Rapid analysis of meat floss origin using a supervised machine learning-based electronic nose towards food authentication

The study addresses the need for accurate, rapid authentication of meat floss origin, particularly to identify pork due to allergy risks and religious dietary restrictions. Conventional biochemical detection methods (DSC, ELISA, NMR, PCR) are accurate but require laboratory infrastructure, reagents, and trained personnel. Given that meat products emit characteristic volatile organic compounds (VOCs), an electronic nose (e-nose) with gas sensor arrays and pattern recognition offers a portable, low-power, rapid alternative for food authenticity. The research question is whether a compact e-nose combined with supervised machine learning and a time window slicing feature-extraction method can reliably discriminate beef, chicken, and pork flosses. The study evaluates unsupervised PCA and four supervised models (LDA, QDA, k-NN, RF) and validates e-nose classifications with FTIR and GC-MS to link sensor patterns to chemical compositions.

Prior work has used various analytical techniques (DSC, ELISA, NMR, PCR) for pork identification but these are lab-bound and resource-intensive. E-noses have been applied to monitor VOCs for air quality, disease screening, and food authenticity/adulteration, showing advantages such as high sensitivity, rapid classification, simple sample prep, low power, and portability. In meat quality and safety evaluation, e-noses have been used for different meats (chicken, pork, goat, fish, duck, goose). PCA is commonly used for dimensionality reduction and initial visualization, while supervised models including LDA, QDA, k-NN, and RF can classify extracted features with relatively simple architectures and low computational cost. Reported mass spectrometry studies indicate overlapping VOC groups among meats (e.g., hydrocarbons and alcohols), with pork often showing distinctive aldehydes, motivating chemometric approaches to separate classes despite overlapping volatiles.

Sample preparation: Beef, chicken, and pork were sourced from local Indonesian animals (chicken 45–75 days, cows 2–4 years, pigs 4–6 months). Selected pure meat cuts (no added fat/bone) were processed into floss in a university laboratory: boiled, shredded, fried to dry, and spun to remove moisture, yielding separate, pure beef, chicken, or pork flosses. For e-nose, 300 total samples (100 per meat type), each 2.0 g, were measured at room temperature. FTIR used KBr pellets; GC-MS samples were methanol-extracted with 150 g per meat type prepared for both FTIR and GC-MS. E-nose hardware and measurement: A portable e-nose with eight metal-oxide semiconductor (MOS) chemoresistive gas sensors (S1–S8) detected a range of gases (hydrocarbons, alcohols, ethers, ammonia, etc.). Signals were acquired via voltage divider circuits, digitized by a 16-bit ADC, and read every 100 ms by a microcontroller, then sent over RS232 to a PC. The setup used two chambers: a sampling chamber (with the floss sample) and a sensing chamber (with sensors). Each run lasted 260 s: 20 s delay (ambient air), 120 s sampling (reference air passed through sample headspace into sensing chamber), and 120 s purging (reference air to clear residuals). Signals were baseline-corrected by subtracting the initial delay-phase value. Feature extraction and windowing: To enhance class separability, a time window slicing method was applied to the sensing phase data (generally 20–200 s). For W0 (no window), features were taken over the entire trace; for W1–W6, the 20–200 s interval was partitioned into equally spaced windows. In each window, basic statistical features were extracted per sensor: maximum, minimum, mean, and median. Additional formulation for window integration is provided (window function K_i(t_k) with parameters defining width/shape/center), though the primary features used for model comparisons were max/min/mean/median per window. Data processing and models: Data were centered and scaled. The dataset (300 samples) was split into 75% training (225) and 25% testing (75) by random sampling. Model development used 10-fold cross-validation with 10 repetitions within the training set; testing used the held-out 25% for external validation. Unsupervised PCA was used for initial visualization and cluster assessment. Supervised models assessed included LDA, QDA, k-NN, and RF. Hyperparameters: k-NN used Euclidean distance; RF tuning included mtry (best case reported mtry=6 with minimum feature at W5). Analyses were implemented in R (v3.5.1) with CARET, MASS, and Kernlab libraries. FTIR and GC-MS: FTIR (Shimadzu Prestige 21) spectra were recorded from 400–4000 cm−1 on dried, KBr-pelleted samples. GC-MS used UHP helium, injector 260 °C, MS transfer 250 °C, ion source 200 °C, specified flows and split ratio; HP-5MS UI column (30 m, 0.25 mm ID). Compounds were identified via MS fragmentation and the NIST 14 library.

- PCA revealed that pork floss samples formed a distinct cluster separable from beef and chicken, while beef and chicken overlapped in PC1 but separated in PC2, consistent with shared volatile groups. - Time window slicing increased inter-cluster distances and reduced overlap. Increasing window numbers generally enlarged distances between class centroids (beef–chicken, beef–pork, chicken–pork, and pork–non-pork), improving separability. - LDA performance (Table 2): With five windows (W5), validation accuracy reached up to 99.9% and testing accuracy reached 100% across max/min/mean/median features. Testing accuracy at W5 and W6 was 100% for all four features; internal validation at W6 slightly decreased relative to W5 (e.g., maximum: 99.4% vs 99.9%), indicating W5 as optimal. - Supervised model comparison at W5 (Table 3): - LDA: Validation 99.4–99.9%; Testing 100% across features, overall >99%. - QDA: Validation 95.9–98.5%; Testing 93.0–99.0% depending on feature. - k-NN: Validation 93.4–97.0%; Testing 89.3–93.3%. - RF: Validation 97.5–98.7%; Testing 96.0–100%. Best RF case achieved 100% testing accuracy for the minimum feature at W5 with mtry=6. - Dataset and splitting: 300 samples (100 per class), 75%/25% train/test split; internal validation via 10-fold CV repeated 10 times. - FTIR: All meats showed typical edible fat/oil spectra with protein-related bands (3290, 3076 cm−1) and lipid bands (2924, 2854 cm−1), and strong ester carbonyl at 1743 cm−1. Pork exhibited higher intensity at 1743 cm−1 (fatty acids) and lower intensity at 1650 cm−1 (protein) than beef/chicken, indicating relatively higher fat and lower protein content in pork floss. - GC-MS: Common compounds across meats included hydrocarbons (e.g., 2,4-dimethylhept-1-ene, 4-methyl-1-decene, (E)-4-dodecene, (Z)-5-tridecene, 3-trifluoroacetoxytridecane), ethers (6-methylheptyl vinyl ether), and alcohols (2-butyl-1-octanol, 2,4-di-tert-butylphenol, ethyl iso-allocholate, 6,11-dimethyl-2,6,10-dodecatrien-1-ol). Beef and chicken shared additional hydrocarbons (1-methylhexyl hydroperoxide, 4-ethyl-octane). Chicken-only markers included 6-methyl-octadecane and methoxyacetic 2-tridecyl ester. Pork-only VOCs included alcohols (trans-2-dodecen-1-ol, 12-methyl-E,E-2,13-octadecadien-1-ol) and aldehydes (octadecanal, 9-octadecenal), aligning with e-nose separability of pork.

The e-nose coupled with supervised learning effectively addressed the need for rapid, portable authentication of meat floss origin. Time window slicing enhanced discrimination by capturing dynamic response segments of the trapezoidal MOS sensor signals. LDA with five windows maximized separation among beef, chicken, and pork, achieving near-perfect validation and perfect external testing accuracy. The clear separability of pork from non-pork aligns with GC-MS findings that aldehydes (e.g., dodecanal-related species) dominate in pork but not in beef or chicken, and with FTIR indications of higher lipid-related signals in pork. Although beef and chicken share many VOCs, windowed features and LDA resolved their clusters, demonstrating that appropriate feature engineering is critical. The findings support e-nose as a practical alternative or complement to FTIR and GC-MS for on-site food authenticity screening, especially for rapid pork vs non-pork discrimination, with strong potential to mitigate food fraud and support dietary compliance.

A compact MOS sensor-based e-nose integrated with supervised machine learning and time window slicing can rapidly and accurately authenticate meat floss origin. LDA with five time windows provided the best balance of internal validation and external testing, achieving >99% validation and 100% testing accuracy across statistical features. Corroboration with FTIR and GC-MS linked sensor patterns to chemical differences, notably pork-specific aldehydes. The system offers a promising, portable, and low-complexity approach for food authenticity testing and pork detection. Future work could expand to mixed/adulterated samples, quantify adulteration levels, evaluate robustness under varying environmental conditions, explore additional classifiers and feature-learning methods, and extend to broader product categories and supply-chain settings.

Explicit limitations are not extensively discussed. The study used pure, separately prepared beef, chicken, and pork flosses from local Indonesian sources and tested three classes only, without mixed/adulterated samples. Measurements were conducted under controlled laboratory conditions with a single e-nose setup. These factors may affect generalizability across diverse product formulations, environments, and supply chains. Further validation on mixed samples, different processing methods, and broader geographic sources would strengthen applicability.