qNMR Summit 2023 - Poster - Triggiani Maurizio - Poliba

Combining NMR and Artificial Intelligence to Improve Food Identification: Insights from a Pilot Project with the Central Inspectorate of Quality Protection and Fraud Repression (ICQRF, part of the Italian Ministry of Agriculture

Maurizio Triggiani,^a,b Biagia Musio,^a,b Stefano Todisco,^a Marica Antonicelli,^a Maria Trisolini,^b Piero Mastrorilli,^a,b Mario Latronico,^a,b Stefania Carpino,^c Vincenzo Di Martino,^c Laura Gambino,^c Vito Gallo^a,b
^a Politecnico di Bari, DICATECh – via Orabona 4, CAMPUS, I – 70125, Bari, Italy
^b Innovative Solutions S.r.l. – Zona H 150/B, 70015, Noci (BA), Italy
^c Ministero dell’agricoltura, della sovranità alimentare e delle foreste, Ispettorato centrale della tutela della qualità e repressione frodi dei prodotti agroalimentari, Ufficio PREF IV – Via Quintino Sella, 42, 00187, Roma, Italy
📧 maurizio.triggiani@poliba.it

Abstract

Food safety and fraud prevention are of great importance in maintaining public health, safeguarding consumer trust, and ensuring the integrity of the import/export fluxes. The increasing complexity and globalization of the food industry have made it increasingly challenging to detect and combat food fraud effectively. This study investigates the potential of integrating Nuclear Magnetic Resonance (NMR) spectroscopy and Artificial Intelligence (AI) techniques to address these challenges and enhance food safety and fraud detection capabilities. NMR spectroscopy is a powerful analytical tool capable of identifying and quantifying metabolites in complex food matrices, providing detailed information on food composition and quality. AI techniques, can process vast amounts of data efficiently, identifying patterns and making predictions with high accuracy.

In collaboration with the Central Inspectorate of Quality Protection and Fraud Repression (ICQRF), a pilot program was conducted to evaluate the effectiveness of this integrated approach. The project focused on two food products with a high potential for fraud: lentils and wheat. These products were selected due to their economic significance and vulnerability to adulteration and mislabeling. A set of authentic and non-authentic samples were provided by ICQRF and analyzed using NMR spectroscopy. The resulting spectral data were then processed using AI algorithms.

The integrated NMR and AI approach demonstrated a high level of accuracy in identifying solid information about the origin of the samples, cultivar, and agronomical practices used. Additionally, the methodology offered meaningful insights into the samples’ attributes, emphasizing details related to seasonal effects and the geographic origin of the fields in which they were grown. This contributed to a more thorough comprehension of the samples’ history and growth conditions.

The findings from this pilot program suggest that the integration of NMR and AI can significantly enhance the capabilities of food safety authorities to detect and prevent food fraud, is also important to consistently update the samples database, ensuring that the classification model remains effective in recognizing new samples with every changing season.

Samplig Procedure

Extraction procedure

50 g of each durum wheat sample was ground using a blender for 1 minute and sieved using a laboratory sieve (sieve size of 0.5 mm). An aliquot of 300 mg of the resulting wheat flour was dissolved in 3 mL of buffer solution [(HC2O4)-/(C2O4)2- (0.25 M) and NaN3 (2.5 mM)] at pH 4.2, sonicated for 10 min at 60°C, vortexed for 5 min at 2500 rpm and, centrifuged for 10 min at 6000 rpm. The resulting supernatant was drawn off and pasteurized for 10 min at 90°C.

After pasteurization, the sample was filtered through a syringe filter equipped with a polytetrafluoroethylene (PTFE) hydrophilic membrane with a pore size of 0.22 µm, to gain a clear sample without any suspended particles that may distort the magnetic field homogeneity. The filtrate was used to fill the NMR tubes.

NMR tubes were filled in by an automated system for liquid handling (Sam-plePro Tube, Bruker BioSpin) according to the following method: 630 μL of the obtained extract and 70 μL of TSP/D2O solution [3-(trimethylsilyl) propi-onic-2,2,3,3-d4 acid sodium salt in D2O (0.2% w/w)]. Two replicates were pre-pared for each sample.

NMR analysis

One-dimensional 1H NOESY spectra were acquired under automaton at 298.1° K using a Bruker Avance 400 MHz spectrometer equipped with a 5 mm in-verse probe and with an autosampler (Bruker, Billerica, MA, USA).

The following acquisition parameters were used: pulse program = noe-sygppr1d; size of fid (TD) = 64 K; spectral width (SW) = 20 ppm; transmitter offset = 4.70 ppm; 90° hard pulse (p1) = 8.16 μs; power level for pre-saturation (pl9) = 62.77 dB; dummy scans (ds) = 4; number of scans (ns) = 64; acquisition time = 4.09 s; mixing time (d8) = 0.01 s; recycle delay (d1) = 10 s. Each spectrum was acquired using TOPSPIN 2.1 software (Bruker BioSpin GmbH, Rheinstetten, Germany) under an automatic process that lasted ca. 22 min and encompassed sample loading, temperature stabilization for 5 min, tuning, matching, shimming, and 90◦ pulse calibration.

NMR raw data (Free Induction Decays, FIDs) were processed using the soft-ware MestReNova 11.0 (Mestrelab Research SL, Santiago de Compostela, Spain).

The 1D 1H NOESY spectra were automatically Fourier transformed, manually phased, and automatic baseline corrected (by a 1st order Bernstein polyno-mial equation). Chemical shifts were referenced to the TSP-d4 peak at δ = 0.00 ppm.

Processing of spectral data for statistical analysis

The processed 1D 1H NOESY spectra were reduced to a numerical matrix called bucket-table. The bucket table was obtained by dividing the entire overlapped spectra in the range of [-0.5, 10.5] ppm into rectangular intervals (buckets) of 0.04 ppm in width, excluding the region [4.30,5.10] ppm corre-sponding to the residual water signal. The underlying area of each bucket was normalized to the total intensity. Thus, in the bucket table, the NMR spectra recorded from samples constituted the observation, and the buckets consti-tuted the x-variables.

The numerical matrix generated for the durum wheat samples provided by ICQRF is described in the paragraphs of the results.
The bucket table was imported into SIMCA 13.0.3 software (Umetrics, Umea, Sweden) to perform multivariate statistical analyses (MVA).

Publications

V. Gallo et All – “Performance assessment in fingerprinting and multi component quantitative NMR analyses” – Analytical chemistry (2015)

Development of a food class-discrimination system by non-targeted NMR analyses using different magnetic field strengths

B. Musio et all – “A community-built calibration system: the case study of quantification of metabolites in grape juice by qNMR spectroscopy” – Talanta, (2019),