A team of Brazilian researchers has developed a multiplatform chemical fingerprinting approach that combines three distinct analytical techniques to classify commercial cigarettes sold in Brazil into their market categories — multinational, regional, smuggled Paraguayan, counterfeit, and products lacking health registration — with a rigor that explicitly accounts for the pitfalls of hidden statistical structure in the data. The study, published in Results in Chemistry, integrates elemental profiles measured by inductively coupled plasma optical emission spectrometry, semiquantitative n-alkane profiles obtained by gas chromatography with flame ionization detection, and bulk carbon isotope and carbon–nitrogen content measurements from elemental analysis–isotope ratio mass spectrometry. By fusing these chemically complementary datasets through a low-level concatenation strategy and feeding the result into multiclass linear discriminant analysis, the researchers have produced one of the most transparent assessments to date of what chemical fingerprinting can — and cannot — legitimately claim in the forensic fight against illicit tobacco.

The stakes in Brazil are considerable. More than 30 percent of the country’s cigarette consumption has been associated with the illicit market, and in regions bordering Paraguay estimates climb to roughly 60 percent. Products that look commercially similar on a shelf can differ radically in legal status, manufacturing history, and concentrations of potentially harmful constituents. The challenge is compounded by the fact that the market categories do not represent chemically isolated production systems: clandestine factories reproducing Paraguayan brands have been documented inside Brazil, and tobacco and industrial inputs circulate through cross-border supply chains linking the two countries. Any forensic chemical method must therefore answer two related questions simultaneously — whether products can be discriminated by market category at all, and whether their multivariate profiles reveal compositional affinities consistent with shared raw materials, blend characteristics, or manufacturing practices.

The analytical logic underpinning the new work rests on the complementary nature of the three platforms. Elemental analysis by ICP-OES probes the soil–plant–product continuum, capturing information tied to soil geochemistry, fertilization, plant uptake, environmental exposure, and processing. The researchers determined ten elements — barium, boron, cadmium, cobalt, copper, iron, manganese, nickel, strontium, and zinc — in tobacco that had been dry-ashed in a muffle furnace at 550 °C following a procedure adapted from AOAC Official Method 930.05, then dissolved in nitric acid. External calibration curves spanning seven standards were used, and method accuracy was verified against the plant-matrix certified reference material NIST-SRM 1573a. The organic block, by contrast, characterized solvent-extractable n-alkanes: tobacco was extracted with dichloromethane in an ultrasonic bath over three consecutive five-minute cycles, and the resulting profiles were semiquantified by GC-FID on a DB-5 ms column, with a deuterated n-C24 internal standard correcting retention times and normalizing chromatographic response. The internal standard, added after extraction, corrected instrumental drift but did not measure extraction efficiency — an honesty the authors preserve by describing their values as method-dependent concentrations rather than absolute recoveries.

The third analytical block came from EA-IRMS. Tobacco was separated from paper and filters, ground, homogenized, and weighed into tin capsules in roughly 0.2 mg aliquots, then combusted at 1020 °C in an elemental analyzer coupled to an isotope-ratio mass spectrometer. Bulk δ13C values, organic carbon content, and total nitrogen content were determined, with performance monitored using the IAEA-USGS40 glutamic acid reference material. Triplicate measurements of the standard yielded −26.33 ± 0.34‰ against a certified value of −26.389 ± 0.042‰ — an absolute bias of just 0.06‰ — while carbon and nitrogen recoveries of approximately 98.9 and 99.2 percent demonstrated the reliability of the acetanilide-based external calibration. Stable isotopes contribute information related to plant physiology, environmental growing conditions, and raw-material origin, filling in dimensions that neither the elemental nor the molecular block can access alone.

The chemometric architecture chosen to bind these datasets together is low-level data fusion, which directly concatenates preprocessed original descriptors after autoscaling within each block and block scaling by the square root of each block’s variable count. Unlike mid-level fusion, which integrates selected latent features, or high-level fusion, which combines outputs from independently constructed models, low-level fusion preserves the chemical identity of every variable — an interpretability advantage the authors argue is essential in forensic contexts where predictive performance alone is insufficient if the chemical basis of discrimination cannot be examined. Three multiclass LDA models were built for comparison: one on the inorganic block alone, one on the combined organic and isotopic descriptors, and one on the fused matrix. Before model construction, a robust principal component analysis procedure, ROBPCA, was applied to the autoscaled fused matrix to screen for multivariate outliers, and stratified sampling divided the data into 60 percent training and 40 percent test subsets.

The sample set comprised 32 commercial cigarette brands — 11 multinational, 8 regional, 7 smuggled Paraguayan, 3 counterfeit, and 3 with no health registration — each analyzed in triplicate, yielding 96 replicate-level datasets. Crucially, the team recognized that these 96 observations are not 96 independent units. All three replicate preparations of a brand share that brand’s chemical identity, and if replicates from the same brand land in both the training and test sets, brand-specific structure can inflate apparent performance — a statistical trap known as pseudoreplication. The researchers confronted the issue head-on rather than quietly ignoring it. They retained replicate-level models because these describe the chemical discrimination present in the analytical dataset and its internal stability, but they explicitly reinterpreted replicate-level accuracy as internal classification consistency, not as generalization to unseen products.

To quantify the optimism directly, the team performed a deliberately conservative brand-level sensitivity analysis. The triplicate measurements were averaged for each brand, collapsing the dataset to 32 brand-level profiles, and a 60:40 train/test partition was repeated 100 times. In each repetition, preprocessing and PCA were fitted exclusively to the training brands, principal components explaining roughly 90 percent of cumulative training variance were retained, and held-out brands were projected into the training-defined PCA space before LDA fitting and prediction. The gap between replicate-level and brand-level accuracy serves as a sensitivity measure of brand clustering and pseudoreplication-related optimism, though the authors caution that lower brand-level performance reflects both the removal of information sharing and the much smaller effective sample size, since the task changes from classifying replicates to extrapolating to previously unseen brand profiles.

Robustness was probed further by repeating the stratified partition 100 times for each of the three datasets — inorganic, organic/isotopic, and fused — independently refitting preprocessing, block scaling, and LDA in every repetition, and summarizing accuracy distributions by means, standard deviations, and ranges. Welch’s two-sample t-test evaluated whether the fused model’s accuracy distribution differed significantly from those of the single-block references. Beyond the supervised models, one-way ANOVA with Tukey’s honestly significant difference post hoc test at α = 0.05 was used to describe univariate intercategory differences in individual descriptors, while unsupervised PCA explored the compositional proximity of smuggled and counterfeit brands to the centroids of the multinational manufacturers represented in the dataset — British American Tobacco, Japan Tobacco International, and Philip Morris. This exploratory question is particularly provocative given the documented cross-border circulation of tobacco and industrial inputs: proximity in the unsupervised chemical space could be consistent with shared raw materials or manufacturing practices, though the authors delimit the level of inference supported by the sample set.

The study also extends a previous GC-FID investigation of the same brand set, which had used a non-targeted chromatographic fingerprint aligned with GCalignR and classified by linear discriminant analysis without compound identification or quantification. That earlier workflow established a rapid screening capability but limited mechanistic interpretation of the discriminant variables and ignored elemental and isotopic information entirely. The new work upgrades the framework in three ways: it uses chemically defined descriptors from all three platforms, explicitly evaluates the effect of the nested replicate structure, and examines proximity relationships between illicit products and multinational manufacturer centroids within the fused unsupervised chemical space. Compared against other published approaches — NIR-based trademark discrimination, combined IRMS–ICP-MS geographical traceability of tobacco, multi-stable-isotope authenticity discrimination with reported accuracies above 96.8 percent, and a 2026 multisensor shredded-tobacco study achieving data-level fusion accuracies of 99.22 percent — the Brazilian work stands out for targeting the harder forensic problem of regulatory and illicit market categories rather than mere product recognition.

Model interpretation relied on discriminant score plots, explained discriminant variance, and variable contributions estimated from absolute scaling coefficients, with predictive performance assessed through confusion matrices, overall accuracy, sensitivity, specificity, precision, F1-scores, and ROC curves with area under the curve computed from posterior LDA probabilities. All chemometric analyses were performed in R, and the complete scripts and processed datasets were deposited in Zenodo, an open repository, allowing independent scrutiny of every preprocessing and modeling decision. That level of transparency — from procedural blanks processed alongside every batch to the frank acknowledgment that no extraction-recovery correction was applied to the n-alkane data — is what distinguishes a defensible forensic workflow from an inflated one.

Ultimately, the study delivers a calibrated answer to a deceptively simple question. Multiplatform chemical fingerprints genuinely do carry the information needed to separate Brazilian cigarette market categories, and their fusion through low-level concatenation preserves the interpretability that courts and regulators require. But by systematically comparing replicate-level and brand-level validation, the researchers have also drawn a bright line around what the numbers mean: replicate-level results characterize internal discrimination within the measured dataset, whereas brand-level validation — the more conservative estimate — is the appropriate basis for claims about previously unseen commercial products. In a market where more than a third of consumption may be illicit and where packaging can deceive, the ability to ground authentication in objectively measured elemental, molecular, and isotopic chemistry represents a meaningful step forward for consumer protection and forensic enforcement alike.

Subject of Research: Chemometric classification and forensic authentication of commercial cigarettes sold in Brazil using low-level data fusion of ICP-OES elemental profiles, GC-FID n-alkane profiles, and EA-IRMS isotopic and elemental measurements

Subject of Research: Chemistry

Article Title: Multiblock chemometric classification of commercial cigarettes sold in Brazil employing low-level data fusion of elemental, organic, and isotopic profiles

Article References: Rodrigues, L. S., Massone, C. G., & de Oliveira Godoy, J. M. (2026). Multiblock chemometric classification of commercial cigarettes sold in Brazil employing low-level data fusion of elemental, organic, and isotopic profiles. Results in Chemistry, 30, Article 103801. https://doi.org/10.1016/j.rechem.2026.103801

Image Credits: AI Generated

DOI: 10.1016/j.rechem.2026.103801

Keywords: cigarette authentication, chemometrics, low-level data fusion, linear discriminant analysis, ICP-OES, GC-FID, EA-IRMS, δ13C, n-alkanes, forensic chemistry, illicit tobacco market, pseudoreplication

Cite Scienmag News
APA MLA Chicago

Bethany Barker. (September 4, 2026). Brazilian cigarette brands classified by fused chemical and isotope data. Scienmag. https://scienmag.com/brazilian-cigarette-brands-classified-by-fused-chemical-and-isotope-data/

Copy citation Download RIS

Tags: application of gas chromatography and isotope ratio mass spectrometry in tobacco analysisBrazilian cigarette brand classificationBrazilian cigarette brands classificationchallenges in tobacco product classification and regulationchemical fingerprinting of cigaretteschemical fingerprinting of tobacco productschemical markers distinguishing legal and illegal tobaccochemical markers for tobacco product origincounterfeit cigarette detection strategiesdetection of illicit and counterfeit cigarettes in Brazildistinguishing legal and illicit tobacco productselemental and isotope profiling of cigaretteselemental profiling of tobacco productsforensic tobacco authentication methodsgas chromatography for cigarette fingerprintingillicit tobacco market in Brazilimpact of illicit tobacco trade in Brazilisotope ratio mass spectrometry in cigarette analysismulti-analytical techniques for tobacco analysismultiplatform analytical techniques in tobacco analysismultivariate analysis in tobacco forensic studiesmultivariate statistical analysis in tobacco forensicstobacco product market segmentation using chemical data