Benchtop NMR spectroscopy for quantitative determination of milk fat and qualitative determination of lactose: from calibration cruve to deep learning

Bench op NMR spec oscopy o quan i a i e de e mina ion o milk a and quali a i e de e mina ion o lac ose: F om calib a ion cu e o deep lea ning Jos´ e Raúl Belmon e-S´ anchez a , Robe o Rome o-Gonz´ alez a , Manuel ´ Angel Ma ínez O osa b , Ma ía Cal o Mo a a b , An onia Ga ido F enich a,* a Depa men o Chemis y and Physics, Analy ical Chemis y A ea, Ag i ood Campus o In e na ional Excellence ceiA3, Uni e si y o Alme ía Resea ch Cen e o Ag icul u al Food Bio echnology (CIAIMBITAL), E-04120, Alme ía, Spain b Ten amus Labo a o io de Con ol S.L.U., Edi icio Las Rozas, 23, C a. de la Co u˜ na, Km 23.200., Las Rozas, E- 28290, Mad id, Spain ARTICLE INFO Keywo ds: Milk a quan i ica ion Bench op NMR spec oscopy Machine lea ning A i icial neu al ne wo ks Pa ial leas squa es eg ession ABSTRACT This s udy compa es h ee di e en me hodologies o he quan i ica ion o he a con en o ul a-high em- pe a u e (UHT) milk using bench op p o on nuclea magne ic esonance ( 1 H NMR) spec oscopy, a lagship o g een, accessible, and s a e-o - he-a echnology sui able o mode n labo a o y en i onmen s. The e alua ed app oaches included adi ional calib a ion cu e and machine lea ning algo i hms, wi h emphasis on pa ial leas squa es eg ession (PLS-R) and a i icial neu al ne wo ks (ANN), o es ima e he a con en in skimmed, semi-skimmed and whole milk. Among hese, ANN p o ided he mos accu a e esul s o all ypes o milk, pa icula ly in skimmed milk, wi h a ela i e s anda d de ia ion (RSD) o 14.9% and an accu acy o −7.3%. The calib a ion cu e showed highe a iabili y, wi h an RSD o 34.1% and ueness o 25.3% o skimmed milk. PLS- R imp o ed accu acy in ela ion o he calib a ion cu e app oach, educing RSD o 18.9% and ueness o −17.7%. The de eloped me hod has been success ully applied o de e mine he a con en in 51 samples o UHT milk pu chased in di e en Spanish supe ma ke s, p o iding adequa e esul s o each o he h ee ca ego ies conside ed, including goa ’s milk, sheep’s milk, and milk co ee. Fu he mo e, he applica ion o machine lea ning has p o en i s alidi y by success ully dis inguishing be ween lac ose and lac ose- ee UHT milk. 1. In oduc ion In he mode n dai y indus y, he assessmen o a con en in ood ma ices, such as ul a-high- empe a u e (UHT) cow’s milk, is o pa a- moun impo ance. UHT milk, cha ac e ized by i s s e iliza ion p ocess ha in ol es hea ing o empe a u es abo e 135 ◦C o a sho pe iod o ime ( ypically 2–5 s), is speci ically designed o elimina e mic obial li e, including spo es. This p ocess signi ican ly ex ends he shel li e o milk, allowing i o be s o ed wi hou e ige a ion un il i is opened (Tamime, 2008). The widesp ead popula i y o UHT milk has led o i s inc eased a ailabili y and consump ion a ound he wo ld, wi h a no able p esence in coun ies such as Spain, whe e 95% o he milk consumed is UHT (Aquini & Gil, 2017). These ac o s highligh he impo ance o UHT milk in ensu ing ood sa e y, imp o ing accessibili y, and mee ing consume p e e ences. Based on he impo ance o UHT milk, Regula ion (EU) No. 1308/ 2013 se s s ic EU s anda ds o milk a labelling. The e o e, whole milk mus con ain a leas 3.25 g/100 mL milk a , semi-skimmed be- ween 1.5 g/100 mL and 1.8 g/100 mL, and skimmed milk less han 0.5 g/100 mL (Ama al e al., 2018). These ules a e designed o main ain anspa ency, suppo consume choice, and ensu e ai compe i ion. Fo he analysis o he a con en in milk and dai y p oduc s, a ious analy ical echniques a e employed. The mos widely adop ed me hod in ol es gas ch oma og aphy coupled wi h lame ioniza ion de ec ion (GC-FID). This echnique is pa icula ly alued o i s abili y o sepa a e and analyze he composi ion o a y acids (FAs), al hough i equi es ex ensi e sample p epa a ion (Danudol & Judp asong, 2022). Ano he a ian o his de e mina ion in ol es he coupling o GC wi h mass spec ome y (GC-MS), which imp o es he speci ici y and sensi i i y o he analysis, bu i comes wi h highe complexi y and cos (Chen e al., 2023). In addi ion o hese ch oma og aphic echniques, nea -in a ed spec oscopy (NIRS) (E angelis a e al., 2021) and Raman * Co esponding au ho . E-mail add ess: [email p o ec ed] (A. Ga ido F enich). Con en s lis s a ailable a ScienceDi ec LWT jou nal homepage: www.else ie .com/loca e/lw h ps://doi.o g/10.1016/j.lw .2024.117000 Recei ed 4 June 2024; Recei ed in e ised o m 25 Oc obe 2024; Accep ed 5 No embe 2024 LWT - Food Science and Technology 212 (2024) 117000 A ailable online 6 No embe 2024 0023-6438/© 2024 Published by Else ie L d. This is an open access a icle unde he CC BY-NC-ND license ( h p://c ea i ecommons.o g/licenses/by-nc-nd/4.0/ ). spec oscopy (Reine e al., 2020) a e well es ablished as non-des uc i e me hods o he analysis o FA in milk and dai y p od- uc s, al hough hey can comp omise on speci ici y and sensi i i y due o luo escence in e e ences. Each o hese me hods has i s own se o applica ions, ad an ages, and limi a ions, which makes he choice o me hod dependen on he speci ic equi emen s o he analysis. The in oduc ion o nuclea magne ic esonance (NMR) spec os- copy, speci ically p o on NMR ( 1 H NMR), as an al e na i e o analyze he a con en o milk and dai y p oduc s, b ings signi ican bene i s. P e ious s udies ha e success ully applied high- esolu ion NMR ech- niques o quan i ying milk a using a ious s a egies, such as wo- dimensional NMR (Hu e al., 2007), ime-domain (P. M. San os e al., 2016) o eg ession models (Monakho a e al., 2012), p o ing he echnique’s e icacy. Howe e , bench op NMR echnology is a ecen ad ancemen and i s applica ion in his con ex is cu en ly unde in es iga ion. Bench op NMR ins umen s, wi h hei educed size, make his so- phis ica ed echnique mo e accessible o smalle labo a o ies and p o- duc ion acili ies, acili a ing mo e equen and ex ensi e es ing (D ape & McCa ney, 2023; Gal an e al., 2021; an Beek, 2021). These cos -e ec i e e sions o adi ional high- esolu ion NMR sys ems o e a iable op ion o ou ine analyses. NMR spec oscopy enables he di ec and non-des uc i e quan i ica ion o a ious componen s in complex mix u es like dai y p oduc s, elimina ing he need o ex ensi e sample p epa a ion o chemical eagen s (Bu ge e al., 2022). The use o bench op NMR also aligns wi h he indus y’s mo e owa d mo e apid and e icien es ing me hodologies ha can p o ide eal- ime da a o quali y con ol and assu ance p ocesses (Ezeanaka e al., 2019). So a , some s udies ha e been published on cow UHT milk using bench op NMR. Soyle e al. di e en ia ed milk samples based on p ope ies such as glyce ol, a , and suga con en using NMR spec a combined wi h an a i icial neu al ne wo k (ANN) model (Soyle e al., 2021). In ano he s udy, a me hod ocused on T 2 elaxa ion imes (acquisi ion o he ime domain) o de e mining he a con en (be- ween 0.1 g/100 mL and 9 g/100 mL) in milk. This app oach equi ed a con as solu ion consis ing o a 6 g/100 mL NaCl solu ion con aining 2 g/100 mL sodium e ic e hylenediamine e aace a e (Fe-EDTA) o adap he elaxa ion a e o wa e (Sø ensen e al., 2022). I is impo an o cla i y ha he dis inc ion be ween s udies does no imply supe io i y o one o e he o he ; a he , i highligh s di e gen me hodological choices based on di e en esea ch objec i es and applica ions. Quan i a i e NMR, including bo h high and low esolu ion mea- su emen s, wo king wi h di e en egions o he spec a, has been applied o simul aneously measu e di e en componen s in a ious dai y p oduc s (Ha zakis, 2019). To achie e his, one o he mos commonly used me hodologies in ol es he use o calib a ion cu es, whe e known s anda ds a e plo ed agains hei esponses o de e mine concen a ions, guided by cu en alida ion guidelines (Bha i & Roy, 2012). Machine lea ning, especially h ough eg ession models, o e s a mo e speci ic app oach o complex da a ha may no i simple uni- a ia e linea ela ionships. Wi hin machine lea ning, ANNs s and ou by emula ing human neu al ne wo ks, allowing o pa e n ecogni ion wi hin la ge da ase s ha migh elude con en ional s a is ical me hods (da Cos a e al., 2021; Joshi, 2023; M. C. San os e al., 2019). Each o hese quan i ica ion me hods has i s own se o ad an ages and appli- ca ions, o en chosen based on he speci ic equi emen s o he analysis, he complexi y o he sample ma ix, and he a ailable ins umen a ion and compu a ional esou ces. In he case o eg ession models applied wi hin machine lea ning con ex s, pa icula ly in chemome ic s udies using NMR, pa ial leas squa es eg ession (PLS-R) s ands ou as one o he mos commonly u ilized and obus me hods, especially alued o i s abili y o handle highly collinea and mul idimensional da a ypical o NMR s udies. The popula i y o PLS-R s ems om i s dual abili y o pe o m dimensionali y educ ion and eg ession simul aneously (Dhaulaniya e al., 2020; Gal- an e al., 2021; Huang e al., 2022). In addi ion o PLS-R, o he models such as suppo ec o machines (SVM) (Hou e al., 2019) o andom o es s (RF) (Zhao e al., 2019) ha e also been used. Howe e , he choice o model is jus he beginning, and adhe ence o specialized alida ion guidelines, such as hose om EUROLAB (Sch¨ onbe ge e al., 2015), is essen ial o ensu ing he model’s eliabili y and gene alizabili y. Fo ANN, he model design p ocess is e en mo e sophis ica ed, gi en he complexi y and “black box” na u e o hese models. In addi ion o gene aliza ion, op imiza ion o ANNs in ol es op imizing he ne wo k a chi ec u e, including he numbe o laye s and nodes, and uning hype pa ame e s, such as he lea ning a e and egula iza ion e ms (Debik e al., 2022). To he bes o ou knowledge, he e a e no guide- lines ha add ess he op imiza ion o alida ion o ANN based quan i- ica ion me hods. Howe e , i is possible o co e his issue om he poin o iew o alida ion o chemome ic me hods (Benzaama e al., 2022; Taylo , 2006). C oss- alida ion echniques, such as k- old c oss- alida ion, a e commonly employed o ensu e ha he model’s pe o mance is obus ac oss di e en subse s o he da a (Co sa o e al., 2022; Pomyen e al., 2020). A emp ing o add ess he di e ences be ween measu emen me hods and o e ing a igo ous app oach applicable o con empo a y analy ical p oblems, his s udy compa es h ee me hods o quan i y he con en o UHT milk a using desk op NMR: adi ional calib a ion cu es, and machine lea ning om PLS-R o ANN (deep lea ning). In addi ion o analyzing hei bene i s, limi a ions, and alida ion o iden i y he mos accu a e, e icien , and eliable app oach o imp o e quali y con ol in he dai y indus y, an analysis o en i onmen al sus- ainabili y was also ca ied ou . This demons a es he syne gy be ween hese ypes o s udies and hei alignmen wi h g een analy ical chem- is y p inciples, showcasing ou commi men o p omo ing sus ainable p ac ices wi hin analy ical esea ch. 2. Ma e ials, eagen s, and me hods 2.1. Milk samples and sample p epa a ion A o al o 55 milk and dai y p oduc samples we e collec ed om di e en Spanish supe ma ke s. Table S-1 shows he samples andomly o de ed as analyzed. The samples included di e en ypes o a con en : skimmed (14), semi-skimmed (23), and whole (18), wi h a con en anging om less han 0.1 g/100 mL o 3.6 g/100 mL, excep o condensed milk wi h g/100 mL a , and a c eam sample wi h 33 g/100 mL a . Mos o he samples we e o cow o igin (51), wi h a mino ep esen a ion o goa and sheep milk ( wo om each), ensu ing a comp ehensi e analysis o a ious ypes o milk. The sample se also included dai y de i a i es such as co ee wi h milk (2), ke i (1), and yoghu (1), p o iding a b oad examina ion o milk and dai y p oduc s. Addi ionally, o he calib a ion cu e me hod, a se o UHT cow milk e e ence s anda d om QSE-GmbH (Wolnzach, Ge many) was pu - chased, co e ing he ange o a con en wi hin he samples (0.06 g/ 100 mL, 1.52 g/100 mL, 3.48 g/100 mL and 4.33 g/100 mL). Be o e NMR measu emen s, 320 μ L o a 1 g/100 mL solu ion o 3- ( ime hylsilyl)p opionic-2,2,3,3-d4 acid sodium sal (TMSP) in deu e- ium oxide (D 2 O), bo h ob ained om Sigma-Ald ich (S Louis, MO, USA), se ing as he in e nal s anda d and locking sol en , espec i ely, we e added o 680 μ L o he sample. The mix u e was o exed a oom empe a u e o 5 min. The esul ing solu ion was hen ans e ed di ec ly in o he NMR ube, which was sealed and p e-wa med a 32 ◦C in an NMR ube hea e o 1 min be o e analysis. 2.2. Ins umen a ion and so wa e NMR expe imen s we e ca ied ou on a 100 MHz bench op NMR sys em (Nanalysis Co p., Albe a, Canada) equipped wi h a 24-posi ion au osample and sample hea e . Spec oscopic analysis was pe o med using he p esa u a ion echnique (P esa ) o supp ess he wa e signal a an ope a ing equency o 102.5 MHz and a spec al wid h o 1500.20 Hz J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 2 (15.0 ppm). A pulse angle o 30◦was used o exci a ion o he sample. To ensu e op imal wa e supp ession, a o al p e-sa u a ion ime o 2 s was applied p io o he acquisi ion o he NMR signal. The bandwid h o he p esa u a ion was se a 150 Hz o co e he wa e esonance wi hou a ec ing he o he signals o in e es in he spec um. A o al o eigh scans we e applied pe analysis, excep o samples labelled as skimmed, whe e 16 scans we e used. While inc easing he numbe o scans a ec s he signal in ensi y, his does no impac ou calcula ions, as hey a e ela i e o he in e nal s anda d used; he e o e, no maliza ion was applied la e o adjus he esul s. An in e scan delay ime o 30 s was used o allow comple e elaxa ion o he nuclei, ensu ing ha each scan was acqui ed unde ully elaxed condi ions. Each milk acquisi ion las ed app oxima ely 4 min unde he condi ions o analysis desc ibed abo e (8 min, app oxima ely, o skimmed sam- ples). Shimming was pe o med e e y h ee samples o ensu e a uni o m magne ic ield h oughou he se ies o measu emen s. All samples we e measu ed a 32.0 ±0.1 ◦C and unde non spin condi ions. Each spec um was manually calib a ed agains he TMSP s anda d a δ 0.0 ppm as a e e ence poin . P io o Fou ie ans o ma ion, he 1 H NMR spec a ecei ed a line b oadening applica ion o 0.3 Hz o smoo h he da a. The spec a we e au oma ically phased, baseline-co ec ed, no malized and in eg a ed using Mes ReNo a 10.0.2 so wa e (Mes e- lab Resea ch SL, San iago de Compos ela, Spain). Mul i a ia e op imiza ion was pe o med using MODDE™ P o so - wa e e sion 13.0.2 (Sa o ius AG, G¨ o ingen, Ge many), which acili- a ed a s uc u ed explo a ion o he expe imen al condi ions (Nguyen e al., 2024; Sagmeis e e al., 2020). Chemome ic analyses (PLS-R) we e pe o med wi h SIMCA 17 so wa e (Ume ics, Umea, Sweden), while he ANN model was de eloped using Py hon ( 3.11) and he Ke as lib a y ( 2.2.4) wi h a Tenso Flow backend ( 1.13.1, CUDA 10.1). 2.3. Calib a ion cu e me hod The p oposed 1 H NMR me hod o de e mine a con en was ali- da ed in e ms o linea i y (R 2 ), selec i i y, p ecision, limi o quan i i- ca ion (LOQ), accu acy ( eco e y), and wo king ange (0.06 g/100 mL, 1.52 g/100 mL, 3.48 g/100 mL and 4.33 g/100 mL) (Bachmann, 2023). In eg a ion o he signals co esponding o a was pe o med wi hin he spec al ange o 0.5–2.5 ppm. Accu acy and p ecision we e e alua ed wi h o i ied samples ( om a ailable samples a a concen a ion o 0.1 g/100 mL) a h ee con- cen a ion le els, skimmed (0.06 g/100 mL), semi-skimmed (1.5 g/100 mL) and whole (3.6 g/100 mL), wi h each o he h ee eplica es. The LOQ was es ablished a he lowes poin o he wo king ange. Impo - an ly, de e mining LOQ equi ed achie ing a signal- o-noise a io g ea e han 10, ensu ing eliable quan i ica ion (Belmon e-S´ anchez e al., 2019; FDA, 2019). 2.4. Pa ial leas squa es eg ession Fo he in es iga ion o UHT milk, an ini ial unsupe ised s a is ical app oach was applied o unco e pa e ns wi hin he da a se . P incipal Componen Analysis (PCA) was employed o explo a o y pu poses, allowing o he isualiza ion o ela ionships be ween he milk samples. Ho elling’s T 2 analysis and he Dis ance o Model in X-space (DModX) plo we e employed o de ec and emo e ou lie s, ensu ing he obus ness o he s a is ical analysis. Following his explo a o y phase, making use o supe ised ech- niques, a PLS-R model was de eloped and e alua ed using an a bi a y da a pa i ioning s a egy, in which he da a se was spli in o a aining se comp ising 80% o he da a and a alida ion se comp ising he emaining 20%. Wi hin he aining se , he model was uned using a se en- old c oss- alida ion, ensu ing ha he model was i e a i ely op imized. Du ing his p ocess, he op imal numbe o componen s in he la en dimensions (NSCs) was de e mined by moni o ing he R 2 Y and Q 2 Y alues as he numbe o NSCs inc eased. The numbe o NSCs was conside ed op imal when hese alues s abilized, meaning ha adding mo e componen s no longe signi ican ly imp o ed he model’s abili y o explain o p edic he da a. The goodness o i o he model was assessed by he de e mina ion coe icien (R 2 X) o he aining se and he explained a iance (R 2 Y), e lec ing he model’s abili y o cap u e he a iance in he da a. Addi- ionally, i s p edic i e pe o mance in he alida ion se was e alua ed using he p edic i e abili y pa ame e (Q 2 Y). Analysis o a iance coupled wi h c oss- alida ion (ANOVA-CV) p o ided a de ailed unde s anding o he model’s a iance componen s, enhancing insigh in o i s p edic i e powe and s a is ical signi icance. The oo mean squa e e o o c oss- alida ion (RMSE-CV) was calcula ed o gauge he model’s p edic i e accu acy by measu ing he a e age de ia ion o p edic ions om ac ual alues du ing c oss- alida ion, o e ing a p ecise measu e o pe o mance on unseen da a. To u he alida e he p ecision and accu acy o he esul s, he de eloped model was es ed agains o i ied samples ( es ing se ) in wa e a h ee concen a ion le els: skimmed (0.06 g/100 mL), semi- skimmed (1.5 g/100 mL), and whole (3.6 g/100 mL), wi h each con- cen a ion es ed in h ee eplica es. 2.5. A i icial neu al ne wo k (deep lea ning) Fo he quan i a i e analysis o UHT p ocessed cow’s milk, a eed- o wa d ANN was u ilized o add ess he nonlinea complexi ies o he da a. The model was ope a ed on an NVIDIA Ti an RTX GPU o imp o ed compu a ional e iciency. The a chi ec u e ea u ed an inpu laye o 21 neu ons and a hidden laye o 32 neu ons wi h Rec i ied Linea Uni (ReLU) ac i a ion, including d opou laye s a a a e o 0.2 o mi iga e o e i ing (Fig. S-1). Compiled wi h he RMSp op op imize and he mean squa ed e o loss unc ion, ANN suppo ed eg ession analysis h ough a single-neu on ou pu laye wi h linea ac i a ion. T aining in ol ed 30 epochs wi h a ba ch size o 8, and model alida ion on unseen da a assessed gene aliza ion capabili ies. RMSE-CV and mean absolu e e o (MAE) measu ed p edic i e accu acy, while lea ning cu e analysis iden i ied op imal complexi y and aining du a ion o balance pa e n ecogni ion and a oid o e i ing, ensu ing obus e alua ion and accu a e p edic ions o he UHT milk a con en . As an addi ional s ep o ensu e he eliabili y o he analysis, he pe o mance o he ANN model was assessed by applying hem o he es se o o i ied samples (see sec ion 2.3) on a g adien o a concen a- ions: skimmed (0.06 g/100 mL), semi-skimmed (1.5 g/100 mL), and whole (3.6 g/100 mL), pe o ming his es in h ee sepa a e measu e- men s o each concen a ion le el. 3. Resul s and discussion 3.1. Op imiza ion o NMR acquisi ion condi ions Conside ing he nume ous a iables in ol ed in his echnique, such as elaxa ion imes, signal supp ession conside a ions, bandwid h o supp ession, he a io o sample o sol en o he numbe o scans, among o he s, a design o expe imen s app oach has been employed o op imiza ion (Pe is-Díaz & K ę˙ zel, 2021). A cen al composi e design (CCD) was used in he esponse su ace me hodology (RSM) amewo k o op imize he pa ame e s o desk op NMR spec oscopy, ocusing on minimizing he spec al dis o ion o he speci ic spec al egion o he p edominan wa e signal a δ 4.7 ppm. The op imiza ion p ocess conside ed ou key pa ame e s: o al p esa u a ion ime, bandwid h, pulse angle, and sample-sol en a io, each wi hin speci ied anges. The o al p esa u a ion ime was a ied om 1 o 4 s, bandwid h om 30 o 200 Hz, pulse angle om 30 o 90◦, and sample-sol en a io om a minimum o 20:80 / o a maximum o 80:20 / . The p ima y objec i e was o iden i y condi ions unde which he in eg a ed a ea o a designa ed spec al egion would be as close o ze o as possible, indica ing minimal spec al dis o ion. This app oach J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 3 emphasized ha he signi ican a iable was he in eg a ed a ea quan- i a i ely measu ed, a he han a cha ac e is ic isually assessed. Gi en he numbe o a iables and he aim o a minimum numbe o deg ees o eedom o 5, wi h a es ic ion on he maximum numbe o uns se a 15, he MODDE™ so wa e ecommended a cen al composi e ace (CCF) design as he mos sui able o his s udy. Table S-2 de ails all expe imen s ca ied ou wi hin he amewo k o his s udy. The op imal condi ions iden i ied h ough his sys ema ic app oach we e a sample-sol en a io o 32:68 / , a pulse angle o 30◦, a o al p esa u a ion ime o 2 s, and a bandwid h o 150 Hz. Fig. S-2 shows he ema kable di e ences ob ained be ween di e en acquisi ion condi- ions, whe e 9 spec a ha e been ep esen ed o a oid sa u a ing he image wi h spec al o e lap. These se ings we e ound o signi ican ly educe he spec al dis o ion, he eby enhancing he quali y and eli- abili y o he NMR spec a ob ained. Rema kably, unde hese op imized condi ions, no dis o ions we e obse ed h oughou he spec um, wi h pa icula emphasis on he c i ical egion o a s in milk, which anged be ween δ 2.8 and δ 0.5 ppm (Soyle e al., 2021), and hey emained undis o ed. 3.2. Fa con en de e mina ion Fi s , o he baseline co ec ion, he Whi ake algo i hm was u i- lized, e ec i ely no malizing he baseline o he spec al signals o ze o. This app oach ensu ed ha any nega i e dis o ions encoun e ed du ing spec um alignmen did no de ac om he analysis, allowing exclusi e conside a ion o posi i e alue peaks (Cobas, 2018). Subsequen ly, selec i i y was e alua ed o iden i y po en ial in- e e ences o impu i ies ha a ise om a ia ions in he ypes o milk (cow, goa , o sheep) o be ween di e en b ands. This in ol ed comp ehensi e spec al p o iling o all inal spec a, using ce i ied UHT cow milk e e ence s anda ds o compa ison. The in es iga ion co e ed he en i e chemical shi ange ele an o he a con en o UHT milk, ensu ing ha no signal o e lap o con use wi h ei he he signals o in e es o he e e ence s anda ds. This p ocess con i med he high selec i i y o he me hod ac oss he ange o selec ed whole milk p od- uc s. Fu he mo e, he analysis was ex ended o dai y de i a i es, such as co ee wi h milk, condensed milk, c eam, ke i , and yoghu . Due o hei unique spec al p o iles, condensed milk, c eam, yoghu , and ke i we e excluded om u he modelling o p ese e he ocus and in eg i y o he analysis, hus limi ing he applica ion o he p esen s udy o he es o he 51 emaining samples (Fig. 1). Pa icula ly o skimmed milk, whe e signal esolu ion was c ucial, he use o addi ional scans signi ican ly imp o ed signal cla i y. Fo example, by doubling he numbe o scans om 8 o 16, a ma ked imp o emen in he quali y o he spec al da a was obse ed. This inc ease, al hough ex ending he analysis du a ion by app oxima ely 4 min pe sample, p o ed o be a aluable adjus men . This s a egic in- c ease in scans ep esen ed a p ac ical app oach o op imizing he analy ical p ocess, ensu ing ha e en samples wi h a low- a con en can be accu a ely quan i ied. 3.2.1. Calib a ion cu e The wo king ange o he calib a ion cu e was es ablished h ough iplica e measu emen s a ou concen a ions (0.06 g/100 mL, 1.52 g/ 100 mL, 3.48 g/100 mL and 4.33 g/100 mL) (Fig. S-3), ensu ing i encompassed he ypical le els o a con en ound in hese milk ca e- go ies h ough he in eg a ion o signals co esponding o a wi hin he spec al ange o 0.5–2.5 ppm (Soyle e al., 2021). The linea i y o he calib a ion cu e was c i ically assessed using R 2 , which e ealed excellen linea i y (R 2 =0.993) wi hin he de ined ange. Fo samples exceeding he uppe limi o he wo king ange, simple wa e dilu ion echniques could be employed, allowing o ac- cu a e analysis wi hou comp omising he me hod’s in eg i y. The p ecision and ueness esul s a e shown in Table 1. Fo semi- skimmed and whole milk, p ecision and accu acy me ics (16.2% and 9.0% o semi-skimmed, 7.9% and 3.7% o whole milk, espec i ely) demons a ed good alignmen wi h he s ingen c i e ia se (Bachmann, 2023; FDA, 2019), suppo ing he me hod’s abili y o deli e consis en and eliable esul s o hese milk a concen a ions. Howe e , he alues ob ained o skimmed milk, wi h a p ecision o 34.1% and an accu acy o 25.3%, we e signi ican ly highe . Fu he mo e, in an icipa ion o c ea ing a p edic ion ensemble o mul i a ia e models wi h samples, an analysis o hese same samples was pe o med using he calib a ion cu e o compa ison pu poses wi h samples o he models. Table 2 shows he esul s o he p edic ion se and he a iance compa ed o he labelled alues. I includes a mean pe - cen age ela i e e o (Di . (%)), which was calcula ed as (obse ed alue− ue alue)/ ue alue ×100, o 5.2% o he calib a ion cu e me hod. 3.2.2. Pa ial leas squa es eg ession Be o e PLS-R analysis, he spec a acqui ed om he UHT milk samples we e subjec ed o a bucke ing p ocess wi h a esolu ion o 0.05 ppm, esul ing in a da a se comp ising 43 dis inc a iables. This spe- ci ic bucke size was chosen on he basis o a ho ough examina ion o he esul ing spec al da a o ensu e ha c ucial in o ma ion, such as signal mul iplici y, was p ese ed. Fu he mo e, e ospec i e analyses conduc ed on a ial basis e ealed ha his bucke ing esolu ion consis en ly yielded he mos eliable and in o ma i e esul s. In a p elimina y phase, a PCA was ca ied ou on milk samples o disce n possible a ia ions be ween he g oups. Loga i hmic ans- o ma ion was used o all da a se s, ob aining alues in a smalle ange wi hou masking he e ec o small alues in he da a se and allowing an op imal a iable ange o he p oposed model (Lubes & Gooda zi, 2017). F om 153 spec a, co esponding o h ee eplica es each o 51 samples, 51 a e aged obse a ions we e ob ained o u he analyses. Using Ho elling’s T 2 and DModX analyses, 7 ou lie s (samples 5-Milk, 12-Milk, 9-Milk, 24-Milk, 36-Milk, 50-Milk, and 55-Milk) we e iden i- ied and excluded o e ine he s a is ical models. These 7 ou lie s co - esponded o 21 obse a ions, esul ing in a o al o 132 samples a e hei emo al. The PCA e ealed signi ican a iabili y in he a con en o milk and he high i o he model, wi h R 2 exceeding 98% and a Q 2 o 91% using only ou NSCs. The sco e plo con i med he ep oducibili y o he an- alyses, bols e ing con idence in he indings (Fig. S-4). The PLS-R anal- ysis iden i ied signi ican componen s o modelling he a con en in milk, leading o he c ea ion o a p edic ion cu e by plo ing p edic ed sample concen a ions agains hei known alues (Fig. 2). ANOVA-CV unde sco ed he model’s e icacy in explaining he a iance o a con en , wi h a o al sum o squa es o 43 and a eg ession sum o squa es o 39.2, highligh ing he model’s explana o y s eng h. Fig. 1. Rep esen a ion o spec al p o iles ob ained om di e en dai y de- i a i es by 1 H NMR, including whole sheep’s milk, whole cow’s milk, co ee wi h whole milk, whole goa ’s milk, condensed milk, ke i and yoghu , showing di e en ia ion o UHT milk and jus i ica ion o i s exclusion (’x’ symbol) om u he analysis. J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 4 An F-s a is ic o 216.6 and a signi ican p- alue o 1.59e-22 emphasized he ole o eg ession in explaining a iance (Table S-3). The dispa i y in s anda d de ia ions be ween he eg ession and esidual componen s u he alida ed he p ecision o he model. Collec i ely, hese esul s unde sco e he obus ness and p ecision o he PLS-R model in quan i ying he dai y a con en , showcasing i s signi ican analy ical po en ial. Wi h 44 a e aged obse a ions (35 o he aining se and 9 o he alida ion se , as indica ed in Table S-1), ou NSCs (Fig. S-5), a R 2 Y o 0.98, a Q 2 o 0.914, and a RMSE-CV o 0.39, he indings highligh he subs an ial capabili y o he model in he analysis o dai y p oduc s. Fu he mo e, he analysis o o e i ing was ein o ced by a pe mu a ion es , which yielded low alues (R 2 =0.16 and Q 2 = − 0.17), u he a i ming he obus ness o he model and i s abili y o gene alize beyond he aining da a. The esul s ob ained o he alida ion se showed a di e ence alue wi h espec o he labelling alue o −2.2%, imp o ing he alues p e iously ob ained by quan i i- ca ion o he calib a ion cu e (Table 2). The applica ion o machine lea ning PLS-R o he milk a con en in he es se a h ee concen a ion le els demons a ed p ecision and accu acy esul s ha , while a iable, ep esen ed a signi ican imp o emen o e he adi ional calib a ion cu e me hod (Table 1). The me hod demons a ed a signi ican imp o emen in he quan i i- ca ion o a con en in ela ion o he calib a ion cu e, wi h p ecision alues o 18.9%, 12.4%, and 4.8% o skimmed (0.06 g/100 mL), semi- skimmed (1.5 g/100 mL), and whole (3.6 g/100 mL) milk, espec i ely. The accu acy ac oss he es ed concen a ions was also no able, showing absolu e alues o 17.7%, 12.6%, and 2.1% o each ype o milk, espec i ely. 3.2.3. A i icial neu al ne wo k A calib a ed eed o wa d ANN was applied, using bucke ing a 0.1 ppm a e showing imp o ed esul s compa ed o he p e ious bucke ing a 0.05 ppm used in he PLS-R analysis, o NMR spec a analysis (21 a iables) o quan i y he a con en o UHT milk. The da a se , con- sis ing o 132 a e aged samples a e he ex ac ion o ou lie s, was ini ially di ided in o a aining se and a alida ion se wi h he emaining samples om he p e ious explo a o y PCA analysis a e ex ac ion o he ou lie (35 o he aining se and 9 o he alida ion se ). The p ocess in ol ed i e a i e pa ame e op imiza ion, ocusing on balancing complexi y, compu a ional e iciency, and pe o mance. Using Py hon, a g id sea ch me hod e alua ed a ious con igu a ions, iden i ying he op imal se up based on ac i a ion unc ion, ba ch size, op imize , and numbe o hidden laye s (Jiang & Xu, 2022). ReLU was selec ed o i s e iciency in main aining g adien low and lea ning complex pa e ns wi hou excessi e compu a ional demands (Sa’adah, 2023). The op imal ba ch size was de e mined o be 8, balancing compu a ional load and lea ning s abili y. RMSp op was chosen as he op imize o i s adap i e lea ning a e, which imp o es he con e gence speed. A single hidden laye p o ed su icien o he model, adhe ing o he pa simony p inciple, while se en- old c oss-- alida ion and s a egic d opou laye s comba ed o e i ing, imp o ing gene alizabili y. The chosen model con igu a ion ea u ed ReLU ac i a ion, an eigh - uni ba ch size, an RMSp op op imize , and one hidden laye , which combined heo e ical and empi ical e icacy o he quan i ica ion o a con en in UHT milk. T aining ended wi hin en epochs, indica ing apid con e gence o he model. Fig. S-6 showcases he model’s pa e n ecogni ion capabili y and p edic i e accu acy, wi h isuals illus a ing loss ends and a sca e plo o p edic ed e sus ac ual a con en . The Table 1 Compa a i e analysis o he esul s o a con en quan i ica ion in h ee eplica es o UHT milk o i ied samples using calib a ion cu e, machine lea ning and ANN me hods. a Milk ype Calib a ion cu e PLS-R ANN P ecision (%) T ueness (%) P ecision (%) T ueness (%) P ecision (%) T ueness (%) Skimmed 34.1 25.3 18.9 −17.7 14.9 −7.3 Semi-skimmed 16.2 9.0 12.4 12.6 10.2 −0.3 Whole 7.9 3.7 4.8 −2.1 6.9 −2.1 a Abb e ia ions: ANN: A i icial neu al ne wo k; PLS-R: Pa ial leas squa es eg ession. Table 2 Fa con en in he p edic ion se o comme cial UHT milk samples ob ained by calib a ion cu e, PLS-R, and ANN, including he pe cen age ela i e e o be ween he obse ed and ue alues (Di .). a ID Labelled a con en Calib a ion cu e Di . (%) PLS-R Di . (%) ANN Di . (%) 1-Milk 3.6 3.5 −3.2 3.6 0.7 3.6 −0.3 6-Milk 1.6 1.7 6.4 1.6 −3.4 1.6 0.2 15-Milk 0.3 0.4 17.3 0.3 −11.7 0.3 2.8 18-Milk 3.6 3.6 0.6 3.5 −1.9 3.6 −1.3 31-Milk 0.3 0.3 24.0 0.2 −3.2 0.3 −0.9 41-Milk 1.6 1.6 −0.3 1.5 −7.3 1.6 0.1 42-Milk 0.3 0.3 −5.3 0.3 12.6 0.3 0.2 44-Milk 3.6 3.5 −1.7 3.5 −1.9 3.5 −1.7 52-Milk 1.6 1.7 8.8 1.5 −3.7 1.6 −0.1 Mean Di . (%) –5.2 –−2.2 –0.1 a Abb e ia ions: ANN: A i icial neu al ne wo k; PLS-R: Pa ial leas squa es eg ession. Fig. 2. Compa ison o expec ed s. ac ual a con en in UHT milk samples ob ained by PLS-R. J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 5 RMSE-CV alue o 0.25 o e ed insigh in o he expec ed e o when he model was applied o no el da a wi hin his ange. The esul s ob ained o he alida ion se , oge he wi h hei di e ences wi h espec o he labelling alue, o each o he me hods applied, a e shown in Table 2. The able shows ha he bes o e all esul s we e ob ained wi h he applica ion o ANNs, wi h an a e age alue o 0.1%. Fu he mo e, he esul s ob ained by applying he ANN app oach o quan i y a con en agains he se o es s ( o i ied samples) a h ee concen a ion le els we e pa icula ly p omising (Table 1). Wi h p eci- sion alues o 14.9% o skimmed, 10.2% o semi-skimmed, and 6.9% o whole milk, oge he wi h accu acy alues o −7.3% o skimmed, −0.3% o semi-skimmed, and −2.1% o whole milk, his me hod demons a ed a signi ican imp o emen in bo h pa ame e s compa ed o p e ious me hodologies. 3.3. Quali a i e de e mina ion o lac ose con en Al hough he p ima y ocus was he de e mina ion o he a con en in milk, spec al egions a ound δ 3.1 and δ 4.0 ppm, ypically ela ed o lac ose (Soyle e al., 2021), played an unexpec edly impo an ole in he de e mina ion o lac ose. In pa icula , wi hou speci ic op imiza ion o hese egions, a bina y classi ica ion model ha could di e en ia e be ween lac ose and lac ose- ee milk was de eloped, wi h no able p elimina y success, such as he PLS Disc iminan Analysis (PLS-DA) model depic ed in Fig. 3. Despi e he ac ha he da a se was no being speci ically balanced o his pu pose, wi h only 15 lac ose- ee samples ou o a o al o 50 (a e emo ing ou lie s), he esul s clea ly indica ed a obus classi i- ca ion be ween lac ose-con aining and lac ose- ee milk. This was pa icula ly e iden in he dis inc ion be ween lac ose- ee milk and bo h semi-skimmed and whole milk, as well as be ween lac ose- ee and skimmed milk. The model’s alida ion p ocess was ca e ully designed ensu ing ha all eplica es o each sample we e ea ed oge he in each old du ing a se en- old c oss- alida ion, which minimized in a-sample a iabili y and allowed o he de ec ion o po en ial ou lie s wi hin he eplica es. In addi ion o he classi ica ion accu acy, which eached 98.67% (Table S-4), he esul s we e u he alida ed wi h Fishe ’s p obabili y es , yielding an o 2.7e-33, con i ming he s a is ical signi icance o he classi ica ion. The alida ion esul s, including he pe mu a ion es o p e en o e i ing, a e summa ized in Table S-5, u he demons a ing he model’s eliabili y. The model pe o med obus ly, and he accu acy o he classi ica ion unde sco es he s eng h o he app oach, sugges ing ha he di e en- ia ion obse ed be ween lac ose-con aining and lac ose- ee milk is eliable and signi ican . This esul opens he doo o u he in- es iga ions in o he unde lying spec al di e ences ha d i e his classi ica ion. 3.4. En i onmen al sus ainabili y analysis To u he emphasize he commi men o sus ainable p ac ices, an en i onmen al sus ainabili y analysis was inco po a ed in o he s udy, compa ed o o he es ablished me hods. This analysis was pe o med using he Analy ical G eenness (AGREE) calcula o , a ool designed o e alua e en i onmen al and occupa ional haza ds associa ed wi h analy ical p ocedu es. AGREE e alua es he “g eenness” o analy ical me hods based on he 12 p inciples o g een analy ical chemis y, p o iding a comp ehensi e o e iew o he en i onmen al impac o he p esen esea ch app oach (Pena-Pe ei a e al., 2020). The GC-MS and GC-FID me hods we e chosen o hei compa ison wi h he me hod p oposed in his s udy in e ms o analy ical g eenness (Chen e al., 2023; Danudol & Judp asong, 2022). The main cha ac- e is ics o he h ee me hods compa ed a e shown in Table S-6. In addi ion, he pic og ams ob ained by using he AGREE so wa e o e alua e hem a e ep esen ed in Fig. 4. En i onmen al sus ainabili y e ealed signi ican ly a o able esul s o he bench op NMR me hod, wi h a sco e o 0.73, compa ed o he lowe sco es achie ed by GC-MS (0.36) and GC-FID (0.33). This supe- io i y o bench op NMR was pa icula ly e iden in c i ical aspec s, such as sample p epa a ion s ages, au oma ion o minia u iza ion, and no ably in he absence o de i a iza ion and he use o oxic agen s in sample p epa a ion. The la e poin highligh ed a signi ican educ ion in was e gene a ion, demons a ing a conside ably “g eene ” en i on- men al p o ile. Howe e , he AGREE analysis highligh ed a eas o imp o emen in all he echniques analyzed, such as he need o o -line measu emen s due o he inabili y o conduc in-si u analysis, and he ene gy demands. I was impo an o men ion ha AGREE calcula o did no dis inguish be ween bench op and con en ional c yogen-cooled supe conduc ing elec omagne NMR sys ems. As such, we conside ed ha he ac ual ene gy consump ion o bench op sys ems would ha e been lowe due o hei use o pe manen magne s. Ne e heless, e en in he wo s -case scena io, whe e he highe ene gy consump ion o con- en ional sys ems was assumed, he bench op NMR me hod s ill demons a ed a highly a o able en i onmen al p o ile. 4. Conclusions This s udy p esen s, o he i s ime, he simul aneous use o h ee di e en quan i ica ion models o de e mine a con en in UHT milk samples using bench op NMR. This compa ison p o ided a be e un- de s anding o he di e ences be ween adi ional calib a ion cu e applica ions, and PLS-R and ANNs based machine lea ning algo i hms o es ima ing a con en in skimmed, semi-skimmed and whole milk. In analyzing he mos challenging cases, such as skimmed milk, he ANN me hod exhibi ed be e p ecision and accu acy (−14.9%, 7.3%) compa ed o bo h he calib a ion cu e me hod (34.1%, 25.3%) and he PLS-R app oaches (18.9%, −17.7%). This end in a quan i ica ion was again obse ed, wi h ANN consis en ly achie ing be e esul s in semi- skimmed (p ecision 10.2% and accu acy −0.3%) and whole milk Fig. 3. Example o he spec al analysis ob ained by bench op NMR in he a ge egion (le ) and he bina y classi ica ion PLS-DA ( igh ) o lac ose and lac ose- ee milk. J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 6 (p ecision 6.9% and accu acy −2.1%). All p oposed me hods ha e been adequa ely alida ed, demons a ing sui able alues o selec i i y, linea i y wi hin he wo king ange, p ecision and accu acy. Rega ding sample analysis, he compa a i e analysis o he me hods demons a ed a commendable le el o accu acy o skimmed, semi- skimmed, and whole milk, wi h he o e all a e age di e ences om he labelled alues being 5.2%, −2.2%, and 0.1%, espec i ely. This accu acy unde sco ed he po en ial o hese me hodologies o mee he igo ous demands o quali y con ol wi hin he dai y indus y. The ex ension o he s udy o goa , sheep and co ee- la ou ed milk u he a es s o he e sa ili y and eliabili y o he me hods in accu a ely de e mining he a con en , co esponding closely wi h he p oduc labels. Addi ionally, he applica ion o machine lea ning o ull spec a, a he han speci ic egions, has shown po en ial o bina y classi ica ion models such as PLS-DA, success ully dis inguishing be ween lac ose and lac ose- ee milk, sugges ing b oade applica ions o u u e esea ch. In ligh o compa a i e analysis wi h es ablished echniques such as GC-MS and GC-FID o quan i ying a con en in milk samples, he AGREE analysis highligh ed he excep ional g eenness sco e o 0.73 o he bench op NMR me hod, showcasing a obus commi men o en i- onmen ally iendly esea ch p ac ices. This compa ison showed he signi ican sus ainabili y ad an ages o bench op NMR o e hese adi ional echniques, wi h bene i s including educed sample p epa- a ion s ages, enhanced au oma ion and minia u iza ion, and he elim- ina ion o oxic eagen s and de i a iza ion p ocesses. By ma kedly ou pe o ming es ablished echniques in e ms o en i onmen al impac , bench op NMR cemen s i s leading posi ion in p omo ing a g eene app oach o quan i ying a con en in milk. CRediT au ho ship con ibu ion s a emen Jos´ e Raúl Belmon e-S´ anchez: Valida ion, So wa e, Me hodology, In es iga ion, Fo mal analysis, Da a cu a ion, Concep ualiza ion. Rob- e o Rome o-Gonz´ alez: W i ing – e iew & edi ing, Visualiza ion, Su- pe ision, Concep ualiza ion. Manuel ´ Angel Ma ínez O osa: W i ing – e iew & edi ing, Fo mal analysis. Ma ía Cal o Mo a a: W i ing – e iew & edi ing, Resou ces, P ojec adminis a ion, Funding acquisi- ion. An onia Ga ido F enich: W i ing – e iew & edi ing, P ojec adminis a ion, Me hodology, Funding acquisi ion, Concep ualiza ion. Decla a ion o compe ing in e es The au ho s decla e ha hey ha e no known compe ing inancial in e es s o pe sonal ela ionships ha could ha e appea ed o in luence he wo k epo ed in his pape . Acknowledgemen s G an p ojec e . CPP2021-008672 unded by MICIU/AEI/ 10.13039/501100011033 and by he “Eu opean Union Nex Gene - a ionEU/PRTR”. Appendix A. Supplemen a y da a Supplemen a y da a o his a icle can be ound online a h ps://doi. o g/10.1016/j.lw .2024.117000. Da a a ailabili y The da a ha has been used is con iden ial. Re e ences Ama al, J. S., Ma a, I., Pissa d, A., Fe n´ andez Pie na, J. A., & Bae en, V. (2018). Milk and milk p oduc s. In FoodIn eg i y handbook (pp. 1–23). Eu o ins Analy ics F ance. h ps://doi.o g/10.32741/ ihb.1.milk. Aquini, S., & Gil, J. M. (2017). Inno a ion and b and e ec s on he consume s’ demand o esh milk in Spain. Ag icul u al and Applied Economics Associa ion, 258131, 1–30. h ps://doi.o g/10.22004/ag.econ.258131 Bachmann, R. (2023). Guide o NMR me hod de elopmen and alida ion-Pa I: Iden i ica ion and quan i ica ion (upda e 2023). h ps://doi.o g/10.13140 /RG.2.2.30200.83208. Belmon e-S´ anchez, J. R., Aguile a-S´ aez, L. M., Rome o-Gonz´ alez, R., Ma ínez Vidal, J. L., A ebola, F. J., & Ga ido F enich, A. (2019). De e mina ion o e id onic acid in ege able-washing wa e by a simple and alida ed quan i a i e 31P nuclea magne ic esonance me hod. Mic ochemical Jou nal, 150, A icle 104083. h ps:// doi.o g/10.1016/j.mic oc.2019.104083 Benzaama, M. H., Menhoudj, S., Mokh a i, A. M., & Lachi, M. (2022). Compa a i e s udy o he he mal pe o mance o an ea h ai hea exchange and seasonal s o age sys ems: Expe imen al alida ion o A i icial Neu al Ne wo ks model. Jou nal o Ene gy S o age, 53, A icle 105177. h ps://doi.o g/10.1016/j.es .2022.105177 Bha i, S. K., & Roy, R. (2012). Quan i a i e 1 H NMR spec oscopy. T AC, T ends in Analy ical Chemis y, 35, 5–26. h ps://doi.o g/10.1016/j. ac.2012.02.007 Bu ge , R., Lindne , S., Rump , J., Do, X. T., Diehl, B. W. K., Rehahn, M., Monakho a, Y. B., & Schulze, M. (2022). Bench op e sus high ield NMR: Compa able pe o mance ound o he molecula weigh de e mina ion o lignin. Jou nal o Pha maceu ical and Biomedical Analysis, 212, A icle 114649. h ps://doi. o g/10.1016/j.jpba.2022.114649 Chen, M., Wang, F., Wu, X., Si, B., Pan, J., Zheng, N., Zhang, Y., & Wang, J. (2023). Upda ing he a y acid p o iles o e ail bo ine milk in China based on an imp o ed GC-MS me hod: Implica ions o nu i ion. F on ie s in Nu i ion, 10. h ps://doi.o g/ 10.3389/ nu .2023.1204005 Cobas, C. (2018). Applica ions o he Whi ake smoo he in NMR spec oscopy. Magne ic Resonance in Chemis y, 56(12), 1140–1148. h ps://doi.o g/10.1002/m c.4747. John Wiley and Sons L d. Co sa o, C., Vasi, S., Ne i, F., Mezzasalma, A. M., Ne i, G., & Fazio, E. (2022). NMR in Me abolomics: F om con en ional s a is ics o machine lea ning and neu al ne wo k app oaches. Applied Sciences, 12(6), 2824. h ps://doi.o g/10.3390/app12062824 da Cos a, N. L., da Cos a, M. S., & Ba bosa, R. (2021). A e iew on he applica ion o chemome ics and machine lea ning algo i hms o e alua e bee au hen ica ion. Food Analy ical Me hods, 14(1), 136–155. h ps://doi.o g/10.1007/s12161-020- 01864-7 Danudol, A., & Judp asong, K. (2022). De elopmen and me hod alida ion o bu y ic acid and milk a analysis in bu e blends and blended milk p oduc s by GC-FID. Foods, 11(22), 3606. h ps://doi.o g/10.3390/ oods11223606 Fig. 4. Resul s o he AGREE analysis o en i onmen al sus ainabili y o he bench op NMR me hod (A) compa ed o he GC-MS (B) and GC-FID (C) me hods. J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 7 Debik, J., Sange mani, M., Wang, F., Madssen, T. S., & Giskeødegå d, G. F. (2022). Mul i a ia e analysis o NMR-based me abolomic da a. NMR in Biomedicine, 35(2). h ps://doi.o g/10.1002/nbm.4638. John Wiley and Sons L d. Dhaulaniya, A. S., Balan, B., Sodhi, K. K., Kelly, S., Canna an, A., & Singh, D. K. (2020). Quali a i e and quan i a i e e alua ion o co n sy up as a po en ial added swee ene in apple ui juices using mid-in a ed spec oscopy assis ed chemome ic modeling. LWT, 131, A icle 109749. h ps://doi.o g/10.1016/j.lw .2020.109749 D ape , S. L., & McCa ney, E. R. (2023). Bench op nuclea magne ic esonance spec oscopy in o ensic chemis y. Magne ic Resonance in Chemis y, 61(2), 106–129. h ps://doi.o g/10.1002/m c.5197 E angelis a, C., Basi ic` o, L., & Be nabucci, U. (2021). An o e iew on he use o nea in a ed spec oscopy (ni s) on a ms o he managemen o dai y cows. Ag icul u e, 11(4), 296. h ps://doi.o g/10.3390/ag icul u e11040296 Ezeanaka, M. C., Nso -A indana, J., & Zhang, M. (2019). Online low- ield nuclea magne ic esonance (LF-NMR) and magne ic esonance imaging (MRI) o ood quali y op imiza ion in ood p ocessing. Food and Biop ocess Technology, 12(9), 1435–1451. h ps://doi.o g/10.1007/s11947-019-02296-w. Sp inge New Yo k LLC. FDA. (2019). Guidelines o he alida ion o chemical me hods in ood, eed, cosme ics, and e e ina y p oduc s. Guidelines o he Valida ion o Chemical Me hods in Food, Feed, Cosme ics, and Ve e ina y P oduc s, 23. h ps:// da. epo /media/121751/Valid a ionNucleicAcidSequenceBasedAnalysisFoodFeedCosme icsVe e ina y.pd . Gal an, D., Tanama i, A. A. C., Casano a, F., Danieli, E., Bona, E., & Killne , M. H. M. (2021). Compac low- ield NMR spec oscopy and chemome ics applied o he analysis o edible oils. Food Chemis y, 365, A icle 130476. h ps://doi.o g/ 10.1016/j. oodchem.2021.130476 Ha zakis, E. (2019). Nuclea magne ic esonance (NMR) spec oscopy in ood science: A comp ehensi e e iew. Comp ehensi e Re iews in Food Science and Food Sa e y, 18(1), 189–220. h ps://doi.o g/10.1111/1541-4337.12408 Hou, X., Wang, G., Su, G., Wang, X., & Nie, S. (2019). Rapid iden i ica ion o edible oil species using supe ised suppo ec o machine based on low- ield nuclea magne ic esonance elaxa ion ea u es. Food Chemis y, 280, 139–145. h ps://doi. o g/10.1016/j. oodchem.2018.12.031 Hu, F., Fu iha a, K., Ka o, Y., & Tanoku a, M. (2007). Nondes uc i e quan i ica ion o o ganic compounds in whole milk wi hou p e ea men by wo-dimensional nm spec oscopy. Jou nal o Ag icul u al and Food Chemis y, 55(11), 4307–4311. h ps:// doi.o g/10.1021/j 062803x Huang, W., Fan, D., Li, W., Meng, Y., & Liu, T. C. (2022). Rapid e alua ion o milk acidi y and iden i ica ion o milk adul e a ion by Raman spec oscopy combined wi h chemome ics analysis. Vib a ional Spec oscopy, 123, A icle 103440. h ps://doi. o g/10.1016/j. ibspec.2022.103440 Jiang, X., & Xu, C. (2022). Deep lea ning and machine lea ning wi h g id sea ch o p edic la e occu ence o b eas cance me as asis using clinical da a. Jou nal o Clinical Medicine, 11(19), 5772. h ps://doi.o g/10.3390/jcm11195772 Joshi, P. B. (2023). Na iga ing wi h chemome ics and machine lea ning in chemis y. A i icial In elligence Re iew, 56(9), 9089–9114. h ps://doi.o g/10.1007/s10462- 023-10391-w Lubes, G., & Gooda zi, M. (2017). Analysis o ola ile compounds by ad anced analy ical echniques and mul i a ia e chemome ics. Chemical Re iews, 117(9), 6399–6422. h ps://doi.o g/10.1021/acs.chem e .6b00698 Monakho a, Y. B., Kuballa, T., Lei z, J., Andlaue , C., & Lachenmeie , D. W. (2012). NMR spec oscopy as a sc eening ool o alida e nu i ion labeling o milk, lac ose- ee milk, and milk subs i u es based on soy and g ains. Dai y Science & Technology, 92 (2), 109–120. h ps://doi.o g/10.1007/s13594-011-0050-5 Nguyen, H.-M., Duong, T.-K., Nguyen, V.-K., Nguyen, T.-K.-L., Dong, T.-H.-Y., Nguyen, C.-H., & Tung, N.-T. (2024). A wo-s ep design o expe imen s app oach o in es iga e he simul aneous e ec s o ion-pai ing and chemical enhance s o imp o e he pe meabili y o lo noxicam in a opical hyd ogel pa ch. Jou nal o Pha maceu ical In es iga ion, 54(2), 229–247. h ps://doi.o g/10.1007/s40005-023- 00660-9 Pena-Pe ei a, F., Wojnowski, W., & Tobiszewski, M. (2020). Ag ee - analy ical GREEnness me ic app oach and so wa e. Analy ical Chemis y, 92(14), 10076–10082. h ps://doi.o g/10.1021/acs.analchem.0c01887 Pe is-Díaz, M. D., & K ę˙ zel, A. (2021). A guide o good p ac ice in chemome ic me hods o ib a ional spec oscopy, elec ochemis y, and hyphena ed mass spec ome y. T AC, T ends in Analy ical Chemis y, 135, A icle 116157. h ps://doi.o g/10.1016/j. ac.2020.116157 Pomyen, Y., Wanich hana ak, K., Poungsomba , P., Fah mann, J., G apo , D., & Khoom ung, S. (2020). Deep me abolome: Applica ions o deep lea ning in me abolomics. Compu a ional and S uc u al Bio echnology Jou nal, 18, 2818–2825. h ps://doi.o g/10.1016/j.csbj.2020.09.033 Reine , J., P o e, K., & Hin ichs, J. (2020). In es iga ion o he applicabili y o Raman spec oscopy as online p ocess con ol du ing consume milk p oduc ion. ChemEnginee ing, 4(3), 45. h ps://doi.o g/10.3390/chemenginee ing4030045 Sa’adah, A. (2023). A i icial in elligence wi h py hon. Technome ics, 65(3), 451–452. h ps://doi.o g/10.1080/00401706.2023.2237826 Sagmeis e , P., Poms, J., Williams, J. D., & Kappe, C. O. (2020). Mul i a ia e analysis o inline bench op NMR da a enables apid op imiza ion o a complex ni a ion in low. Reac ion Chemis y & Enginee ing, 5(4), 677–684. h ps://doi.o g/10.1039/ d0 e00048e San os, M. C., Nascimen o, P. A. M., Guedes, W. N., Pe ei a-Filho, E. R., Fille i, ´ E. R., & Pe ei a, F. M. V. (2019). Chemome ics in analy ical chemis y - an o e iew o applica ions om 2014 o 2018. Ecle ica Quimica, 44(2), 11–25. h ps://doi.o g/ 10.26850/1678-4618eqj. 44.2.11-25. A lan is Li os L da. San os, P. M., Pe ei a-Filho, E. R., & Colnago, L. A. (2016). De ec ion and quan i ica ion o milk adul e a ion using ime domain nuclea magne ic esonance (TD-NMR). Mic ochemical Jou nal, 124, 15–19. h ps://doi.o g/10.1016/j.mic oc.2015.07.013 Sch¨ onbe ge , T., Monakho a, Y. B., Lachenmeie , D. W., Walch, S., Kuballa, T., Team, N.- P. E., & Ge many, (NEXT) -NMR wo king g oup. (2015). EUROLAB Technical Repo 01/2015 -Guide o NMR me hod de elopmen and alida ion-Pa II: Mul i a ia e da a analysi. 01, 1–20. h p://www.eu olab.o g/documen s/NMRValGuidelineIIV6.pd . Sø ensen, M. K., Balsga , N. M., Beye , M., Jensen, O. N., & Nielsen, N. C. (2022). On- Si e measu emen o a and p o ein con en s in milk using mobile NMR echnology. Molecules, 27(3), 583. h ps://doi.o g/10.3390/molecules27030583 Soyle , A., Cik ikci, S., Ca da oglu, C., Bouillaud, D., Fa jon, J., Gi audeau, P., & Oz op, M. H. (2021). Mul i-scale bench op 1H NMR spec oscopy o milk analysis. LWT, 139, A icle 110557. h ps://doi.o g/10.1016/j.lw .2020.110557 Tamime, A. Y. (Ed.). (2008). Milk p ocessing and quali y managemen . Wiley. h ps://doi. o g/10.1002/9781444301649. Taylo , B. J. (2006). Me hods and p ocedu es o he e i ica ion and alida ion o a i icial neu al ne wo ks. Kluwe Academic Publishe s. h ps://doi.o g/10.1007/0-387- 29485-6 an Beek, T. A. (2021). Low- ield bench op NMR spec oscopy: S a us and p ospec s in na u al p oduc analysis. Phy ochemical Analysis, 32(1), 24–37. h ps://doi.o g/ 10.1002/pca.2921 Zhao, L.-L., Qiu, X.-J., Wang, W.-B., Li, R.-M., & Wang, D.-S. (2019). NMR me abolomics and andom o es s models o iden i y po en ial plasma bioma ke s o blood s asis synd ome wi h co ona y hea disease pa ien s. F on ie s in Physiology, 10. h ps:// doi.o g/10.3389/ phys.2019.01109 J.R. Belmon e-S´ anchez e al. LWT 212 (2024) 117000 8