PLS-DA Model Research Articles

From Spectra to Signatures: Detecting Fentanyl in Human Nails with ATR-FTIR and Machine Learning.

Human nails have recently become a sample of interest for toxicological purposes. Multiple studies have proven the ability to detect various analytes within the keratin matrix of the nail. The analyte of interest in this study is fentanyl, a highly dangerous and abused drug in recent decades. In this proof-of-concept study, ATR-FTIR was combined with machine learning methods, which are effective in detecting and differentiating fentanyl in samples, to explore whether nail samples are distinguishable from individuals who have used fentanyl and those who have not. PLS-DA and SVM-DA prediction models were created for this study and had an overall accuracy rate of 84.8% and 81.4%, respectively. Notably, when classification was considered at the donor level-i.e., determining whether the donor of the nail sample was using fentanyl-all donors were correctly classified. These results show that ATR-FTIR spectroscopy in combination with machine learning can effectively differentiate donors who have used fentanyl and those who have not and that human nails are a viable sample matrix for toxicology.

Identification and discrimination of human keratinized tissues using ATR-FTIR and chemometrics.

In forensic investigations, human keratinized tissues like skin and nails are commonly encountered as trace evidence, yet the use of vibrational spectroscopy for their identification and differentiation has been underexplored. This research utilized ATR-FTIR to distinguish between human nails and skin samples collected from a group of 50 participants, employing advanced chemometric analysis techniques. The spectral signatures of human keratinized tissues, such as nails and skin, exhibit similarities consistent with previous studies. Chemometric analysis aimed at distinguishing these tissues showed that the PLS-DA model achieved an overall accuracy of 67 % with an AUC score of 0.65, while the SVM model had an overall accuracy of 56 % with an AUC score of 0.71. For sex identification, the PLS-DA model demonstrated an overall accuracy of 83 % with an AUC value of 1, whereas the SVM model achieved an overall accuracy of 100 % with an AUC score of 1. The study underscores the potential of ATR-FTIR coupled with chemometrics in the precise identification and differentiation of human keratinized tissue, thereby enhancing the capabilities of forensic investigations.

Non-invasive diagnosis of papillary thyroid microcarcinoma using a novel metabolomics analysis of urine.

The severity of thyroid cancer is judged on the basis of histologic and clinical features. A limited number of studies have considered urinary metabolite signatures for its diagnosis, and no reliable urinary metabolite biomarkers have been proposed. This diagnostic method would be particularly valuable because of its non-invasive nature. A nuclear magnetic resonance (NMR)-based metabolomics approach was used as the analytical platform to study the urine samples of patients with PTMC. Urine samples collected from 41 PTMC patients, 52 healthy subjects, and 13 patients with benign tumors were analyzed using 1H-NMR spectroscopy to identify metabolic changes. PLS-DA, or partial least squares discriminant analysis, was used to analyze the NMR spectra. A double cross-validation method and randomization tests were used to validate PLS-DA models. Clear discriminations between PTMC patients and healthy controls, as well as between PTMC patients and patients with benign tumors were obtained. Collectively, pi-methyhistidine, trimethylamine, myo-inositol, acetate, suberate, azelate, mannitol, tau-methylhistine, ascorbate, 3-aminoisobutyric acid, 2-oxoglutarate, and methanol contributed to the discrimination. Apart from myo-inositol and methanol, all of these metabolites exhibited increased levels in the urine samples of PTMC patients as compared to that of patients with benign tumors. The application of this NMR-based metabolomics approach allowed the detection of anomalous metabolic traits directly connected PTMC, potentially yielding a more sensitive and comprehensive diagnostic results for PTMC.

Metabolomic study for the identification of symptomatic carotid plaque biomarkers

Carotid artery stenosis is mainly produced due to the progressive accumulation of atherosclerotic plaque in the vascular wall. The atherosclerotic plaque is characterized by the accumulation of lipids, low density proteins, expression of chemokines and adhesion molecules, and migration of monocytes and lymphocytes into the plaque. Its rupture can produce stroke, but embolic propensity depends principally on the composition and vulnerability of plaque rather than the severity of stenosis. It is important, then, to ascertain which patients with carotid artery stenosis have a greater risk of developing neurological symptomatology. Here, we present a metabolomic study by using nuclear magnetic resonance (NMR) spectroscopy in atheroma plaque and serum samples from patients with recently symptomatic and asymptomatic carotid stenosis to search for metabolites that could be used as biomarkers associated with plaque vulnerability and subsequent risk of rupture. Thirty-eight atheromatous plaque samples (24 asymptomatic patients and 14 symptomatic) and 70 serum samples (43 asymptomatic and 27 symptomatic) were studied by NMR spectroscopy. The data were analysed using multivariate statistics (PLS-DA) to determine a model to discriminate between symptomatic and asymptomatic samples (atheroma plaques and sera). The calculated PLS-DA models showed a 100 % sensitivity and a 96.6 % specificity for the cross validation to discriminate between symptomatic and asymptomatic plaques, and 88.37 % sensitivity and 77.78 % specificity when serum samples were analysed. According to the results of our multivariate and univariate analysis, the most discriminative metabolites for plaque vulnerability were threonine in serum samples, and glutamate in plaque samples. Also, an analysis of the main metabolic pathways involved in plaque vulnerability revealed that d-glutamine and d-glutamate metabolism, and phenylalanine, tyrosine, and tryptophan biosynthesis were the most affected pathways in plaque and serum, respectively.

Attenuated Total Reflection Fourier Transform Infrared Spectroscopy and Chemometrics for the Discrimination of Animal Hair Fibers for the Textile Sector.

Analyzing the composition of animal hair fibers in textiles is crucial for ensuring the quality of yarns and fabrics made from animal hair. Among others, Fourier transform infrared (FT-IR) spectroscopy is a technique that identifies vibrations associated with chemical bonds, including those found in amino acid groups. Cashmere, mohair, yak, camel, alpaca, vicuña, llama, and sheep hair fibers were analyzed via attenuated total reflection FT-IR (ATR FT-IR) spectroscopy and scanning electron microscopy techniques aiming at the discrimination among them to identify possible commercial frauds. ATR FT-IR, being a novel approach, was coupled with chemometric tools (partial least squares discriminant analysis, PLS-DA), building classification/prediction models, which were cross-validated. PLS-DA models provided an excellent differentiation among animal hair of both camelids and eight animal species. In addition, the combination of ATR FT-IR and PLS-DA was used to discriminate the cashmere hair from different origins (Afghanistan, Australia, China, Iran, and Mongolia). The model showed very good discrimination ability (accuracy 87%), with variance expression of 94.88% and mean squared error of cross-validation of 0.1525.

The Pattern of Gene Expression (Igf Family, Muscle Growth Regulatory Factors, and Osteogenesis-Related Genes) Involved in the Growth of Skeletal Muscle in Pikeperch (Sander lucioperca) During Ontogenesis.

The pikeperch (Sander lucioperca) is an economically important freshwater fish and a valuable food with high market acceptance. It is undergoing important changes in growth and regulatory metabolism during the ontogeny. Hence, the current study aims to investigate the mRNA expression of the growth hormone (gh)/insulin-like growth factor (igf) axis (ghr, igfI, igfbp, igfr), muscle regulatory factors (pax7, myf5, myod, myogenin, mrf, mymk, mstn), and osteogenesis-related genes (colla1a, fib1a, on, op, ostn) from hatching through day 40th post-hatching (DPH). The average total length (TL) of larvae measured at hatching was 3.6 ± 0.4 mm (67 degree days), and at the end of the experiment (40 DPH, 777 degree days), it was 27.1 ± 1.1 mm. The results showed three phases of gene expression in day 0 (egg), larval, and juvenile stages of pikeperch, which can be a progression or transition from the initial state toward the juvenile state. The expression pattern of myf5, mymk, and fib1a genes showed the highest value at day 0. The growth hormone receptor gene (ghr) and igfbp5 were raised to 1 DPH, whereas increased expression of igfI, igfII, igf1bp4, igf1rb, myod2, and mrf4 was detected at 14 DPH. The myod1, pax7, op, ostc, on, igf1ra, and col1a1a genes were highly expressed at 21 DPH and juvenile stages. According to the PLS-DA model, the most relevant VIPs are myf5 and mymk as best markers of earlier stages and igf1ra, ostc, pax7, and ghr as markers of later stages of ontogeny. Results from this study suggest that basal metabolism, growth of body cells and muscles, and bone proliferation and development can be regulated by the dynamic changes in gene expression patterns in this species. The identified genes will help to understand the basic biological process of pikeperch larvae and development, which is very important in pikeperch farming.

Ultrasound and its coupled oak blocks treatment based on absorbed energy density for comprehensive insight and scale-up consideration of phenolic and astringency profiles in red wines

Ultrasound and ultrasound-coupled oak blocks techniques on the phenolic and astringency profiles of Petit Verdot wines were discussed in this study. The relationship between techniques at varying absorbed energy density (AED) gradients and astringency was revealed, elucidating potential molecular mechanisms regarding compound interactions. Ultrasound was found to promote phenolics degradation and condensed tannins maturation, while oak blocks improved ellagitannins release into wines, facilitating polycondensation reactions to form larger complexes. Phenolics binding to salivary proteins decreased at low AED values (0–33.07 J/mL), enhanced at high AED values (66.14–165.34 J/mL), and reduced salivary proteins' precipitation ability. Treated samples significantly altered astringency's global terminology and sensory attributes across oral regions, with oak blocks enriching the astringency hierarchy. Multivariate analysis identified an optimal ultrasound AED of 0 to 33.07 J/mL, and an effective PLS-DA model was developed for industrial scale-up considerations, aiming to cost-effectively enhance the organoleptic quality of wines.

Electrochemical sensors for fast classification of different Cannabis sativa L. samples according to total Δ9-tetrahydrocannabinol content

In this work, we investigated the ability of an electrochemical sensor to recognize Cannabis sativa L. samples with different total content of Δ9-tetrahydrocannabinol (Δ9-THC), determined by the levels of the psychoactive cannabinoid and of its biosynthetic precursor Δ9-tetrahydrocannabinolic acid (Δ9-THCA), using a multivariate approach. The voltammetric responses recorded with screen-printed electrodes modified with carbon black reflected the compositional differences from the different samples, in terms of cannabinoids of the vegetal material. PLS-DA models allowed for the correct classification of most C. sativa samples into the classes of legal and illegal samples according to total Δ9-THC content, based on threshold limits defined by the EU/US (0.3 % w/w) and Italian (0.6 % w/w) regulations. Satisfactory results were achieved in both cases, obtaining classification efficiency values in prediction of the external test set equal to 85 % and 100 % for the EU/US and Italian thresholds, respectively. The obtained results suggest the possibility to consider the proposed method as a starting point for the implementation of an automated device for rapid prescreening of total Δ9-THC content directly on site.

Rapid and accurate identification of Gastrodia elata Blume species based on FTIR and NIR spectroscopy combined with chemometric methods

Different varieties of Gastrodia elata Blume (G. elata Bl.) have different qualities and different contents of active ingredients, such as polysaccharide and gastrodin, and it is generally believed that the higher the active ingredients, the better the quality of G. elata Bl. and the stronger the medicinal effects. Therefore, effective identification of G. elata Bl. species is crucial and has important theoretical and practical significance. In this study, first unsupervised PCA and t-SNE are established for data visualisation, follow by traditional machine learning (PLS-DA, OPLS-DA and SVM) models and deep learning (ResNet) models were established based on the fourier transform infrared (FTIR) and near infrared (NIR) spectra data of three G. elata Bl. species. The results show that PLS-DA, OPLS-DA and SVM models require complex preprocessing of spectral data to build stable and reliable models. Compared with traditional machine learning models, ResNet models do not require complex spectral preprocessing, and the training and test sets of ResNet models built based on raw NIR and low-level data fusion (FTIR + NIR) spectra reach 100 % accuracy, the external validation set based on low-level data fusion reaches 100 % accuracy, and the external validation set based on NIR has only one sample classification error and no overfitting.

Detection of physical hazards from fruit processed products using hyperspectral imaging and prediction based on PLS-DA and logistic regression machine learning models

The current study aims to investigate a prediction model from reflection values and the spectral angle mapper (SAM) obtained from hyperspectral imaging (HSI) technology for the detection of three types of foreign materials – a branch, a knife, and rubber – that are associated with the various fruit-processed products. The study found that the maximum and minimum reflection values in the 900 to 1700 nm range for juices (apple, grape, tomato) and jams (strawberry, peach, tomato) differed depending on the presence or absence of physical hazards. The presence of physical hazards was also confirmed by the color difference in the SAM image. The partial least squares discriminant analysis model (PLS-DA) and logistic regression implemented on the Jupyter Notebook platform through the Anaconda prompt provided accuracy, F1 score, specificity, and sensitivity based on the confusion matrix. The maximum value was 100.0 %, while the minimum value was 97.6 % of the result of the PLS-DA modeling in the testing set. Logistic regression modeling also had a similar result: the maximum value was 100.0 %, while the minimum value was 97.8 % in the testing set. The sensitivity value in the testing set, which is a meaningful result for detecting physical hazards, was a maximum of 100.0 % for both PLS-DA and logistic regression. Results from the current study suggest that the reflection value and SAM data obtained through hyperspectral imaging could build a big data platform for early determination of physical hazards during agricultural product processing.

Authentication of premium Asian rice varieties: Stable isotope ratios and multi-elemental content for the identification of geographic fingerprints

Over 90 percent of the world's rice is produced and consumed in the Asia-Pacific region. Varieties such as Thai Jasmine rice and Paw San (or “Myanmar pearl rice”) are globally recognised as premium, while more local high-grade varieties include the Indonesian Ciherang and Inpari. Being able to trace the origin of these products has become necessary, since they are marketed at relatively higher prices compared to other cultivars, and they often become the target of fraudulent activities. In this work, we aimed to identify variables that could distinguish the premium-producing regions within each country, by Isotope Ratio Mass Spectrometry (IRMS) and Inductively Coupled Plasma- Mass Spectrometry (ICP-MS). Low-Level Data Fusion (LLDF) followed the analysis of more than 300 authentic samples, and (O)PLS-DA models yielded very high accuracy values. The most important geo-differentiating variables (VIP>1.4) were identified as: δ13C, δ18O, δ2H, δ34S, Co, Rb, Cu, Ba and Zn.

Discrimination between Covid-19 positive and negative blood serum based on excitation-emission matrix fluorescence spectroscopy and chemometrics

The outbreak of the disease caused by the SARS-CoV-2 virus (Covid-19) has resulted in a global health emergency that has caused millions of deaths in recent years. The control of the pandemic was significantly impacted by the availability of inputs and qualified labor to correctly diagnose the population. The challenges faced by numerous countries in conducting this extensive diagnosis, utilizing methods such as RT-PCR, emphasize the necessity for alternative testing strategies that are less reliant on expensive raw materials and can be implemented on a larger scale. This paper proposes a methodology for classifying blood serum samples as either positive or negative for Covid-19 infection using excitation-emission matrix (EEM) fluorescence spectroscopy associated with multivariate analysis. The proposed methodology uses EEM spectra of samples diagnosed by the reference method (RT-PCR) to train and validate classification models. Two approaches were tested: the first using PARAFAC and the second by unfolding the excitation-emission matrices. The DD-SIMCA model performed best in the PARAFAC approach, with an error rate of 0.05, sensitivity of 0.98 and specificity of 0.96. The PLS-DA and PCA-DA models in the second approach effectively distinguished between classes. The PCA-DA model performed the best with an error rate of 0.06 and sensitivity and specificity of 0.94. Fluorescence spectroscopy was found to be effective in analyzing serum samples and obtaining discrimination models to determine if a patient is infected with SARS-CoV-2. The findings are encouraging and could aid in the development of an inexpensive and reliable auxiliary diagnostic method.

Chemometrics and neural networks for estimating the chilling injury severity of kimchi cabbage (Brassica rapa L. ssp. pekinensis) based on hyperspectral images

The objective of this study was to develop prediction models for the severity of chilling injury in kimchi cabbages (KCs) using short-wave infrared hyperspectral imaging (HSI) combined with supervised classifications (partial least square discriminant analysis, PLS-DA; and artificial neural network, ANN) at early stages. Classifications were implemented for the two-class (i.e., normal and chilled), three-class (i.e., normal, semi, and severe) and four-class (i.e., normal, slight, moderate, and severe) classifications at full and selected optimal wavelengths (associated with the C–H stretch of sugar and cellulose). The PLS-DA classifier with Savitzky-golay (SG) preprocessing showed improved average classification accuracy of 0.912 and 0.896 for three-class and four-class classification developed with optimal wavelength, respectively. The best classification results based on ANN using optimal wavelengths and SG achieved overall accuracies above 0.924 for all-class classifications. ANN models have higher potential and are simplified to predict chilling damage characteristics of KCs compared to PLS-DA models. By visualizing injured areas, we showed the association between the internal changes as detected by HSI. These results demonstrate the capability of HSI and the central role of chemometrics and neural networks in developing efficient models for prediction of chilling damage properties of KCs.

An Integrated Approach Based on Clinical Data Combined with Metabolites and Biomarkers for the Assessment of Post-Operative Complications after Cardiac Surgery.

Background: Early diagnosis of post-operative complications is an urgent task, allowing timely prescribing of appropriate therapy and reducing the cost of patient treatment. The purpose of this study was to determine whether an integrated approach based on clinical data, along with metabolites and biomarkers, had greater predictive value than the models built on fewer data in the early diagnosis of post-operative complications after cardiac surgery. Methods: The study included patients (n = 62) admitted for planned cardiac surgery (coronary artery bypass grafting with cardiopulmonary bypass) with (n = 26) or without (n = 36) post-operative complications. Clinical and laboratory data on the first day after surgery were analyzed. Additionally, patients' blood samples were collected before and on the first day after surgery to determine biomarkers and metabolites. Results: Multivariate PLS-DA models, predicting the presence or absence of post-operative complications, were built using clinical data, concentrations of metabolites and biomarkers, and the entire data set (ROC-AUC = 0.80, 0.71, and 0.85, respectively). For comparison, we built univariate models using the EuroScore2 and SOFA scales, concentrations of lactate, the dynamic changes of 4-hydroxyphenyllactic acid, and the sum of three sepsis-associated metabolites (ROC-AUC = 0.54, 0.79, 0.62, 0.58, and 0.70, respectively). Conclusions: The proposed complex model using the entire dataset had the best characteristics, which confirms the expediency of searching for new predictive models based on a variety of factors.

Detection of kernels in maize forage using hyperspectral imaging

In this study, the use of hyperspectral imaging was examined to enhance the kernel processing assessment of maize forage, which is crucial for optimizing the production of dairy cow feed and biogas production. As the visual contrast between the kernels and other crop particles is quite limited, spectral imaging could provide better kernel classification. A pushbroom hyperspectral imaging system was used, scanning samples collected during maize harvesting in the 400–1000 nm range. A PLSDA model for pixel classification was developed to distinguish kernel spectra from the other particles in the forage, achieving a pixel-level classification accuracy of 95.2 %. Next, the most important wavelengths were identified by means of a stepwise procedure using the Wilks Lambda criterion. A Pixel classification using the top five discriminating wavelengths achieved nearly the same accuracy as the full spectrum wavelength model (95.2 % compared to 93.5 %) and did considerably better than the RGB classifier’s 86.3 % accuracy. Finally, these top 5 discriminating wavebands were applied in an object detection deep learning model, more specifically a Faster R-CNN model. While object detection based on these 5 wavebands stilled outperformed object detection based on RGB wavebands, detection performance, as measured by AP50, was rather low. This weak performance resulted from the low resolution of the hyperspectral imaging camera.

Comparison and identification of aroma components in 21 kinds of frankincense with variety and region based on the odor intensity characteristic spectrum constructed by HS–SPME–GC–MS combined with E-nose

Frankincense is an important seasoning and spice known for its distinctive and intricate flavor profile. Considering the considerable variation in the aromatic quality of frankincense due to geographical origin, species diversity and cultivation conditions, frankincense from major global origins was characterized holistically for the first time. The electronic nose (E-nose) with headspace solid-phase microextraction gas chromatography–mass spectrometry (HS–SPME–GC–MS) and sensory evaluation were implemented to characterize the aroma components of 21 commercial varieties of frankincense from around the world. The results showed that a total of 149 volatile organic compounds (VOCs) of 10 categories were identified in frankincense, among which the numbers of alcohols, terpenes and esters compounds accounted for 22.15 %, 18.79 % and 15.44 % of the total VOCs of frankincense, respectively. The PLS-DA model effectively distinguished frankincense from Oman/Somalia and other origins. Furthermore, the study identified two differential VOCs with VIP > 1 in three Asian countries and five in six African countries. The total VOCs content and sensory characteristic score of “Lemon/Citrus” in Oman frankincense is significantly higher than other regions. The OAV results showed that 61 substances (e.g., Diacety, alpha-Pinene, Camphene, Myrcene) as key aroma compounds and OICS model indicated that p-Cymenol was found to contribute significantly to the citrus aroma in frankincense. This study identified the fundamental components of frankincense flavor and revealed different flavor descriptors of frankincense, which are crucial for reconstructing frankincense flavor and improving flavor quality.

Physicochemical, volatiles, non-volatiles, and microbial dynamics in dried versus traditional fresh red pepper paste fermentation: A comparative study

The utilization of dried red peppers (DRPs) in fermented red pepper paste creation serves as a solution to the season-dependent availability of fresh red peppers (FRPs). Nevertheless, the current understanding of the fermentation disparities between DRPs and FRPs is limited, posing challenges to DRPs application in red pepper paste manufacturing. This study conducted a side-by-side analysis of physicochemical indicators, volatiles and non-volatiles, and microbial composition within DRPs and FRPs during fermentation, deploying a comprehensive multi-omics technique. The study results signified that after two weeks of fermentation, pH levels were significantly heightened (p < 0.05) in FRPs than in DRPs, with total acid presenting an opposite trend. Concurrently, total flavonoid content and the a* value were considerably elevated in FRPs (p < 0.05) during the entire fermentation process. Furthermore, the drying process led to substantial changes (p < 0.05) in red pepper metabolites, with higher total volatiles and non-volatiles found in FRPs throughout fermentation. The PLS-DA modeling indicated that 20 volatiles and 37 non-volatiles were screened as differential metabolites between the two pastes, respectively. The microbial composition analysis showed a similar fungal population in both DRPs and FRPs during fermentation, but drying significantly altered the bacterial composition, resulting in Weissella and Ligilactobacillus becoming dominant post-fermentation, whereas Pantoea and Ligilactobacillus dominated in DRPs paste. This study enriches our understanding of the fermentation process using various raw materials and paves the way for the development of DRPs-based red pepper paste that can be produced year-round.

Real-Time Monitoring of Laser-Layered Paint Removal from CFRP Based on the Synergy of Laser-Induced Breakdown Spectroscopy and PLS-DA Models

Real-Time Monitoring of Laser-Layered Paint Removal from CFRP Based on the Synergy of Laser-Induced Breakdown Spectroscopy and PLS-DA Models

Forensic Discrimination of White Architectural Paints via ATR FT-IR Spectroscopy and Chemometrics

The present work entailed the comparative examination and discrimination of white architectural paint samples of different brands/manufacturers. The analysis of this evidence can facilitate exhibiting a linkage between the victim, and the crime scene and assist in the reconstruction of the crime scene. Architectural paint traces have evidential value, especially in cases of burglary, acts of vandalism, property crimes, etc. In the present study, a total of 102 samples of 34 distinct brands (3 samples of each brand) of white commercially available architectural paints were purchased and analysed using ATR FT-IR spectroscopy and multivariate analytical tools. The use of visual analysis of FTIR spectral data resulted in 12 undifferentiated/ indistinguishable sample pairs with a discrimination power of 97.86%. Adding PCA and PLS-DA in concert with visual analysis of FTIR spectra resulted in 99.4% (3 sample pairs undifferentiated) and 100% discrimination power, respectively. In conclusion, samples of each brand were completely differentiated using ATR FT-IR spectra subjected to the PLS-DA model. To assess the impact of different substrates on the examination of paint chips, simulated samples were produced on different substrates and the obtained results from the substrates study have been summarized. The ATR-FTIR spectroscopy therefore suggests an objective and rapid approach for the discrimination of white architectural paints for forensic purposes.

Characterization of the flavor profile and dynamic changes in Chinese traditional fish sauce (Yu-lu) based on electronic nose, SPME-GC-MS and HS-GC-IMS

In this study, flavor characteristics and dynamic change of Chinese traditional fermented fish sauce (Yu-lu) with different fermentation time (2, 4, 6, 8, and 12 months) were analyzed. The electronic nose analyses confirmed a notable flavor change in fish sauce samples from different stages. During the 12-months fermentation, the total volatile compounds in fish sauce increased from 3.9 mg/L to 13.53 mg/L. Acids, aldehydes, esters and phenols were the main aroma substances and their contents gradually increased during the fermentation process. The PCA of GC–MS and GC-IMS showed that fish sauce samples from different fermentation periods can be well distinguished. A total of 110 volatile compounds identified by GC–MS, and 102 volatile compounds were detected by GC-IMS. Among them, 13 compounds were identified by both GC–MS and GC-IMS. The most varieties (49) of volatiles appeared after 8 months of fermentation. The odor activity value (OAV) analysis showed that 10 volatile compounds were considered as characteristic flavor in traditional fish sauce. The variable influence on projections (VIPs) in PLS-DA models constructed by GC–MS and GC-IMS identified 5 and 10 volatile compounds as biomarkers, respectively. Our results revealed the dynamic changes of characteristic flavor in fish sauce in combination of GC–MS and GC-IMS, which provides theoretical basis for the production and flavor regulation of fish sauce.