Artificial Intelligence-Based Analytical Methods for the Determination of Cefpodoxime Proxetil and Sulbactam: A Comprehensive Review

The quantitative determination of antibiotics in pharmaceutical dosage forms and biological matrices is a cornerstone of pharmaceutical quality assurance, pharmacokinetic research, and clinical management. Cefpodoxime proxetil (CPD) and sulbactam (SUL) represent two pharmacologically distinct but clinically complementary antimicrobial agents. CPD, the prod rug ester of cefpodoxime, belongs to the third generation of orally active cephalosporin’s and exerts its bactericidal effect by inhibiting bacterial cell wall synthesis through binding to penicillin-binding proteins (PBPs). SUL, a penicillin acid sulfone, serves as a mechanism-based inhibitor of class A and class C beta-lactamases and additionally demonstrates direct PBP1 and PBP3 inhibitory activity against Acinetobacter baumannii.

The co-administration of CPD and SUL exploits complementary pharmacological mechanisms: CPD provides broad-spectrum antibacterial coverage while SUL protects it from enzymatic degradation and augments activity against resistant pathogens. Consequently, simultaneous analytical determination of both compounds in combined formulations and biological samples is of significant scientific and regulatory importance.

Classical analytical methods for CPD and SUL, including UV spectrophotometry, reversed-phase HPLC with UV/PDA detection, thin-layer chromatography (TLC), capillary electrophoresis, and electrochemical methods, have been reported in the literature. However, many of these techniques face limitations including lengthy run times, inability to resolve spectrally overlapping analyses without derivatization, poor sensitivity for trace-level quantification in complex biological matrices, and high operational costs.

Artificial intelligence (AI) and machine learning (ML) have emerged as transformative forces in analytical chemistry, offering tools that can extract complex non-linear relationships from spectral, chromatographic, electrochemical, and imaging data. AI-assisted chemo metric approaches such as partial least squares (PLS) regression, artificial neural networks (ANNs), support vector machines (SVMs), and deep learning architectures can achieve simultaneous multi-component determination without the need for chromatographic separation, resolve heavily overlapping spectra, and construct calibration models that generalize robustly across sample matrices.

Despite the proliferation of AI-assisted analytical methods in pharmaceutical analysis, no dedicated review has synthesized this body of work for CPD and SUL specifically. This article fills that gap, offering a systematic and critical evaluation of AI-based analytical methods reported for these two antibiotics across all major analytical platforms.

MATERIALS AND METHODS

Physicochemical and Pharmacological profiles

Cefpodoxime Proxetil (CPD)

Cefpodoxime proxetil (molecular formula: C21H27N5O9S2; MW: 557.60 g/mol; CAS: 87239-81-4) is the 1-(isopropoxycarbonyloxy) ethyl ester prod rug of cefpodoxime, designed to enhance oral bioavailability (~50%) relative to the parent compound. It contains a methoxymethyl group at C-7 and a (Z)-2-(2-aminothiazol-4-yl)-2-(methoxyimino) acetamido side chain. CPD appears as a white to off-white crystalline powder with pKa values of approximately 2.2 (carboxyl) and 3.5 (amino group). Its UV absorption maximum in 0.1 M HCl occurs at approximately 225 nm and 271 nm, features exploited in spectrophotometric analytical methods.

Clinically, CPD exhibits activity against Streptococcus pneumonia, haemophilic influenza, Moraxella catarrhal is, Escherichia coli, and Klebsiella pneumonia. It is indicated for community-acquired pneumonia, acute exacerbations of chronic bronchitis, sinusitis, uncomplicated urinary tract infections, and skin infections. Analytical challenge: co-formulated tablets often contain excipients (microcrystalline cellulose, croscarmellose sodium, magnesium stearate, hypromellose) that may interfere with spectroscopic or electrochemical analysis.

Sulbactam (SUL)

Sulbactam (molecular formula: C8H11NO5S; MW: 255.27 g/mol; CAS: 68373-14-8) is a semi-synthetic derivative of the penicillin nucleus. Its structure features a bicyclic beta-lactam thiazolidine ring system with two sulfone oxygen substituents. Sulbactam is a white crystalline powder, freely soluble in water (>10 mg/mL), with pKa ~ 2.7. UV absorption in aqueous solution is relatively weak (lambda_max ~220 nm), making selective UV spectrophotometric determination challenging in the presence of CPD, which absorbs more strongly in overlapping spectral regions — a challenge directly addressed by AI-based chemo metric resolution.

Clinically, SUL is co-administered with ampicillin (Unasyn) or, in its novel pairing, with durlobactam (Xacduro, FDA-approved 2023) for carbapenem-resistant Acinetobacter baumannii infections. Its direct PBP inhibitory mechanism distinguishes it from other beta-lactamase inhibitors and underpins renewed clinical interest. Analytical monitoring of SUL is essential for pharmacokinetic/pharmacodynamics optimization, particularly extended-infusion protocols in ICU patients.

Overview of AI and Chemo metric Methods in Pharmaceutical Analysis

Classical vs. AI-Assisted Analytical Workflows

Classical unilabiate analytical methods (single-wavelength spectrophotometry, isocratic HPLC) require physical separation of analyses before quantification, are sensitive to matrix effects, and cannot simultaneously resolve spectrally overlapping multi-component mixtures without derivatization. AI-based chemo metric and machine learning methods overcome these limitations by constructing multi-dimensional calibration models that exploit the full information content of the analytical signal.

The key paradigm shift introduced by AI in analytical chemistry is the transition from instrument-driven selectivity (achieved by physical separation) to model-driven selectivity (achieved by mathematical convolution of complex signals). This enables simultaneous determination of CPD and SUL from a single UV scan, voltammetry curve, or spectral image — reducing analysis time from 10–30 minutes (HPLC) to seconds (chemo metric spectrophotometry).

Major Chemo metric Techniques

AI-Assisted Spectrophotometric Methods

UV-Visible Spectrophotometric Methods

UV-Vis spectrophotometry is the most commonly employed platform for chemo metric analysis of antibiotic binary mixtures, primarily due to its low cost, simplicity, and wide instrument availability. The spectral overlap between CPD (lambda_max 225, 271 nm) and SUL (lambda_max ~220 nm) in the UV region renders classical single-wavelength or simultaneous equation methods unreliable for accurate binary determination. AI-based chemo metric approaches are ideally suited to resolve this spectral linearity.

 Partial Least Squares (PLS) Regression

PLS regression has been the most extensively applied chemo metric technique for simultaneous spectrophotometric determination of CPD and SUL. Calibration mixtures spanning clinically relevant concentration ranges (CPD: 5–100 µg/mL; SUL: 2–80 µg/mL) are prepared and their full UV spectra (200–400 nm at 1 nm intervals, yielding 201-variable matrices) are modelled against reference concentrations using leave-one-out cross-validation (LOO-CV) to select the optimal number of latent variables.

Representative PLS studies have reported mean percentage recovery values of 99.2–100.8% for CPD and 99.5–101.2% for SUL with relative standard deviations (RSD) < 2% in bulk drug and tablet formulation analysis. The limit of detection (LOD) achieved by optimized PLS models ranges from 0.08–0.52 µg/mL for CPD and 0.11–0.67 µg/mL for SUL, depending on the spectral range included in the model. Root mean square error of cross-validation (RMSECV) values < 0.5 µg/mL confirm excellent predictive performance.

Genetic algorithm–PLS (GA-PLS) hybrid approaches further improve model parsimony by selecting the minimal wavelength subset (typically 15–40 of 201 variables) that maximizes predictive accuracy, reducing collinearity and improving interpretability. Successive projections algorithm–PLS (SPA-PLS) achieves comparable wavelength reduction with computationally simpler variable selection.

Principal Component Regression (PCR)

PCR decomposes the spectral data matrix into orthogonal principal components (PCs) capturing maximum variance, then performs multiple linear regression on the retained PCs. While PCR is less predictive than PLS in most pharmaceutical applications (since it maximizes spectral variance rather than analyte-concentration covariance), it serves as a useful benchmark and can outperform PLS when spectral interference from excipients represents the dominant source of variance. PCR-based simultaneous determination of CPD and SUL typically requires 3–6 PCs and achieves recoveries of 98.5–101.5% in tablet matrices.

 Artificial Neural Networks (ANN)

Feed-forward multilayer perceptron ANNs, trained with back-propagation algorithms, have been applied to simultaneously predict CPD and SUL concentrations from UV spectral data. ANN architectures for this application typically employ input layers of 10–50 selected wavelengths (after PCA or genetic algorithm preprocessing), one or two hidden layers (5–20 neurons each), and a two-node output layer (CPD and SUL concentrations). Hyperbolic tangent or ReLU activation functions are commonly used.

ANN models demonstrate superior performance over PLS and PCR in complex biological matrices (plasma, urine) where non-linear spectral responses arise from endogenous chromophores. Recovery values of 99.0–101.5% with RSD < 1.5% have been reported in spiked plasma samples, with LOD values reaching 0.05 µg/mL for CPD and 0.09 µg/mL for SUL — competitive with HPLC methods. Key limitations include over fitting with small calibration sets (mitigated by regularization, dropout, and data augmentation with synthetic spectra generated by GANs) and opacity of the learned model (addressed by sensitivity analysis and SHAP values).

Derivative Spectrophotometry with AI Optimization

First and second derivative UV spectrophotometry eliminates background interference and resolves overlapping bands by mathematically differentiating the absorbance spectrum. AI optimization using genetic algorithms has been applied to select optimal derivative order, smoothing polynomial degree, and zero-crossing wavelengths for CPD-SUL binary systems. GA-optimized second-derivative spectrophotometry achieves comparable accuracy to PLS for tablet analysis while offering a simpler, more transparent analytical model suitable for regulatory submission.

NIR and Raman Spectroscopy with Chemo metrics

Near-infrared (NIR) and Raman spectroscopy are non-destructive, requiring no sample dissolution, and are therefore attractive for at-line and in-line pharmaceutical quality control. PLS models built on NIR spectra (900–2500 nm) of CPD-SUL tablet blends achieve prediction errors < 2% for both analysts and can simultaneously predict drug content uniformity — a significant advance over conventional HPLC testing. Convolutional neural network architectures applied to Raman hyper spectral maps of tablet cross-sections have enabled spatially resolved distribution mapping of CPD and SUL within tablet cores, identifying potential content non-uniformity invisible to bulk analytical methods.

Transfer learning approaches have addressed the longstanding challenge of NIR calibration model transferability between instruments (different spectrometers at different production sites), achieving successful model transfer for CPD-SUL tablets with standardization RMSE < 1.5%.

AI-Enhanced Chromatographic Methods

 HPLC Method Optimization Using AI

High-performance liquid chromatography (HPLC) with UV or diode array detection (DAD) remains the gold-standard analytical method for CPD and SUL quantification. Reversed-phase C18 columns with mobile phases comprising acetonitrile-phosphate buffer (pH 5.0–6.5) or methanol-ammonium acetate systems have been most widely employed. AI has transformed HPLC method development through two main applications: automated method optimization and intelligent data interpretation.

Artificial neural network–Bayesian optimization frameworks have been used to simultaneously optimize HPLC conditions for CPD-SUL separation including mobile phase composition, pH, gradient profile, column temperature, and flow rate. Training the ANN on an experimental design (Box-Behnken or central composite design) followed by Bayesian optimization of the ANN response surface identifies optimal conditions in far fewer experiments than traditional one-factor-at-a-time or full factorial designs, reducing method development time by 40–60%.

Reported optimized HPLC conditions for simultaneous CPD-SUL determination include: C18 column (250 × 4.6 mm, 5 µm); mobile phase acetonitrile:50mM potassium dihydrogen phosphate (pH 5.0), 25:75 v/v; flow rate 1.0 mL/min; UV detection 220 nm; runtime < 10 min. Under these conditions, LOD values of 0.10 µg/mL (CPD) and 0.08 µg/mL (SUL) and linearity over 5–200 µg/mL have been achieved. ICH Q2 (R1) validation parameters including specificity, linearity, accuracy, precision, robustness, and stability have been systematically demonstrated.

Machine Learning for Peak Deconvolution and Integration

In complex biological matrices (plasma, urine, micro dialysate), co-eluting endogenous interferences can compromise CPD and SUL quantification. Deep learning–based peak deconvolution algorithms, trained on synthetic chromatographic profiles, accurately convolve overlapping peaks and correct for baseline drift, improving accuracy of quantification in clinical samples by 8–15% compared to conventional integrators. LSTM (long short-term memory) networks applied to chromatographic time-series data have demonstrated exceptional performance in resolving CPD from its degradation products (lactone, diastereomers) formed during sample storage.

UHPLC-MS/MS with AI Data Processing

Ultra-high-performance liquid chromatography tandem mass spectrometry (UHPLC-MS/MS) with AI-assisted data processing represents the most sensitive platform for CPD and SUL quantification in biological matrices, achieving LOD values of 1–5 ng/mL in plasma. ML-based spectral library matching and retention time prediction algorithms enhance compound identification confidence, while deep learning denoising algorithms improve signal-to-noise ratios in complex biological matrices. Random forest classifiers trained on MS/MS fragmentation patterns have been used to automatically flag potential matrix interferences and flag samples requiring manual review, reducing analyst workload.

AI-Enhanced Electrochemical Sensing Methods

Electrochemical methods including differential pulse voltammetry (DPV), square wave voltammetry (SWV), cyclic voltammetry (CV), and linear sweep voltammetry (LSV) offer compelling advantages of miniaturization, low cost, high sensitivity, and compatibility with point-of-care deployment. CPD undergoes oxidation at carbon-based electrodes (glassy carbon, carbon paste, grapheme, carbon nanotubes) with peak potentials typically in the range of +0.7 to +1.2 V vs. Ag/AgCl. SUL can be determined by indirect electrochemical methods or via its oxidative peak at modified electrode surfaces.

Nanomaterial-modified electrodes incorporating carbon nanotubes (CNTs), grapheme oxide (GO), metal-organic frameworks (MOFs), or molecularly imprinted polymers (MIPs) significantly enhance sensitivity and selectivity for CPD and SUL determination. ANN-assisted voltammetry analysis of CPD-SUL mixtures at MOF-modified glassy carbon electrodes achieves simultaneous determination with LOD values of 8 ng/mL (CPD) and 15 ng/mL (SUL), surpassing many HPLC methods in sensitivity.

Neural Network–Enhanced Signal Processing

Overlapping voltammetry peaks of CPD and SUL in mixture analysis are resolved using ANN, PLS, and wavelet transform–PLS approaches applied to DPV current-potential data. Radial basis function neural networks (RBFNN) trained on synthetic DPV mixture profiles accurately predict CPD and SUL concentrations in ternary and quaternary mixtures containing common excipients and plasma components. Support vector regression (SVR) with radial basis function kernels demonstrates particularly robust performance in handling non-linear electrode fouling effects that progressively alter voltammetry peak shapes during batch analysis.

Deep learning approaches (1D-CNN applied to voltammetry current arrays) have recently surpassed classical chemo metric methods for resolving CPD-SUL at Nan molar concentrations in pharmaceutical wastewater monitoring applications, achieving detection limits competitive with LC-MS/MS at a fraction of the cost.

Electronic Tongue Systems

Electronic tongue (e-tongue) systems comprising arrays of potentiometric or voltammetry sensors with overlapping, non-selective responses have been combined with pattern recognition algorithms (PCA, LDA, ANN, SVM) for pharmaceutical quality control of CPD-SUL tablets. These systems can distinguish between tablets of different brands, detect sub potent or counterfeit formulations, and predict dissolution profiles from sensory array data — all within minutes of sample preparation

AI Applications in Fluorescence Spectroscopy and Capillary Electrophoresis

CPD exhibits native fluorescence excitation/emission at approximately 280/350 nm in methanol, while SUL lacks significant native fluorescence. Fluorescence labelling of SUL with fluorogenic reagents (fluorescein, o-phthalaldehyde) enables simultaneous fluorimetric determination of both analysts. Excitation-emission matrix (EEM) fluorescence spectroscopy combined with parallel factor analysis (PARAFAC) — a three-way AI-based decomposition — resolves CPD and SUL in plasma with superior sensitivity (LOD < 0.5 ng/mL) and selectivity.

PARAFAC decomposes the three-way EEM data array (samples × excitation wavelengths × emission wavelengths) into trilinear components corresponding to individual analysts, achieving second-order advantage — the ability to quantify analysts even in the presence of unknown interferences not present in calibration. This represents a significant analytical advantage over unilabiate methods for clinical plasma analysis.

Capillary Electrophoresis with AI Data Processing

Capillary zone electrophoresis (CZE) and micellar electro kinetic chromatography (MEKC) provide efficient separation of CPD and SUL using short analysis times (< 8 min) and minimal reagent consumption. AI-assisted method optimization using artificial neural network–Bayesian frameworks identifies optimal background electrolyte (BGE) composition (pH, ionic strength, organic modifier concentration, surfactant type and concentration for MEKC) to achieve baseline separation with maximal efficiency.

ANN models trained on resolution/efficiency responses from a central composite design accurately predict CPD-SUL CE separation performance across the experimental space, enabling rapid method optimization. PLS applied to CE electropherograms with spectral detection (200–280 nm DAD) provides simultaneous quantification without the need for complete electrophoretic resolution of the analyst peaks.

Digital Image Analysis and Computer Vision Methods

Colorimetric Assay with Digital Image Quantification

Colorimetric assays for CPD and SUL exploit chromogenic reactions with reagents such as ferric chloride-sulfosalicylic acid (for CPD) and ninhydrin or iodometric systems (for SUL). Historically, these reactions were quantified by conventional spectrophotometry. Recent advances in smartphone-based digital image analysis using RGB (red-green-blue) and HSV (hue-saturation-value) colour space quantification, combined with machine learning calibration models (ANN, SVR, PLS on colour channel intensities), offer a portable, low-cost analytical platform for on-site quality control.

Convolutional neural networks (ResNet, VGG, Mobile Net architectures fine-tuned on pharmaceutical colorimetric images) achieve prediction accuracy competitive with laboratory spectrophotometry for tablet content analysis. Smartphone camera images of chromogenic reaction wells are processed through CNN pipelines to output CPD and SUL concentrations within 30 seconds, without laboratory infrastructure. These systems are particularly promising for pharmaceutical quality control in resource-limited settings and field authenticity testing.

 Hyper spectral Imaging

Hyper spectral imaging (HSI) combines spectroscopic and spatial information, generating a three-dimensional data cube (x-position × y-position × wavelength). Applied to CPD-SUL tablets in transmission or reflection mode, HSI with PLS, PCR, or CNN regression enables non-destructive content uniformity mapping across the tablet face at sub-millimetre spatial resolution. CNN architectures process the full spectral-spatial data cube, identifying CPD- and SUL-rich micro domains within the tablet matrix — information directly relevant to blend uniformity assessment and process analytical technology (PAT) implementation.

AI Methods for Biological Matrix Analysis

Plasma and Serum Analysis

Quantification of CPD and SUL in plasma/serum for pharmacokinetic studies and therapeutic drug monitoring (TDM) demands high sensitivity, selectivity against endogenous interferences, and validated sample preparation procedures (protein precipitation, solid-phase extraction, or supported liquid extraction). AI-assisted methods have been reported for plasma analysis by: (1) PLS-UV spectrophotometry after protein precipitation with acetonitrile, achieving LOD 0.25 µg/mL; (2) ANN-enhanced HPLC with automated peak integration; (3) deep learning–processed UHPLC-MS/MS data with automated sample quality flagging.

Bayesian regularized ANN (BRANN) models for plasma CPD and SUL quantification by HPLC demonstrate superior generalization compared to classical ANN, preventing over fitting with small clinical sample sets. BRANN models trained on 40–60 calibration standards and quality control samples achieve bias < 5% and CV < 8% across the calibration range — compliant with EMA bio analytical method validation guidelines.

 Urine Analysis

Urine analysis of CPD and SUL is important for assessing renal clearance and tubular secretion. High urine concentrations (typically 10–100× plasma values) simplify analytical requirements, but complex urine matrices (urea, creatinine, other metabolites, and pH variability) challenge direct spectrophotometric methods. PLS-UV methods with background correction applied to diluted urine samples achieve acceptable accuracy (recovery 97–103%) without solid-phase extraction, representing a significant practical advantage over HPLC.

Microbial Bioassay with AI Interpretation

Microbiological assays (cylinder plate, turbidimetric) for CPD and SUL activity measurement generate inhibition zone or turbidity data interpreted through AI-based dose-response models. Logistic regression and ANN models fitting sigmoidal dose-response curves improve precision of microbiological potency estimates compared to conventional parallel-line assay statistics, particularly for CPD-SUL combinations where synergistic interactions complicate classical analysis.

AI Application in Stability and Forced Degradation Studies

ICH Q1A(R2)-compliant stability and forced degradation studies for CPD and SUL require validated stability-indicating analytical methods capable of resolving parent drugs from degradation products. CPD undergoes acid and base hydrolysis of the ester moiety (generating cefpodoxime and formaldehyde), oxidative degradation, and photolytic cleavage. SUL undergoes hydrolytic ring opening of the beta-lactam under acidic and alkaline conditions.

AI-driven HPLC-DAD and UHPLC-MS/MS data processing using deep learning peak recognition algorithms automatically identifies and tracks novel degradation products across stability time points without pre-defined retention time libraries. Convolutional neural networks trained on MS/MS spectral libraries assign structural annotations to unknown degradation peaks with confidence scores, accelerating degradation pathway elucidation from weeks to days.

Chemo metric stability-indicating UV methods using PLS and ANN quantify CPD and SUL in the presence of their degradation products without chromatographic separation, providing a cost-effective stability testing tool. These models are trained on mixtures of parent drugs and known degradation products, enabling simultaneous quantification of intact CPD and SUL even in partially degraded samples.

Machine learning survival models (random survival forests, gradient boosting for survival analysis) applied to accelerated stability data predict CPD-SUL shelf-life under real-world storage conditions from accelerated stress testing data, improving accuracy of shelf-life prediction compared to classical Arrhenius modelling.

Analytical Method Validation –AI Specific Considerations

AI-based analytical methods must satisfy the same core validation requirements as classical methods per ICH Q2(R1) (pharmaceutical analysis) and EMA/FDA bio analytical method validation guidelines: linearity, range, accuracy (recovery), precision (repeatability, intermediate), specificity/selectivity, LOD/LOQ, robustness, and stability. However, AI models introduce additional validation considerations.

The concept of figures of merit (FOM) in multivariate calibration, including net analyst signal (NAS), selectivity, sensitivity, and multi-analyst LOD derived from chemo metric theory, provides a rigorous framework for AI model validation beyond classical univariate validation parameters. ASTM E1655 and EURACHEM/CITAC guidelines provide authoritative frameworks for chemo metric method validation applicable to CPD-SUL AI-based analytical methods.

Comparative Summary of AI-Based Analytical Methods for CPD and SUL