers. Second, intratablet spatial independence is achieved by collecting spectra from 20 nonoverlapping positions on each tablet. Third, batch uniformity is ensured by using tablets from a single batch for each brand, which eliminates interference from batch-related factors. This systematic sampling design meets the strict requirements of multivariate analysis and supports the reliability of the PCA and SVM classification results.

Spectral preprocessing was systematically executed to eliminate background interference, mitigate noise artifacts, and ensure interspectral comparability across all measurements. These actions are critical prerequisites for reliable multivariate data analysis, including subsequent PCA-SVM modeling.

To improve the accessibility of concentration data for interdisciplinary readers, the rationale for selecting distinct concentration units (ppm, mol/L, and mmol/L) is clarified herein. The utilization of these units is scientifically justified based on the concentration characteristics of the target substances: specifically, ppm was adopted to quantify the trace-level concentration of the SERS substrate modifier, whereas mol/L and mmol/L were designated for determining the higher concentrations of LEV stock solutions and working solutions, respectively. This unit assignment strategy ensures that concentration values are presented in a format that is both numerically intuitive and contextually appropriate for each experimental component.

A 4 ppm Rhodamine 6G solution and 3 mol/L solutions of potassium iodide (KI), sodium chloride (NaCl), sodium sulfate (Na₂SO₄), and sodium bromide (NaBr) were prepared. Additionally, KI solutions with concentrations of 3 mol/L, 2 mol/L, 1.5 mol/L, 1 mol/L, 0.5 mol/L, and 0.1 mol/L were prepared.

2.4. SERS Measurements, Data Processing, and Analysis Were Performed

Background subtraction and baseline correction were executed using BWSPEC software, with baseline correction via a fifth-order polynomial fitting algorithm. The workflow was standardized: raw conventional Raman and SERS spectra were imported into BWSPEC, where each spectrum’s baseline was automatically recognized. A fifth-order polynomial was fitted to the baseline profile, and the fitted background was subtracted from the raw spectrum to eliminate nonspecific interference. This methodology preserved LEV’s characteristic Raman peaks, preventing the loss of critical spectral features indispensable for brand discrimination.

Subsequent to baseline correction, spectral smoothing and cosmic ray removal were sequentially performed in BWSPEC. Smoothing was achieved via the Savitzky-Golay method (5-point window size, second-order polynomial), suppressing random noise while maintaining peak shape integrity. Cosmic ray spikesaberrant signals that could undermine data reliabilitywere automatically detected and removed using BWSPEC’s built-in adaptive thresholding algorithm, safeguarding the spectral data set’s integrity.

Normalization was conducted in Origin 2024b using vector normalization, which scales all processed spectra to a consistent 0–1 range. This step effectively eliminated intensity variability from laser power fluctuations, sample dosage differences, and minor SERS substrate inconsistencies. Importantly, the identical preprocessing pipelineencompassing algorithms, parameter settings, and operational sequencewas applied to both data sets, eliminating biases from disparate protocols and ensuring valid comparative analyses between the two spectral modalities.

Postpreprocessing, the 500–2000 cm^–1 range was selected for each spectrum to construct the final data set. This range was chosen for encapsulating LEV’s key characteristic Raman peaks while excluding irrelevant spectral regions that might introduce noise or confounding signals.

3. Results and Discussion

3.1. Characterization of Silver Sol

The properties of the Ag sol substrate were characterized. Silver nanoparticles were synthesized using a previously reported method. The SEM image of the Ag sol, presented in Figure A, shows nearly spherical and uniformly dispersed particles with an average size of approximately 30–40 nm.

1.

Characterization of the properties of silver nanoparticles (A) SEM image of silver nanoparticles assembled on silicon wafers; (B) UV–vis absorption spectrum of silver nanoparticles; (C) SERS spectrum of silver sols detected using a 4 ppm R6G probe; (D) Particle size distribution of the modified silver sols.

SERS-active substrates are typically composed of nanoscale metal arrays that induce the polarization of surface plasmon excitons, significantly enhancing the Raman signals of analytes located on or near the substrate. , Furthermore, these substrates exhibit distinct characteristic peaks, which are associated with the molecular vibration modes. The ultraviolet–visible (UV–vis) absorption spectra of the silver nanoparticles were characterized by surface plasmon resonance (SPR). As shown in Figure (B), the absorption spectrum displays a single peak with a maximum at approximately 445 nm, suggesting that the nanoparticles are predominantly spherical and exhibit good monodispersity. This further suggests that they exhibit good monodispersity. To minimize the impact of fluorescence interference, a 785 nm light source was employed during the experiments.

To further assess the sensitivity of the silver colloid, the probe molecule R6G (4 ppm) was utilized to evaluate the detection performance of the silver colloid. The corresponding SERS spectrum of R6G, shown in Figure (C), demonstrates the significant enhancement and high sensitivity of the silver colloid. It can be concluded that the silver colloid exhibits an excellent SERS enhancement effect and demonstrates relatively high sensitivity. Statistical analysis of the particle size distribution, presented in Figure (D), reveals an average particle size of 35.9 nm.

3.2. Calculation of SERS Enhancement Factors

The Enhancement Factor (EF) is a definitive metric for quantifying the performance of SERS-active substrates. In this study, Rhodamine 6G (Rh6G) was utilized as the model probe, with 35.9 nm Ag colloid serving as the SERS substrate. All spectral measurements were conducted under 30 mW laser power, with meticulous attention paid to the consistency of instrumental parameters to ensure reliable intensity comparison.

A streamlined concentration-based formula was utilized for Analytical Enhancement Factor (AEF) calculation

$$\mathrm{AEF}=\frac{I_{\mathrm{SERS}}/c_{\mathrm{SERS}}}{I_{\mathrm{Raman}}/c_{\mathrm{Raman}}}$$

Here, I _SERS and I _Raman denote the baseline-corrected integrated intensities of the target R6G SERS and corresponding bulk Raman peaks, respectively. C _SERS corresponds to the 4 ppm R6G solution (converted to 8.35 × 10^–6 mol/L using R6G’s molar mass of 479.01 g/mol), while C _Raman = 28 mmol/L ((bulk R6G concentration)).

Substituting I = 1.5 × 10⁵ a.u.cm^–1 and I = 3.0 × 10³ a.u.cm^–1 into the formula yielded AEF ≈ 1.68 × 10⁵, which lies squarely within the typical SERS enhancement range (10³–10⁶), validating the Ag colloid’s robust SERS capability.

3.3. Attribution of Raman Characteristic Peaks of LEV

To validate the spectra of LEV, the molecular and three-dimensional (3D) spatial structures of LEV were first optimized using Molview, theoretical spectra were subsequently calculated using density functional theory (DFT) with the B3LYP/6–31G­(d,p) method from the Gaussian09 package (Gaussian, Inc.), as shown in Figure (A)–(C). Concurrently, the SERS spectra of an aqueous LEV solution at a concentration of 28 mmol/L were recorded, as shown in Figure (C). It is evident that the experimentally measured positions of the characteristic peaks exhibited slight shifts in frequency when compared to the theoretical values; however, the peak positions largely aligned. The DFT calculations were based on complexes consisting of molecular ligands and active sites on the metal surface, modeled as heteroatoms or heteroclusters with one or a few atoms, which were assumed to be nearly isolated on the metal surface However, the experimentally obtained Raman characteristic peaks were influenced by external factors, including solvent effects and interactions between the silver substrate and the molecules, resulting in measurable shifts. ,

2.

Structure and spectra of LEV. (A) Molecular structure of LEV. (B) Optimized 3D spatial structure of LEV. (C) Theoretical spectrum, Raman, and SERS spectrum of LEV aqueous solution.

Raman scattering is induced by distinct bending vibrational frequencies associated with various chemical bonds, enabling the identification of molecular structural features through the analysis of scattering peaks and vibrational modes. As summarized in Table , the primary vibrational peaks of LEV in the theoretical Raman spectrum correspond to N–H and C–H bending vibrations, C–C stretching vibrations, and C–H bending vibrations, with calculated frequency shifts at 734, 1014, 1053, and 1474 cm^–1, respectively. Furthermore, the characteristic Raman peaks of LEV appear at 706, 1014, 1043, and 1492 cm^–1, and the characteristic SERS peaks of LEV appear at 753, 1004, 1032, and 1443 cm^–1, which are attributed to the corresponding vibrational modes of the chemical bonds.

1. Calculated and Experimental Raman Frequencies Assigned to the Vibrational Modes of LEV.

Raman shift (cm–1)			Vibrational modes
Theory	Raman	SERS
734	706	753	N–H, C–H bending vibration
1014	1014	1004	C–C Telescopic vibration
1053	1043	1032	C–H bending vibration
1474	1492	1443	C–H bending vibration

3.4. Selection of Coagulants

The enhancement of SERS signals can be attributed to two primary mechanisms. First, the electromagnetic field enhancement mechanism, in which the vibration of conduction electrons amplifies the localized electromagnetic field, thereby significantly increasing the Raman scattering intensity of molecules in close proximity. Second, photoinduced charge transfer (CT) occurs between the substrate and adsorbed molecules, resulting in the formation of metal-molecule intermediates that significantly increase the polarizability of the molecules, thereby enhancing the Raman scattering intensity.

As an SERS substrate, silver sol exhibits SERS enhancement mainly through the synergistic effect of electromagnetic enhancement and chemical enhancement, and the adsorption-induced selection rules (AISR) are crucial for interpreting the intensity differences of characteristic peaks in SERS spectra. LEV molecules are physically or chemically adsorbed on the surface of silver nanoparticles through polar groups (e.g., amide groups) in their molecular structure, and this adsorption process can alter the electron cloud distribution and vibrational mode symmetry of the molecules. According to the adsorption-induced selection rules, after molecular adsorption, the vibrational modes that are forbidden or weakly intense in conventional Raman spectra may be activated due to the reduction of molecular symmetry and charge transfer between silver nanoparticles and molecules, which is manifested as a significant enhancement of characteristic peak intensity in SERS spectra. On the contrary, the Raman intensity of some vibrational modes perpendicular to the adsorption direction may be suppressed. In this study, the intensity differences of the characteristic SERS peaks of LEV are a direct reflection of the molecules following the adsorption-induced selection rules under the combined action of adsorption orientation and adsorption force on the surface of silver sol. This also further confirms the effective adsorption of LEV molecules on the silver sol substrate and the reliability of the SERS enhancement effect.

In this study, silver nanoparticles were utilized as the SERS substrates, with an inorganic salt solution employed as an intermediate to improve their adsorption capacity.

The introduction of inorganic salt ions to silver nanoparticles displaces citrate from the nanoparticle surface. As the salt ions exhibit a higher adsorption affinity for the silver surface. This results in the formation of chemical bonds between the ions and silver, disrupting the equilibrium of the sol–gel system and leading to the aggregation of the sol–gel particles, which enhances the metallosol properties. , Consequently, the reliability and accuracy of aggregated silver nanoparticles induced by inorganic salt ions in the detection of SERS is well established.

In this study, four inorganic saltsKNO₃, NaCl, NaBr, and KIwere selected as coagulant promoters for comparative analysis. The results, shown in Figure (A), indicate that the addition of a coagulant promoter significantly improved the silver sol. At a concentration of 3 mol/L, the KI solution exhibited the most pronounced enhancement among the four coagulants. This can be attributed to the distinct adsorption properties of the anions (NO₃ ^–, Cl^–, Br^–, and I^–). While the addition of a moderate amount of inorganic salts provides a beneficial enhancement effect, excessive salt concentrations lead to excessive aggregation of the silver sol particles, which is detrimental to the enhancement process. As illustrated in Figure (B), when the concentration of the KI solution was 3 mol/L, the relative intensity of the characteristic peak located at 1004 cm^–1 was 17761 au, indicating the optimal enhancement effect. Additionally, the impact of varying volume mixing ratios of the sample, silver sol, and coagulant promoter on the SERS enhancement was also investigated. The results demonstrated that the highest SERS enhancement was achieved when the volume ratio of silver sol, sample, and coagulant promoter was 2:1:1. Accordingly, a 3 mol/L KI aqueous solution was employed as the coagulant promoter, with a volume ratio of 2:1:1 for the silver sol, sample, and coagulant promoter, respectively, for each test. Following thorough mixing, 20 μL of each of the three liquids was then subjected to the SERS test.

3.

Selection of Optimal Coagulant Promoter Concentration (A) SERS spectra of LEV in different kinds of inorganic salt ions.(B) SERS spectra of LEV at different mass concentrations of KI solution.

3.5. Processing and Characterization of Three Types of LEV Tablets

To ensure the reliability of spectral data, the outer coating of LEV tablets was first removed to eliminate potential interference, with three tablets pretreated following the simplified protocol described by Shi et al., (Figure (A)). The pretreatment process involved sequential ethanol immersion for 3 min, residual ethanol removal via filter paper blotting, and careful scraping of the coating with a blade (B); the processed tablet was then crushed, flattened onto tin foil tape (C), after which Raman spectra were collected (20 scans per measurement). This procedure was repeated for the remaining two tablets, generating 60 Raman spectra in total.

4.

Experimental Procedure (A) Dissolution in ethanol within a beaker. (B) Removal of the film-coated. (C) Flattening the tablet on tin foil tape. (D) Mixing of silver sol, sample, and 3 mol/L KI solution. (E) Real-time observation of the testing process.

For SERS analysis, 100 ppm solutions of the three tablet formulations were prepared by water dissolution and 5 min sonication to ensure completeness. The sample solution, silver sol, and coagulation promoter were mixed at a 2:1:1 volume ratio (D), and 20 SERS scans were recorded per sample after depositing the mixture onto tin foil tape (E). Comparative analysis of the SERS spectra of the three LEV samples against the theoretical spectrum confirmed consistent characteristic peaks, further validating that the model specifically targets LEV-derived signals. While trace residual excipients or coating-related effects may persist postethanol pretreatment, these factors do not significantly interfere with the model’s core differentiation capability. Given the study’s focus on LEV-specific recognition, additional verification of excipients’ impact on classification outcomes is unwarranted, ensuring the model’s robust and precise performance for LEV differentiation.

3.6. SERS Spectra of LEV at Variable Concentrations

To assess the concentration-dependent behavior and analytical sensitivity of our SERS platform for LEV, we systematically acquired spectra from a single pharmaceutical formulation diluted across a broad concentration range (100 ppm to 0.1 ppm). It is noteworthy that while absolute peak intensities diminished in a predictable manner as concentration decreased, the relative intensity ratios of diagnostic fingerprint bands, notably those at 759 cm^–1 and 1442 cm^–1, exhibited remarkable conservation across the entire dynamic range. This finding serves to substantiate the efficacy of the methodology employed, which has been demonstrated to accurately capture the intrinsic molecular signature of LEV, thereby negating the potential for concentration-induced spectral artifacts to compromise the analysis. The consistent spectral pattern, coupled with a monotonic intensity-concentration relationship, validates the robustness of the approach for qualitative identification (Figure ).

5.

SERS spectra of LEV aqueous solutions at various concentrations.

3.7. Repeatability Assessment of LEV

To evaluate the reproducibility of the proposed method, 15 replicate measurements were performed on an aqueous solution of LEV at a concentration of 100 ppm. The solution was prepared by accurately weighing and dissolving the appropriate amount of LEV in deionized water. All measurements were conducted under consistent experimental conditions. The results of these measurements are presented in Figure (A), which illustrates the typical SERS spectra obtained from the LEV solution. A total of 15 data sets were analyzed, and the relative standard deviation (RSD), defined as the ratio of the standard deviation to the mean of the spectral peak intensities, was calculated. The RSD was determined to be 6.4% (Figure (B)), indicating minimal dispersion in the SERS spectra and suggesting high consistency among the replicate measurements. The experimental results confirm that the SERS technique provides high reproducibility for the analysis of LEV at known concentrations. Consequently, the method is stable, reliable, and well-suited for accurate and reproducible analysis of LEV in aqueous solution at known concentrations, making it a promising tool for routine pharmaceutical analysis.

6.

Stability Assessment of LEV Aqueous Solution­(A) Stability determination of LEV aqueous solution across 15 batches.(B) Distribution of characteristic peak intensities at 753 cm^–1 for 15 batches of LEV aqueous solution.

3.8. Measurement of Conventional SERS of Three Types of LEV

As shown in Figure (A), the conventional Raman spectra of the three LEV formulations exhibit characteristic peaks at 706, 748, 856, 1015, 1215, and 1492 cm^–1. These spectra are highly similar, making differentiation between drugs based on visual inspection alone difficult. In Figure (B), the SERS spectra of JiYiKe (tablet 1) display characteristic peaks at 762, 820, 1006, 1028, 1238, and 1445 cm^–1, with the peak at 1445 cm^–1 showing the highest intensity. For PuLuo (tablet 2), the SERS peaks are observed at 810, 860, 1239, and 1446 cm^–1, with a notable increase in intensity at 860 cm^–1. Keppra (tablet 3) shows peaks at 808, 859, 1241, and 1448 cm^–1, with the peak at 859 cm^–1 displaying the most pronounced intensity. The SERS spectra of all three formulations share common peaks at 859, 1239, and 1446 cm^–1, with variations in intensity and position across formulations. This indicates the SERS technique’s ability to enhance spectral differences between samples. However, the observed variations were too subtle to differentiate reliably by visual inspection alone. To address this limitation, multivariate analysis using the PCA-SVM method was employed to improve classification accuracy.

7.

Comparison of Raman Spectra of Three LEV Formulations (A) Conventional Raman Spectra of three LEV formulations. (B) SERS of Three LEV formulations with Silver Sol.

3.9. Qualitative Identification of Three LEV Formulations Using Principal Component Analysis (PCA) and Support Vector Machine (SVM)

Principal component analysis (PCA) is a widely used technique for dimensionality reduction in data analysis. Its simplicity and effectiveness have made it a prevalent tool across various fields of application. − As a machine learning algorithm, PCA efficiently processes high-dimensional data, extracts key features, and projects the data into a lower-dimensional space for simplification and visualization. This approach retains essential information while eliminating noise and redundant details. The Support Vector Machine (SVM), a supervised learning algorithm, functions by mapping linearly inseparable points into a higher-dimensional space to define a discriminative boundary. By applying linear classification within this transformed feature space, SVM can effectively solve highly nonlinear classification and regression problems. This ability makes SVM particularly suitable for distinguishing complex sample content. Its simplicity and effectiveness have made it a prevalent tool across various fields of application. − As a machine learning algorithm, PCA efficiently processes high-dimensional data, extracts key features, and projects the data into a lower-dimensional space for simplification and visualization. This approach retains essential information while eliminating noise and redundant details. The Support Vector Machine (SVM), a supervised learning algorithm, functions by mapping linearly inseparable points into a higher-dimensional space to define a discriminative boundary. By applying linear classification within this transformed feature space, the SVM can effectively solve highly nonlinear classification and regression problems. This ability makes SVM particularly suitable for distinguishing complex sample content. Furthermore, SVM demonstrates considerable advantages in small-sample, nonlinear, and high-dimensional pattern recognition tasks. ,

As shown in Figure (A),(B), the Raman and SERS spectra of the three LEV formulations exhibit a significant degree of similarity, which poses a challenge to manual identification and increases the risk of subjective error. To address this limitation, a PCA-SVM approach was employed for automated identification and qualitative classification. The initial step in the analysis was to apply PCA to the data set, with the objective of reducing dimensionality and enhancing the classification accuracy. For this study, 60 conventional Raman spectra and 60 SERS spectra were selected for the training and testing sets. The Raman data set was segmented into a training set comprising 45 spectra and a testing set consisting of 15 spectra. Similarly, in the SERS data set, 45 spectra were used for training, and 15 for testing. Subsequently, the SVM algorithm delineated the decision boundaries, thereby effectively distinguishing the data points into discrete categories.

In this paper, the data were analyzed using PyCharm Community Edition 2021, resulting in Figure , Table , and Table . The percentage associated with each principal component indicates its contribution to the overall variance in the data set, reflecting the amount of information retained during dimensionality reduction. A higher percentage signifies that the corresponding principal component retains more of the original data’s variability. This is particularly important in high-dimensional data sets, where dimensionality reduction facilitates more efficient processing and analysis without significant loss of information. PCA identifies and extracts the most significant sources of variation within the data set, focusing on the directions that explain the highest variance.

8.

PCA-SVM model was employed to classify different brands of LEV­(Regions with different background colors are generated by the Support Vector Classification algorithm; points of varying colors represent different sample types: circular = training set, triangular = test set). (A) PCA-SVM scatterplot of the conventional Raman spectral data set of three LEV formulations. (B) PCA-SVM scatterplot of the SERS spectral data set of three LEV formulations.

2. Accuracy, Precision and Sensitivity of the PCA-SVM Model in Classifying Three Types of LEV Tablets Based on Conventional Raman Spectra.

Parameter	Tablet1	Tablet2	Tablet3
Accuracy	100%	100%	100%
Precision	100%	100%	100%
Recall	100%	100%	100%

3. Accuracy, Precision and Sensitivity of the PCA-SVM Model in Classifying Three Types of LEV Tablets Based on SERS Spectra.

Parameter	Tablet1	Tablet2	Tablet3
Accuracy	100%	100%	100%
Precision	100%	100%	100%
Recall	100%	100%	100%

Regarding the use of PC1 and PC2 for data visualization, these two principal components were selected because they capture the most chemically meaningful sources of variation among samples and account for the majority of variance in each data set. As shown in Figure , for the conventional Raman data set, PC1 and PC2 contribute 56.8% and 8.4% of the variance, respectively, yielding a cumulative value of 65.2%. For the SERS data set, PC1 and PC2 account for 48.9% and 9.1% of the variance, respectively, giving a total of 58%. Although these cumulative variance values are relatively modest, the two-dimensional score plots clearly illustrate distinct clustering and separation of the three LEV formulations. This level of variance retention is sufficient to encapsulate the essential information from the original spectra, and the successful classification with 100% test accuracy confirms that PC1 and PC2 provide adequate discriminatory power for the present analysis.

As presented in Table , the PCA-SVM algorithm demonstrated 100% accuracy, precision and sensitivity for the three categories of LEV ordinary Raman spectra, demonstrating the exceptional classification performance of the PCA-SVM model. These results underscore the model’s capability to differentiate between the formulations with flawless precision and reliability. Table shows that the PCA-SVM algorithm was used to evaluate the accuracy, precision and sensitivity of three LEV SERS spectra, achieving 100% accuracy. This confirms the efficacy of the PCA-SVM model in identifying and classifying these samples. The findings highlight the robustness and reliability of the PCA-SVM approach, which consistently delivers high performance across both conventional Raman and SERS spectral data sets.

In order to evaluate the classification performance of the constructed PCA-SVM model for the Raman and SERS spectra of the three LEV drugs, 10-fold stratified cross-validation was employed to test 60 Raman spectra (20 per drug class, 1500 Raman shift variables) and 60 SERS spectra (20 per drug class, 1500 Raman shift variables). The findings indicated that the accuracy, precision and recall rate for all folds of both Raman and SERS spectra attained 1.0000. This finding suggests that the model displays perfect and fully reproducible classification capability on the current data set, thereby demonstrating exceptional robustness.

In conclusion, the PCA-SVM model achieved 100% accuracy in both the identification and qualitative classification of LEV across various tablet formulations. This result underscores the model’s exceptional precision and efficiency, making it well-suited for rapid, on-site detection in clinical and pharmaceutical environments. Its capacity to differentiate between formulations without relying on complex instrumentation highlights the practicality and robustness of the PCA-SVM model in real-world applications. Furthermore, the integration of SERS with PCA-SVM enhances its sensitivity, allowing for the detection of subtle variations in the chemical composition of LEV tablets. This combination provides a fast, highly sensitive method for guiding personalized medication management in epilepsy patients, enabling more accurate and timely treatment decisions. The incorporation of SERS also facilitates the detection of impurities or counterfeit drugs, thus extending the method’s utility from therapeutic applications to include drug safety and quality assurance.

While the established SERS coupled PCA-SVM strategy achieves satisfactory classification accuracy for discriminating levetiracetam brands, this study still contains several inherent limitations that deserve objective clarification. The analysis in this work is performed based on a limited number of samples with only one tablet selected to represent each brand and relevant verification of different batch samples is not included in the current experimental design. Although multiple spectral collections have been implemented at different detection positions, the lack of additional tablet samples and cross batch validation may impose certain constraints on the statistical reliability and practical extrapolation of the constructed model. Moreover, this study does not set up an independent validation set to further evaluate model performance, and the data splitting strategy for training and testing sets is only implemented under the existing data set framework, which leaves certain room for improving the generalizability of the method. In addition, this work mainly focuses on the application feasibility of the existing analytical scheme and does not carry out in depth comparative discussion with alternative classification methods under the current research scope. We acknowledge the constrained coverage of the adopted data set and recognize the potential limitations in the generalization capability of the proposed model, and such objective conditions should be taken into account when extending the present findings to more realistic application scenarios. Future research will concentrate on expanding sample size, introducing multi batch specimens, supplementing independent validation data sets, and conducting comparative analysis with other classical classification algorithms.

4. Conclusion

This study presents a rapid and reliable qualitative method for brand identification of different LEV formulations by combining SERS, conventional Raman spectroscopy, and machine learning. Silver sol was employed as the SERS substrate and systematically optimized. Theoretical Raman calculations were performed to support the assignment of LEV characteristic peaks. The reproducibility of the method was verified with a relative standard deviation (RSD) of 6.4% based on 15 replicate measurements. The PCA-SVM classification model achieved 100% accuracy, precision and sensitivity for both Raman and SERS data sets, enabling clear and robust discrimination among the three LEV brands. These results demonstrate that the combination of SERS and machine learning is a feasible and effective strategy for the brand-level identification of LEV tablets. This method provides a sensitive, rapid, and convenient analytical tool with promising application prospects for the rapid screening, quality monitoring, and classification of pharmaceutical products. It can serve as a complementary approach to support stable medication use and drug safety management in clinical and regulatory settings.

This study also acknowledges the limited scope of the current data set and recognizes the potential limitations regarding the generalization ability of the developed model. The findings are derived from a relatively constrained sample framework, which may restrict the broader extrapolation of the classification performance to more diverse practical scenarios. Future work will expand the data set coverage and further validate the model to enhance its generalizability and practical applicability.

All core data supporting the findings of this study are presented within the manuscript and Supporting Information.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c12399.

• SEM micrographs of silver nanoparticles at increasing magnifications with scale bars of 1 μm, 500, 200, and 50 nm; size distribution histogram of silver nanoparticles based on TEM imaging and ImageJ analysis, showing a number-average diameter of 35.9 nm (PDF)

H.Z.: Investigation, writingoriginal draft. W.L.: Methodology, writingreview and editing. N.L.: Formal analysis, editing. W.H.: Formal analysis, resources.

The authors declare no competing financial interest.