ARTICLE   Open Access    

Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains

More Information
  • To swiftly and noninvasively assess the freshness of bighead carp heads within simulated cold chain environments, an excitation-emission matrix fluorescence spectroscopy coupled with a long short-term memory network (EEM-LSTM) model was developed. Through the parallel factor algorithm based on analysis of residuals, diagnosis of core consistency, and split-half evaluation, three key fluorescent components from fish fillets were extracted, with the most significant components linked to tryptophan and NADH, both indicative of fish freshness. The EEM-LSTM model exhibited coherent trends in freshness indicators and demonstrated exceptional predictive capabilities for four freshness indicators simultaneously, achieving R2 values exceeding 0.8840 in simulated cold chain situations. Relative errors in the supermarket direct sales cold chain were less than 10%, surpassing those of the long-distance transport cold chain. Hence, the EEM-LSTM model stands validated for predicting fish freshness in simulated cold chains, holding promise for real-world aquatic product freshness forecasting within cold chain scenarios.
  • The Trichoderma genus encompasses a wide-ranging collection of filamentous fungi, prevalent in various natural ecosystems[1]. Trichoderma species within this genus have earned acclaim for their exceptional capacity to inhabit plant roots, stimulate plant growth, and showcasing biocontrol attributes against a spectrum of fungal adversaries[2]. Employing tactics like mycoparasitism, antibiosis, competitive resource acquisition, and plant resistance induction, these species effectively manage fungal diseases[3]. Notably, they are increasingly utilized in agriculture as biofertilizers and biopesticides[1]. Trichoderma-based bio-fungicides, available in different formulations like wettable powders, granules, and flowable concentrates, offer a convenient application to seeds, seedlings, soil, and foliage[4,5]. Besides their disease-fighting properties, these bio-fungicides promote plant growth through various mechanisms such as phytohormone production, nutrient solubilization, and stress tolerance enhancement[3]. Recent progress in Trichoderma-based formulations has led to innovative materials, advanced nanotechnology strategies, and genetic engineering techniques aimed at boosting stability, shelf life, and efficacy[4]. Among these advancements, biochar has shown promise as an ideal carrier for Trichoderma formulations due to its high porosity, surface area, and soil stability maintenance abilities[6]. New research indicates that biochar can strengthen Trichoderma's biocontrol properties[7,8]. Experiments show that using Trichoderma bio-fungicides on soil blended with biochar is more effective in fungal disease control than on unamended earth[9]. Likewise, applying these bio-fungicides on biochar-coated seeds provides better resistance against fungal diseases in seedlings[7,10]. Such biotechnological advancements in Trichoderma-based formulations can promote sustainable agricultural practices by reducing reliance on chemical pesticides[11]. This, in turn, helps mitigate the ecological impact of agricultural activities and enhances food and feed safety[12]. By protecting plants from fungal diseases and improving soil fertility, Trichoderma-based bio-fungicides hold promise for enhancing crop yield[1]. Trichoderma formulations play a crucial role in minimizing harm to non-target organisms while maximizing the effectiveness of the active ingredient[13]. While Trichoderma is significant in ensuring agronomic safety, challenges in their formulation persist due to potential degradation of the biomass or bioactive metabolite caused by factors like exposure to air, light, and temperature[14]. Additionally, these products need to be easy to handle, apply, and produce[15,16]. To address this objective, the present study aims to offer a comprehensive examination of various technological advancements that enhance the efficiency of natural preparations. Distinguishing itself from typical literature reviews that predominantly delve into the biological attributes of metabolites, this review incorporates a bibliometric analysis of biopesticides and their formulations[17]. This analysis employs quantitative and statistical indicators to identify patterns related to the most critical pest issues, agriculture's susceptibility, sources of biological control, innovative methodologies, and the current status of Trichoderma formulations. The insights presented in this analysis significantly contribute to the bibliometric methodology, potentially promoting positive strides in the advancement of technology for Trichoderma formulation. Additionally, it offers valuable suggestions for researchers engaged in this field.

    Bibliometric analysis is a technique employed to scrutinize the characteristics and evolving patterns within academic literature using various mathematical and statistical methods[10]. Through this approach, we can quantitatively assess the overall state of the literature, collaborative relationships, research areas of interest, and the development trends in a specific research field[18]. A descriptive analysis of the corpus of published research pertaining to Trichoderma formulations was conducted. This analysis entailed the examination of co-occurring terms within the body of published articles, allowing for the elucidation of evolutionary trends in scientific themes[19]. The fundamental aim of this research is to conduct an exhaustive review of the existing body of literature on Trichoderma formulations and to project the areas of highest interest and potential for future investigation[20]. One of the primary objectives of bibliometric analysis is to assess the trends in research related to Trichoderma formulations, and to identify the most influential authors and institutions in the field of Trichoderma research. Determining the impact of research in terms of citations, patents, or applications in real-world scenarios.

    As a result, this study seeks to investigate the following research objectives: In the field of Trichoderma formulations, what are the key research themes and trends observed from 2016 to 2023 include:

    (1) How is research on Trichoderma formulations distributed geographically, and what regions exhibit the most active contributions to the field? (2) Can bibliometric analysis predict future trends and potential innovations in Trichoderma formulations research based on historical patterns? (3) The article follows a well organised structure[21]. Initially the research methodology adopted for the study is outlined. Subsequently, a well-organized article is crucial for effectively communicating research findings to the intended audience[22].

    Literature retrieval was performed online through the Science Citation Index Expanded (SCI-E) of the Web of Science Core Collection (WoSCC, Clarivate Analytics) from 2016 to 2023[21,22]. Scopus is a preferred data source for bibliometric analysis, and it provides comprehensive information and data from a multi-disciplinary field of literature[23]. To retrieve literature comprehensively and accurately on Trichoderma formulations, different search terms and retrieval strategies were assembled in this study. Finally, the optimal search items were set as follows: TS = ('Trichoderma formulation*') OR ('Bio formulation of Trichoderma*') OR ('Bio control') OR ('Antagonist') OR ('Rhizosphere fungus')[24]. The data range was set from 2016 to 2023, to collect all relevant publications. It is worth noting that as the Scopus database data network is constantly updated, the results may vary depending on the exact retrieval date.

    A detailed literature retrieval process was performed online through Science Citation Index Expanded (SCI-E) of the Web of Science Core Collection (WoSCC, Clarivate Analytics) from 2016 to 2023, and considered Scopus as a preferred data source for bibliometric analysis. Additionally, we outlined the search terms and retrieval strategies that are used in our study and set the range of data from 2016 to 2023 to collect all relevant publications. If we sum up our descriptions narrate on the following key points, like data sources, search terms, data range, constantly updated data base, and optimal search items[25]. The search strategy was designed to capture relevant literature on Trichoderma formulations. The search terms included Trichoderma formulation, bio-formulation of Trichoderma, biocontrol, antagonist and Rhizosphere fungus. The asterisks, in the search terms are used as wildcard characters to capture different word endings. The data range was set from 2016 to 2023 to collect all relevant publications within that timeframe. This was the period during which the literature retrieval was performed[25]. It is mentioned that as the Scopus database data network is constantly updated, and the results may vary depending on the exact retrieval date. This indicates that the study acknowledged the dynamic nature of the database and its potential impact on the results[26]. After assembling different search terms and retrieval strategies, the study determined the optimal search items, which were the selected search terms that would yield the most comprehensive and accurate results for the study's objectives. Overall, the present study took a systematic approach to literature retrieval, considering multiple data sources and employing a combination of search terms to ensure the retrieval of relevant publications on Trichoderma formulations. It is worth noting that as the Scopus database data network is constantly updated, to add upon, the results may vary depending on the exact retrieval date.

    To ensure the credibility of the research conclusions, this study gathered peer-reviewed English journal articles to summarize global research perspectives. It's important to mention that articles not aligned with this paper's purpose were manually omitted in the final phase of data collection[27]. For instance, some articles explored the relationship between plants and microorganisms on leaves. Eventually, a total of 287 articles that met all criteria were sourced from Scopus. These pieces represented almost all top-tier experimental studies on Trichoderma formulations from 2016 to 2023 worldwide. The variability among these articles could effectively indicate the trend of related research development. Therefore, these publications were prioritized for further analysis and assessment. Figure 1 illustrates the flowchart of the literature retrieved in this study[28].

    Figure 1.  Flow chart of literature review methodology.

    Table 1 presents the top 10 countries/regions, institutions, authors, and journals that published the most studies on Trichoderma formulations. As indicated in Table 1, China emerged as the country making the most significant contribution, with 1,231 publications, accounting for 20.33% of the total. India and Pakistan followed closely, ranking second and third, with 1,096 (18.10%) and 618 (10.20%) publications, respectively. Among the institutions, the Chinese Academy of Sciences held the top spot, boasting 160 (2.64%) publications. Following closely was the University of Agriculture, Faisalabad (144, 2.37%), and Nanjing Agricultural University (137, 2.26%). In the realm of scholarly contributions, prolific authors often set the tone for research trends. Identifying these influential scholars can shed light on the direction of the research field[29]. The leading author in the study of rhizosphere microorganisms was Wang Y, with 102 publications. Li Y, Zhang Y, and Hkan M were also highly prolific, each publishing nearly 90 studies. These works were predominantly featured in prominent journals in Ecology and Botany, such as Frontiers in Microbiology (3.73%), Frontiers in Plant Science (2.36%), and Plant and Soil (2.03%).

    Table 1.  Leading journals contributing to the existing body of knowledge in the field of formulation of Trichoderma.
    Sl. no. Name of journal No. of publication Citations
    1 Journal of Applied Microbiology 2 40
    2 Applied Microbiology and Biotechnology 2 16
    3 Indian Phytopathology 3 2
    4 Biological Control 2 59
    5 Crop Protection 2 30
    6 Frontiers in Microbiology 2 23
    7 Journal of Biological Control 2 1
    8 Medicinal Plants 2 5
     | Show Table
    DownLoad: CSV

    The analysis of scientific production on Trichoderma formulations demonstrated the trend of publications per year on Trichoderma formulation studies. It was observed that published research showed a significant increase of 71.24% over the last decade (2016–2023) (Fig. 1). The increasing trend is possibly related to economic support from government programs, since funding for innovative, sustainable, and ecological research is being considered to meet the demand for food and mitigate environmental pollution[30]. Figure 2 illustrates a co-citation map of authors collaborating in the field of Trichoderma formulation. The purpose of conducting this co-citation analysis is to visually portray the knowledge base of the specific area of review. The analysis identifies three distinct clusters, each represented by different colored nodes: blue, red, and green. The green cluster stands out as it is associated with Harman, who has the highest collaboration, working with nine researchers on Trichoderma formulation research. The red cluster, on the other hand, signifies the second-highest collaboration, led by Mukherjee et al.[31] with six researchers. Lastly, the pink cluster represents the lowest level of collaboration among researchers, with only two researchers working together in this area.

    Figure 2.  Co-citation analysis of cited authors as the unit of analysis in the field of Trichoderma formulation.

    Table 2 showcases the country–wise citation analysis, and the series presented here seems to have large variability in distribution. In terms of total citations of studies dedicated to Trichoderma formulations. Citations count of articles by country as a unit of analysis represents the popularity of a field of research in a particular region. India with 20 publications having 115 citations topped the list and, therefore, is the most impactful country contributing to the existing body of knowledge in the said domain followed by Italy with four publications having 69 citations and Brazil with three publications having 51 citations. From the viewpoint of the total number of publications, India holds first position, having 20 publications, followed by Italy having four publications.

    Table 2.  Leading countries contributing to the existing body of knowledge in the field of formulation of Trichoderma.
    Sl. no. Country No. of publication Citations
    1 Argentina 1 20
    2 Belgium 2 30
    3 Brazil 3 51
    4 China 2 82
    5 Croatia 1 30
    6 Finland 1 37
    7 India 20 115
    8 Italy 4 69
    9 New Zealand 2 21
    10 Portugal 1 67
    11 South korea 1 54
     | Show Table
    DownLoad: CSV

    The top researchers working in the field of Trichoderma formulation are presented in Table 2. The authors' citation count represents the recognition of their research work in a particular field of research. It is quite clear from the list of 13 author's citations that all the authors have at least 10 citations to their name based on the total citation count. Among the 13 authors Park et al.[32] has the highest number of citations of 57 followed by Herrera-Téllez et al.[33] having 47 and Hewedy et al.[34] having 44 citations. The analysis of bibliographic coupling in the Trichoderma formulation domain is represented in Fig. 3. This technique utilizes references from existing publications to elucidate the relevant literature[35]. For this study, five thematic clusters have been identified, labeled green, red, blue, yellow, and purple. Among these interconnected groups, India stands out as the country with the most extensive collaboration network, linked with 15 other countries. Given that India also holds the highest number of published documents (115), it was anticipated to be the central node in this cooperation network. The findings demonstrate the strong relationships between researchers and their respective institutional affiliations, emphasizing the scientific cooperation aimed at developing sustainable and ecologically sound strategies for crop protection in a competitive manner[36].

    Figure 3.  Bibliographic coupling of articles in the field of Trichoderma formulation.

    In Fig. 4, a co-occurrence analysis of keywords with a minimum threshold of five occurrences is displayed. The network illustrates the most frequently utilized terms within the 'Trichoderma formulation' research domain, capturing the essence of the article's core content. The prevalence of these keywords can be indicative of the research direction and content within this specific field[37]. This analysis allows for the identification of developmental trends within a field and a comprehensive understanding of its current research status[38]. The co-occurrence graph of keywords reveals a total of four co-occurrence clusters (Fig. 4), encompassing themes like biocontrol, formulation, Trichoderma, and shelf life. Each cluster is further examined below, providing an in-depth portrayal of the prominent topics within the Trichoderma formulations landscape during the research period. Annual publication number leading years contributing to the existing body of knowledge in the field of formulation of Trichoderma is presented in Table 3.

    Figure 4.  Co-occurrence analysis based upon keywords from articles in the field of Trichoderma formulation.
    Table 3.  Annual publication number leading years contributing to the existing body of knowledge in the field of formulation of Trichoderma.
    Sl. no. Year No. of publication
    1 2016 8
    2 2017 13
    3 2018 21
    4 2019 16
    5 2020 20
    6 2021 17
    7 2022 15
    8 2023 8
     | Show Table
    DownLoad: CSV

    The shift in annual publication counts serves as a vital benchmark for gauging the progress of a research field, lending insights into potential development trends[28]. Figure 4 provides a clear portrayal of the publication distribution in Trichoderma formulation from 2016 to 2023, illustrating a noticeable increase in annual article output. This surge suggests a heightened interest in the field over the past few years. Scientific research fields typically undergo a four-stage evolution[39] ; (1) the inception phase, characterized by the introduction of novel research areas or directions by notable scientists; (2) the expansion phase, where scientists gravitate toward the emerging research direction, leading to a proliferation of discussion topics; (3) the stabilization phase, marked by the amalgamation of new knowledge to form a distinct research context; and (4) the contraction phase, in which the number of new publications diminishes. Notably, the research on Trichoderma formulation seems to be currently in the expansion phase[40].

    The publication pattern reveals a growing research interest in Trichoderma formulations, a relatively new field that is attracting increasing enthusiasm among scholars. Notably, the top contributors to this area, as identified through VOS viewer analysis, include key individuals, organizations, sources, and countries. Park emerges as the leading author based on citation count, followed by Herrera-Téllez and Hewedy. China, Portugal, Italy, and South Korea are recognized as major contributors to research in this field, as reflected in their citation counts. Conversely, India leads in terms of document count, demonstrating a significant contribution to the literature.

    Co-citation and bibliographic coupling analyses have identified three distinct thematic clusters. In the co-citation analysis, these clusters relate to application methods, types of Trichoderma formulations, and their biocontrol efficacy. Additionally, insights from the bibliometric analysis of biopesticide formulations have facilitated the integration of methods and strategies aimed at enhancing the effectiveness of Trichoderma formulations.

    This paper conducts a bibliometric analysis to critically examine articles related to biological control, focusing specifically on those published in various indexed journals, and offers a comprehensive overview of the evolution of Trichoderma formulations over time. The primary objective of this research is to investigate and characterize the key literature on this topic, covering historical, current, and emerging developments in this dynamic field. Bibliometric techniques are employed to visualize the Trichoderma formulation landscape.

    To achieve this goal, the study analyses a dataset of articles obtained from the core collection databases of Scopus and Web of Science. Within the scope of this study, significant publications on Trichoderma formulations are meticulously reviewed to highlight the potential trajectory of Trichoderma's role in biological disease control in plants. The study identifies and examines the various developmental phases of Trichoderma formulations, offering a comprehensive analysis that can shape the future of this critical research area. Given the scarcity of comprehensive bibliometric studies on Trichoderma formulation research, this study seeks to fill this gap, making a significant contribution to the extensive readership interested in Trichoderma.

    This study has several limitations and challenges that must be considered by future researchers. First, the study relies on a single database, which could restrict the amount of available data. Additionally, the search criteria were limited to research articles, and only those with specific phrases in the title were included, which may not represent the complete dataset. However, Trichoderma's biological control mechanisms against plant fungal and nematode diseases involve various strategies, including competition, antibiosis, antagonism, and mycoparasitism. In addition to these, Trichoderma enhances plant growth and induces systemic resistance in plants, making it effective in controlling a wide range of plant fungal and nematode diseases[41]. Although biological control is effective, it generally requires time to become established in the environment, making it a slower process. Therefore, optimizing the formulation of Trichoderma-based products is essential to ensure their stability and efficacy. It is important to ensure that these formulations are compatible with other agricultural treatments, such as chemical fertilizers and pesticides, to maximize their overall effectiveness. Proper formulation can improve the shelf life, ease of application, and survival of Trichoderma under varying environmental conditions. Ongoing research is necessary to refine these formulations for broader application in integrated pest management programs[42]. This is a significant limitation as the analysis was restricted to articles published in journals, excluding other valuable sources like reviews, conferences, books, and book chapters. To overcome this limitation, future researchers should consider utilizing additional databases such as Scopus and Science Direct, which can provide more comprehensive data. This limitation may have impacted the study's ability to provide a comprehensive overview of the field.

    The authors confirm contribution to the paper as follows: conceptualization: Kumar V, Mishra KK, Panda SR; writing − original draft preparation: Kumar V, Wagh AK, Mishra KK; writing − review and editing: Panda SR, Kumar V, Wagh AK; supervision: Kumar V, Panda SR, All authors have read and agreed to the published version of the manuscript.

    The data that support the findings of this study are available on request from the corresponding authors.

  • The authors declare that they have no conflict of interest.

  • [1]

    Wang H, Zheng Y, Shi W, Wang X. 2022. Comparison of Arrhenius model and artificial neuronal network for predicting quality changes of frozen tilapia (Oreochromis niloticus). Food Chemistry 372:131268

    doi: 10.1016/j.foodchem.2021.131268

    CrossRef   Google Scholar

    [2]

    Shi X, Zhang J, Shi C, Tan Y, Hong H, et al. 2022. Nondestructive prediction of freshness for bighead carp (Hypophthalmichthys nobilis) head by Excitation-Emission Matrix (EEM) analysis based on fish eye fluid: comparison of BPNNs and RBFNNs. Food Chemistry 382:132341

    doi: 10.1016/j.foodchem.2022.132341

    CrossRef   Google Scholar

    [3]

    Zhao H, Liu S, Tian C, Yan G, Wang D. 2018. An overview of current status of cold chain in China. International Journal of Refrigeration 88:483−95

    doi: 10.1016/j.ijrefrig.2018.02.024

    CrossRef   Google Scholar

    [4]

    Yu J, Xiao K, Xue W, Shen Y, Tan J, et al. 2020. Excitation-emission matrix (EEM) fluorescence spectroscopy for characterization of organic matter in membrane bioreactors: principles, methods and applications. Frontiers of Environmental Science & Engineering 14:31

    doi: 10.1007/s11783-019-1210-8

    CrossRef   Google Scholar

    [5]

    Li L, Wang Y, Zhang W, Yu S, Wang X, et al. 2020. New advances in fluorescence excitation-emission matrix spectroscopy for the characterization of dissolved organic matter in drinking water treatment: A review. Chemical Engineering Journal 381:122676

    doi: 10.1016/j.cej.2019.122676

    CrossRef   Google Scholar

    [6]

    ElMasry G, Nagai H, Moria K, Nakazawa N, Tsuta M, et al. 2015. Freshness estimation of intact frozen fish using fluorescence spectroscopy and chemometrics of excitation-emission matrix. Talanta 143:145−56

    doi: 10.1016/j.talanta.2015.05.031

    CrossRef   Google Scholar

    [7]

    Rahman MM, Shibata M, ElMasry G, Nakazawa N, Nakauchi S, et al. 2019. Expeditious prediction of post-mortem changes in frozen fish meat using three-dimensional fluorescence fingerprints. Bioscience Biotechnology and Biochemistry 83:901−13

    doi: 10.1080/09168451.2019.1569494

    CrossRef   Google Scholar

    [8]

    Liao Q, Suzuki T, Kohno Y, Al Riza DF, Kuramoto M, et al. 2018. Potential of using uric acid fluorescence in eye fluid for freshness assessment on Red Sea bream (Pagrus major). Spectroscopy Letters 51:431−37

    doi: 10.1080/00387010.2018.1501395

    CrossRef   Google Scholar

    [9]

    Marini F. 2009. Artificial neural networks in foodstuff analyses: Trends and perspectives A review. Analytica Chimica Acta 635:121−31

    doi: 10.1016/j.aca.2009.01.009

    CrossRef   Google Scholar

    [10]

    Chu Y, Tan M, Yi Z, Ding Z, Yang D, et al. 2021. Shelf-life prediction of glazed large yellow croaker (Pseudosciaena crocea) during frozen storage based on Arrhenius model and long-short-term memory neural networks model. Fishes 6(3):39

    doi: 10.3390/fishes6030039

    CrossRef   Google Scholar

    [11]

    Guo L, Xu C, Yu T, Wumaier T, Han X. 2024. Ultra-short-term wind power forecasting based on long short-term memory network with modified honey badger algorithm. Energy Reports 12:3548−65

    doi: 10.1016/j.egyr.2024.09.021

    CrossRef   Google Scholar

    [12]

    Ren H, Du L, Peng C, Yang J, Gao W. 2024. The composite drought index incorporated solar-induced chlorophyll fluorescence enhances the monitoring capability of short-term drought. Journal of Hydrology 637:131361

    doi: 10.1016/j.jhydrol.2024.131361

    CrossRef   Google Scholar

    [13]

    Matenda RT, Rip D, Marais J, Williams PJ. 2024. Exploring the potential of hyperspectral imaging for microbial assessment of meat: A review. Spectrochimica Acta Part a-Molecular and Biomolecular Spectroscopy 315:124261

    doi: 10.1016/j.saa.2024.124261

    CrossRef   Google Scholar

    [14]

    Huang W, Zhao J, Wang X. 2024. Model-driven multimodal LSTM-CNN for unbiased structural forecasting of European Union allowances open-high-low-close price. Energy Economics 132:107459

    doi: 10.1016/j.eneco.2024.107459

    CrossRef   Google Scholar

    [15]

    Gao K, Huang Z, Lyu C, Liu C. 2024. Multi-scale prediction of remaining useful life of lithium-ion batteries based on variational mode decomposition and integrated machine learning. Journal of Energy Storage 99:113372

    doi: 10.1016/j.est.2024.113372

    CrossRef   Google Scholar

    [16]

    Zhang Y, Chu G, Shen D. 2021. The role of investor attention in predicting stock prices: The long short-term memory networks perspective. Finance Research Letters 38:101484

    doi: 10.1016/j.frl.2020.101484

    CrossRef   Google Scholar

    [17]

    Boutheina B, Leila K, Besbes N, Messina CM, Santulli A, et al. 2023. Evaluation of the qualitative properties and consumer perception of marinated sardine Sardina pilchardus: The effect of fucoxanthin addition. International Journal of Gastronomy and Food Science 31:100611

    doi: 10.1016/j.ijgfs.2022.100611

    CrossRef   Google Scholar

    [18]

    Koel M. 2024. Developments in analytical chemistry initiated from green chemistry. Sustainable Chemistry for the Environment 5:100078

    doi: 10.1016/j.scenv.2024.100078

    CrossRef   Google Scholar

    [19]

    Takefuji Y. 2023. An updated tutorial on reproducible PyPI applications for advancing chemometrics and boosting learner motivation. Chemometrics and Intelligent Laboratory Systems 241:104941

    doi: 10.1016/j.chemolab.2023.104941

    CrossRef   Google Scholar

    [20]

    Smilde AK, Doornbos DA. 1992. Simple validatory tools for judging the predictive performance of parafac and three-way PLS. Journal of Chemometrics 6:11−28

    doi: 10.1002/cem.1180060103

    CrossRef   Google Scholar

    [21]

    Câmara ABF, da Silva WJO, Neves ACdO, Moura HOMA, de Lima KMG, de Carvalho LS. 2024. Excitation-emission fluorescence spectroscopy coupled with PARAFAC and MCR-ALS with area correlation for investigation of jet fuel contamination. Talanta 266:125126

    doi: 10.1016/j.talanta.2023.125126

    CrossRef   Google Scholar

    [22]

    Mangalgiri K, Cheng Z, Cervantes S, Spencer S, Liu H. 2021. UV-based advanced oxidation of dissolved organic matter in reverse osmosis concentrate from a potable water reuse facility: a parallel-factor (PARAFAC) analysis approach. Water Research 204:117585

    doi: 10.1016/j.watres.2021.117585

    CrossRef   Google Scholar

    [23]

    Yu Y, Si X, Hu C, Zhang J. 2019. A review of recurrent neural networks: LSTM cells and network architectures. Neural Computation 31:1235−70

    doi: 10.1162/neco_a_01199

    CrossRef   Google Scholar

    [24]

    Smagulova K, James AP. 2019. A survey on LSTM memristive neural network architectures and applications. European Physical Journal-Special Topics 228:2313−24

    doi: 10.1140/epjst/e2019-900046-x

    CrossRef   Google Scholar

    [25]

    Tran TTK, Bateni SM, Ki SJ, Vosoughifar H. 2021. A review of neural networks for air temperature forecasting. Water 13:1294

    doi: 10.3390/w13091294

    CrossRef   Google Scholar

    [26]

    Bekhit AEDA, Holman BWB, Giteru SG, Hopkins DL. 2021. Total volatile basic nitrogen (TVB-N) and its role in meat spoilage: A review. Trends in Food Science & Technology 109:280−302

    doi: 10.1016/j.jpgs.2021.01.006

    CrossRef   Google Scholar

    [27]

    Lorentzen G, Ageeva TN, Heide M, Esaiassen M. 2020. Temperature fluctuations in processing and distribution: Effect on the shelf life of fresh cod fillets (Gadus morhua L.). Food Control 112:107102

    doi: 10.1016/j.foodcont.2020.107102

    CrossRef   Google Scholar

    [28]

    Gayer AV, Yakimov BP, Sluchanko NN, Shirshin EA. 2023. Multifarious analytical capabilities of the UV/Vis protein fluorescence in blood plasma. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 286:122028

    doi: 10.1016/j.saa.2022.122028

    CrossRef   Google Scholar

    [29]

    Wang X, Xie J, Qian Y. 2020. A non-invasive method for quantitative monitoring of quality changes and water migration in bigeye tuna (Thunnus obesus) during simulated cold chain logistics using low-field nuclear magnetic resonance coupled with PCA. Food Science and Technology International 26:475−84

    doi: 10.1177/1082013220903148

    CrossRef   Google Scholar

    [30]

    Wu Y, Deng J, Xu F, Li X, Kong L, et al. 2023. The mechanism of Leuconostoc mesenteroides subsp. IMAU:80679 in improving meat color: Myoglobin oxidation inhibition and myoglobin derivatives formation based on multi enzyme-like activities. Food Chemistry 428:136751

    doi: 10.1016/j.foodchem.2023.136751

    CrossRef   Google Scholar

    [31]

    Karoui R, Thomas E, Dufour E. 2006. Utilisation of a rapid technique based on front-face fluorescence spectroscopy for differentiating between fresh and frozen–thawed fish fillets. Food Research International 39:349−55

    doi: 10.1016/j.foodres.2005.08.007

    CrossRef   Google Scholar

    [32]

    Wu B, Dahlberg K, Gao X, Smith J, Bailin J. 2018. A rapid method based on fluorescence spectroscopy for meat spoilage detection. International Journal of High Speed Electronics and Systems 27:1840025

    doi: 10.1142/S0129156418400256

    CrossRef   Google Scholar

    [33]

    Sun F, Zong W, Liu R, Chai J, Liu Y. 2010. Micro-environmental influences on the fluorescence of tryptophan. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy 76:142−45

    doi: 10.1016/j.saa.2010.03.002

    CrossRef   Google Scholar

    [34]

    Zhang YY, Li XS, Ren KD, Peng J, Luo XJ. 2023. Restoration of metal homeostasis: a potential strategy against neurodegenerative diseases. Ageing Research Reviews 87:101931

    doi: 10.1016/j.arr.2023.101931

    CrossRef   Google Scholar

    [35]

    Hassoun A, Karoui R. 2015. Front-face fluorescence spectroscopy coupled with chemometric tools for monitoring fish freshness stored under different refrigerated conditions. Food Control 54:240−49

    doi: 10.1016/j.foodcont.2015.01.042

    CrossRef   Google Scholar

    [36]

    Shibu A, Jones S, Tolley PL, Diaz D, Kwiatkowski CO, et al. 2023. Correlating structure and photophysical properties in thiazolo[5, 4-d]thiazole crystal derivatives for use in solid-state photonic and fluorescence-based optical devices. Materials Advances 4:6321−32

    doi: 10.1039/D3MA00686G

    CrossRef   Google Scholar

    [37]

    Chicco D, Warrens MJ, Jurman G. 2021. The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation. Peerj Computer Science 7:e623

    doi: 10.7717/peerj-cs.623

    CrossRef   Google Scholar

  • Cite this article

    You J, Sun Z, Li X, Ying X, Shi C, et al. 2024. Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains. Food Innovation and Advances 3(4): 405−415 doi: 10.48130/fia-0024-0037
    You J, Sun Z, Li X, Ying X, Shi C, et al. 2024. Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains. Food Innovation and Advances 3(4): 405−415 doi: 10.48130/fia-0024-0037

Figures(5)  /  Tables(1)

Article Metrics

Article views(1259) PDF downloads(253)

ARTICLE   Open Access    

Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains

Food Innovation and Advances  3 2024, 3(4): 405−415  |  Cite this article

Abstract: To swiftly and noninvasively assess the freshness of bighead carp heads within simulated cold chain environments, an excitation-emission matrix fluorescence spectroscopy coupled with a long short-term memory network (EEM-LSTM) model was developed. Through the parallel factor algorithm based on analysis of residuals, diagnosis of core consistency, and split-half evaluation, three key fluorescent components from fish fillets were extracted, with the most significant components linked to tryptophan and NADH, both indicative of fish freshness. The EEM-LSTM model exhibited coherent trends in freshness indicators and demonstrated exceptional predictive capabilities for four freshness indicators simultaneously, achieving R2 values exceeding 0.8840 in simulated cold chain situations. Relative errors in the supermarket direct sales cold chain were less than 10%, surpassing those of the long-distance transport cold chain. Hence, the EEM-LSTM model stands validated for predicting fish freshness in simulated cold chains, holding promise for real-world aquatic product freshness forecasting within cold chain scenarios.

    • Bighead carp (Aristichthys nobilis) is a primary freshwater fish in the aquaculture industry with a low price, large catch quantity, and short growth cycle[1]. According to the 2022 Fishery Statistical Yearbook of China, the total output of bighead carp reached 3.177 million tons in 2021, ranking third among freshwater aquaculture species. The head of bighead carp accounts for approximately one-third or even one-half of the body length, and the delicacies made from it, such as chopped pepper fish head and blanched silver carp, are popular among consumers. However, due to the inherent high protein and moisture content of fish, post-mortem microbial growth occurs rapidly, resulting in accelerated deterioration[1]. Shi et al. developed an EEM-BPNN model to predict the TAC, TBARS, TVB-N, and K value of bighead carp heads in the storage periodiod[2]. Recently, the Chinese government has been increasingly focused on the protection of food in the cold chain, which is beneficial for not only reducing the losses of producers but also strengthening consumer confidence in food safety[3]. Temperature fluctuations are however almost inevitable during processing, transport services, and retailing, which is not conducive to the total quality and freshness of bighead carp heads. Therefore, the fish industry must assess how different temperatures throughout the food chain influence the freshness of fish.

      Nondestructive spectroscopy including infrared spectroscopy, hyperspectral imaging spectroscopy, Raman spectroscopy, and fluorescence spectroscopy has been successfully used to determine the freshness of fish with the advantages of high specificity, convenience, rapidity, non-destruction, and non-invasion. Compared with near-infrared spectroscopy, Raman spectroscopy, and hyperspectral imaging, the intensity of the fluorophores in fluorescence spectroscopy represents the amount of fluorescent substances, so it can accurately quantify the given fluorescent substances. Excitation-emission matrix (EEM) spectroscopy, a novel fluorescence spectroscopy technique has been gradually used in the field of food because of its ability to detect trace substances and enable the dialectical differentiation of various fluorescent substances[4]. The features of EEM spectra, including the average position of the fluorescence maximum and the span of emission and excitation wavelengths of fluorescent substances have been successfully applied to evaluate the freshness of aquatic products[5]. Studies have demonstrated the potential of EEM to characterize fluorophores in aquatic products[6], including the determination of niacinamide adenine dinucleotide (NAD and NADH)[7], amino acid, and uric acid content[8], which can be used to track the changes in fish freshness after death. These results indicated that EEM has attracted attention in the field of aquatic products and has an increasing number of applications for indication, characterization, and prediction.

      An artificial neural network, such as the backpropagation neural network (BP-NN), radial basis function neural network (RBF-NN), and LSTM network are nonlinear mathematical models consisting of a considerable amount of interconnected neurons that can match the intricate connections between input layer and output layer data through its learning ability[9]. BP-NN is the most commonly used artificial neural network, but due to its fixed learning rate, it often has unfavorable self-learning and multi-output capabilities[10]. The RBF-NN is an architecture based on a feed-forward network with one hidden layer that is much better in terms of learning speed and stability than BP-NN but more prone to underfitting phenomenon[11]. The LSTM network has a unique recurrent network structure associated using a suitable gradient-based optimization algorithm and can be used to solve complex algorithms and artificial tasks with extended time delays that previous cyclic network algorithms have not addressed. Therefore, the LSTM network results in more successful running and faster learning than real-time cyclic learning, time backpropagation, Elman networks, and neural sequence chunking[12], and has been used to forecast short-term metro ridership[13], coal stock[14], the remaining useful life of lithium-ion batteries[15], and the stock market[16]. However, the application of the LSTM network in aquatic product freshness prediction has rarely been reported. Furthermore, the prediction of food freshness presents significant challenges due to the influence of intrinsic and environmental factors, whereas the LSTM network has certain advantages in exploring the correlations among these variables.

      Therefore, the current study aims to not only probe the correlation between fluorescence components and fish quality during storage of bighead carp heads but also establish the LSTM network combined with fluorescent substances obtained from EEM data through parallel factor algorithm to simultaneously and precisely predict TVB-N, TBARS, K, and TVC values in simulated cold chains, which is beneficial for further developing portable fluorescent devices.

    • Live bighead carp were obtained from the Beijing Wholesale Market of Aquatic Products (Beijing, China) and transported directly to the laboratory in oxygenated water. They were humanely slaughtered after being rendered unconscious by a blow to the head with wooden sticks. The heads were separated from the fresh bighead carp by cutting vertically with the dorsal fin as the cutting point, and the scales, gills, and internal organs were removed. After being washed under running water and drained, bighead carp heads were placed into polyethylene valve bags (35 cm × 45 cm). The entire bighead carp treatment process followed the Regulations released by the State Council of China in 1988 and the Guidelines for the Treatment of Experimental Animals published by the Ministry of Science and Technology of China in 2006.

      This project was divided into two parts: the isothermal experiment and the nonisothermal experiment for the measurement of K values, TVB-N, TVC, and TBARS along with the scanning of the excitation-emission matrix fluorescence spectrum of fish fillets, respectively. In the isothermal experiment, the heads of bighead carp were placed in low-temperature incubators set at 16 ± 0.5 °C, 12 ± 0.5 °C, 8 ± 0.5 °C, 4 ± 0.5 °C, and 0 ± 0.5 °C respectively. Regarding the analysis, the sampling procedure was as follows: eight samples were randomly chosen every 3 days for those stored at 0 ± 0.5 °C, every 2 days for those at 4 ± 0.5 °C, and every 1 day for those stored at 16 ± 0.5 °C, 12 ± 0.5 °C, and 8 ± 0.5 °C respectively. For the nonisothermal experiment, two simulated cold chains were considered namely the supermarket direct sales (SDS) cold chain and the long-distance transport (LDT) cold chain. The bighead carp heads were stored in refrigerators for the SDS cold chain using the temperature schedule outlined below: the first stage for simulating the storage process: 0 ± 0.5 °C for 3 d, the second phase for simulating the transport process: 4 ± 0.5 °C for 1 d, the third phase for simulating the transport process: 12 ± 0.5 °C for 1 d, and the fourth phase for simulating the sale process: 4 ± 0.5 °C for 1 d. The LDT cold chain was mainly simulated with alternating storage and transport processes using the temperature schedule outlined below: the first stage for simulating the storage process 0 ± 0.5 °C for 3 d, the second phase for simulating the transport process: 12 ± 0.5 °C for 1 d, the third phase for simulating the storage process: 4 ± 0.5 °C for 1 d, the fourth phase for simulating the transport process: 16 ± 0.5 °C for 1 d, and the fifth phase for simulating the sale process: 4 ± 0.5 °C for 1 d. Eight samples were picked at random for evaluation at the specified time intervals.

    • The TVB-N was determined through semi-micro steam distillation using some adjustments[2]. Three grams of minced fish was appended to 30 mL distilled water, stirred, and shaken in a constant-temperature oscillator for 30 min. After centrifugation, 5 mL supernatant and 5 mL MgO suspension (10 g·L−1) were distilled in a Kjeldahl apparatus (KDY-9820, Beijing, China). The distillate was gathered in a conical bottle with boric acid and mixed indicator (g·L−1: methyl red solution 2; methylene blue solution 1) and titrated with 0.01 M hydrochloric acid.

    • The TBARS value was measured as described by Boutheina et al.[17] with certain adjustments. The minced samples (2 g) were added to 16 mL, trichloroacetic acid solution (5% weight/volume) along with 100 μL of butylated hydroxytoluene (2 g·L−1), and homogenized for 1 min. After centrifugation, 5 mL supernatant was blended with 1 mL thiobarbituric acid solution (0.01 M) and incubated in a boiling water bath for 40 min. The mixture was brought down to room temperature, and its absorbance was measured at 532 nm with an ultraviolet spectrophotometer (UV2600A, Shanghai, China). The TBARS value was computed by multiplying the optical density measurements by multiplying by 10.2 obtained from a standard curve established with 1, 1, 3, and 3-tetraethoxy-propane.

    • The K value is characterized as the proportion of non-phosphorylated ATP metabolites to the overall products resulting from ATP breakdown[6]. The K value was measured according to the manufacturers instructions for Freshness Assessment Device (QS-3201, Sendai, Japan). A total of 0.2 g of minced fish was extracted with reagent A and then centrifuged. To adjust the pH to neutral, 200 μL of supernatant was added to reagents B and C followed by centrifugation. Three microliters of supernatant droplets were absorbed on the soaked filter paper and examined through electrophoresis at 800 v for 5 min. After drying in a drying oven and imaging under an ultraviolet lamp, The Spot Analyzer application (QS-Solution, Tokyo, Japan) was utilized for automatic image analysis and the K value was calculated.

    • The TVC value was measured following the Chinese National Standard Method GB 4789.2-2022 with some alterations. Fish tissues (5 g) were blended with 45 mL sterile sodium chloride solution (0.9%). After a tenfold serial dilution of the homogenized solution was carried out, 1 mL diluted liquid was added to a 3M™ Petrifilm™ Aerobic Count Plate (3M Company, USA) and incubated in a biochemical incubator maintained at 30 °C for 72 ± 2 h. Between 30 and 300 colonies were calculated on the tablets, expressed as log CFU/g.

    • Fluorescence fingerprints of each fish fillet, with a fillet size of approximately 4 cm × 3 cm × 1 cm (length × width × thickness), were obtained using a fluorescence spectrometer (FS5, Edinburgh, UK). Each fish fillet was placed on a bracket in the SC-10 solid sample bin for scanning. The relevant instrument parameters were adjusted as described by Shi et al.[2] with some modifications. The EEM spectra were gathered the range of excitation wavelengths between 250 and 580 nm (a span of 5 nm) and the range of emission wavelengths between 260 and 600 nm (a span of 5 nm), and with a rate of scanning of 100 nm·s−1, an excitation and emission bandwidth of 3 nm, a dwell time of 3 ms, and a xenon Lamp source without ozone. To avoid interference from the external environment, the black case housing the instrument was not opened during sample scanning. Each EEM consisted of 67 excitation (columns) and 69 emission (rows) wavelength data, with dimensions of 67 × 69. The excitation-emission matrices were displayed as contour maps, of which the vertical axis represented the excitation wavelength, the horizontal axis represented the emission wavelength, and the contour line represented the fluorescent signal strength.

    • After scattering the original EEM data, the characteristic fluorescent components of the fish could be extracted by PARAFAC analysis. The PARAFAC model was decomposed into three matrices (A, B, and C), which was an improvement upon the Tucker3 model[18], as shown in Eqn (1). xijk is the fluorescence signal of the k-th sample corresponding to the i-th excitation wavelength along with the j-th emission wavelength. Ain, Bjn, ckn are the elements related to the matrices A, B, and C, respectively. eijjk is an element of the residual matrix. The minimum factor N of the trilinear decomposition refers to the most suitable number of components in the sample.

      xijk=Nn=1ainbjnckn+eijjki=1,2,...,I;j=1,2,..,J;k=1,2,..,K (1)

      When PARAPAC was used to solve the excitation-emission matrix data, the minimum factor N of a trilinear decomposition was required. Three methods, including residual analysis, core reliability diagnostic, and half-split assessment, were used to determine N. Residual analysis is an effective method for comparing different component models and uses the sum of squared error, obtained by the alternating least squares regression method, to evaluate the most authentic three-dimensional fluorescence spectra[19,20]. The core consistency diagnostic (CORCONDIA) originated from the single-component model to assess the degree of superdiagonality[21]. Half-split assessment is employed to confirm the reliability of the PARAFAC model by creating and comparing various models after dividing data sets in two different ways[22].

    • In this study, the LSTM network algorithm was employed to create predictive models that connect (i) the fluorescent component derived from bighead carp head fillets under different storage conditions, and (ii) the measurements of TVB-N, TBARS, K value, and TVC. The construction of the freshness prediction model and the determination of related parameters were as follows.

      The LSTM network could update the current state according to the past state combined with the input data and identify the relationship between the front and back dimensions with too many input sequences[23]. The unit consisted of a cell, an input gate, an output gate, and a forget mechanism. The forget gate played an important role in the LSTM network model by allowing the deletion and forgetting of unnecessary information to avoid the influence of outdated information on the input of current data and improve the computing power of the model.[24]

      The LTSM network presents a novel internal state (ctRD) for linear loop transmission, and outputs information to the hidden layer's external state of the hidden layer (htRD) for nonlinear operation. The specific formula is as follows:

      ct=ftct1+it~ct (2)
      ht=ottanh(ct) (3)

      The three gates, ft[0,1]D, it[0,1]D, and ot[0,1]D, are the paths for controlling information transmission. is the product of the vector element. ~ctRD is obtained by the nonlinear function candidate status.

      ~ct=tanh(Wcxt+Ucht1+bc) (4)

      where, ct-1 represents the current memory unit, W*, U*, b* represents the learning network parameters ({i,f,o,c}), xt is the input at present, and ht−1 is the external state at the preceding time.

      At each instance t, ct logs the past information until the present time. The forget gate ft determines ct−1 at the previous time and thus how much data should be forgotten. The input gate it determines ct, and thus how much data should be saved. The output gate ot controls the amount of information ct should transmit to the external state ht at the current time[25]. The computation formulas for the three gates are listed below:

      it=σ(Wixt+Uiht1+bi) (5)
      ft=σ(Wfxt+Uiht1+bf) (6)
      io=σ(Woxt+Uoht1+bo) (7)

      where, σ is a logistic function, and its output interval is (0,1).

      The LSTM cycle unit can be succinctly described as follows:

      [~ctotitft]=[tanhσσσ](W[xtht1]+b) (8)
      ct=ftct1+it~ct (9)
      ht=ottanh(ct) (10)

      The root squared error (RMSE), average absolute deviation (MAE), average absolute percentage deviation (MAPE), relative error (RE), and mean relative error (MRE) are usually used to evaluate the modeling performance. The calculation formulas are as follows:

      RMSE=1nni=1(CmeaCpre)2 (11)
      MAE=ni=1|CmeaCpre|n (12)
      MAPE=1nni=1|CpreCmeaCpre| (13)
      RE(%)=|CpreCmeaCpre|×100% (14)
      MRE(%)=1nni=1|ypreymeaymea|×100% (15)
    • The excitation-emission matrix data were acquired by using the fluorescence spectrometer (FS5, Edinburgh, UK). Excel 2010 by Microsoft Corporation (Redmond, WA, USA) was employed for calculating and analyzing freshness indicators data statistics. The PARAFAC and LSTM algorithms were utilized for data handling by MATLAB 2020b (MathWorks, Natick, Massachusetts, USA), and mapping was performed by OriginLab 2018 (OriginLab, Northampton, MA, USA).

    • The TVB-N figure indicates the degradation of proteins and other N-containing substances due to microorganisms and other factors[26]. The TVB-N figure of fish samples preserved at 0, 4, 8, 12, and 16 °C rose with extended storage duration, and the increasing rate of this value of samples stored at 16 °C surpassed the levels observed at other temperatures (Fig. 1a). The TVB-N value of bighead carp heads kept at 0 °C on the 18th day, 4 °C on the 12th day, 8 °C on the 6th day, 12 °C on the 5th day, and 16 °C on the 3rd day, exceeded 20 mg/100 g, which was considered to indicate fish spoilage[2] . In the SDS and LTD cold chains, the TVB-N value rose with the prolongation of storage days. The TVB-N value improved from 10.73 to 11.67 mg/100 g in the third stage of the SDS cold chain and then increased to 12.83 mg/100 g at the end of the SDS cold chain, which did not indicate fish spoilage. The TVB-N value improved from 14.35 to 16.10 mg/100 g, higher than that stored at 4 °C on the 6th day, in the fourth stage (16 °C) of the LDT cold chain. The TVB-N value was up to 20.18 mg/100 g in the 5th phase of the LDT cold chain, which was considered to indicate fish spoilage. The accumulating rate of the TVB-N value may be affected by the range and number of temperature fluctuations[27].

      Figure 1. 

      Changes of (a) TVB-N, (b) TBARS, (c) K value, and (d) TVC of bighead carp under isothermal conditions (0, 4, 8, 12, and 16 °C) and the simulated cold chains (SDS and LDT cold chains). Note: SDS represents the supermarket direct sales cold chain; LDT represents the long distance transport cold chain.

      The TBARS value is frequently utilized as an indicator of fat oxidation levels. The initial TBARS measurement was 0.19 mg·kg−1 and the TBARS values for all samples remained below 0.60 mg·kg−1 (Fig. 1b). The TBARS values of samples stored at 0, 4, 8, 12, and 16 °C rose with extended storage time. The increasing rate of the TBARS value of samples preserved at higher temperatures was substantially less than that of samples stored at lower temperatures. The TBARS value improved from 0.28 to 0.36 mg·kg−1 in the third stage of the SDS cold chain, and the increasing rate of TBARS value in this stage was significantly greater than that in the other stages. The improving rate of the TBARS value significantly improved in the second and fifth stages of the LDT cold chain. The pace of lipid oxidation accelerated as the range of temperature fluctuations increased, which was similar to the research of Gayer et al.[28].

      The K value is used to determine the freshness index of fish regarding the decomposition products of adenosine triphosphate. The freshness of fish could be divided into three levels: a K value of < 20% represents extreme freshness, a K value of 20%−60% represents sub-class freshness, and a K value of > 60% represents initial decomposition. The initial K value was 9.46%, which was akin to the initial K value documented by Shi et al.[2]. The K value stored at 16, 12, 8, 4, and 0 °C increased and the rate of increase for the K value samples stored at 16 °C was the highest (Fig. 1c). The K value reached 42.7% in the fourth stage of the SDS cold chain. The K value rapidly increased from 35.5% to 52.3% in the fourth stage of the LDT cold chain, and the increasing rate of K value was similar to that of the samples stored at 16 °C. Then the K value reached 61.3%, which was considered to indicate fish spoilage in the fifth stage of the LDT cold chain. The decomposition ability of enzymes was accelerated with increasing temperature fluctuation, leading to the increase of K values.

      Fish is a nutrient-rich substrate, providing a good living environment for the growth and reproduction of microorganisms. The TVC value of bighead carp heads during storage exceeded 7.00 log10 CFU/g, which was considered an unacceptable level for fish[13]. The TVC value kept at 0, 4, 8, 12, and 16 °C increased during storage(Fig. 1d). The increasing rate of the TVC value decreased during the post-storage period, mainly due to the limited nutrients and living space for microorganisms. The TVC value significantly increased from 5.42 log10 CFU/g to 6.65 log10 CFU/g in the third stage of the SDS cold chain. The TVC value rapidly increased from 5.32 log10 CFU/g to 6.62 log10 CFU/g in the second stage of the LDT cold chain, at which point the increasing rate was higher than that in the second stage of the SDS cold chain. The increasing rate of the TVC value was accelerated with increasing fluctuating temperature, which agreed with the results of Wang et al.[29].

    • After the preprocessing of the original excitation-emission matrix data to remove Raman scattering and Rayleigh scattering, the optimum number of components in fish fillets was determined. The fluorescence components were then extracted through the PARAFAC method using the tool.

      In this study, PARAFAC analysis in conjunction with residual analysis, core consistency diagnostic, and half-split analysis methods were used to comprehensively estimate the optimal quantity of components. By contrasting the 2−4 component models (Fig. 2a), the sum of squared error changed obviously from the 2- to 3-component model, whereas the sum of squared error showed little change from the 3- to 4-component model, indicating that 3 components were sufficient for these data[30]. When the component was 3, the core consistency value was 99.11%, When the component was 4, the core consistency value dropped to −8.18% (Fig. 2b). A core consensus of more than 90% can be interpreted as 'very trilinear', while a core consensus value of less than 50% indicates a problem with the model. When the optimum number of components is exceeded, the core consistency value decreases sharply[19], indicating that three components were sufficient for these data. Half-split analysis was used to further verify the reliability of the 3-component model, the results of which nearly perfectly corresponded between the excitation and emission loadings of the three parts in two split-half data sets of fish fillets (Fig. 2c). Therefore, the 3-component model had the best-fitting effect.

      Figure 2. 

      PARAFAC analysis of fish fillets. The sum of squared error of (a) emission, (b) excitation, and wavelengths corresponding to the 2−4 component model in fish fillets. (c)−(f) Split half analysis for model validation of 3-component in fish fillets. (g) Core consistency diagnostic results of 2 components, (h) 3 components, and (i) 4 components in fish fillets. Note: (c)−(f) comparison of the excitation and emission loading changes of two split-half sets of fish fillets excitation emission matrices respectively. (g)−(i) the red dot represents the number of components, while the position of the green dot represents the degree of deviation from the yellow line parallel to the X-axis.

      According to the results of PARAFAC analysis, three fluorescent components of fish fillets were extracted, corresponding to Peak A (λEx/λEm: 300 nm/335 nm), Peak B (λEx/λEm: 290 nm/320 nm), and Peak C (λEx/λEm: 340 nm/460 nm), as shown in Fig. 3. Karoui et al.[31] found that the maximum fluorescent signal intensity (λEx/λEm: 290 nm/326 nm) of fresh cod fillets and the maximum fluorescent signal intensity (λEx/λEm: 290 nm/330 nm) of frozen-thawed cod fillets could be attributed to tryptophan. Peak A and Peak B could also be attributed to tryptophan according to the maximum fluorescence intensity and the excitation and emission wavelength range. The fluorescence peak movement between Peak A and Peak B was assigned to a displacement of the tryptophan of the samples to a more hydrophilic environment and the difference in fluorescence properties of the samples may also be related to the denaturation of proteins during the preservation of bighead carp heads[31]. Peak C could be attributed to NADH, which was consistent with the results reported by Wu et al.[32].

      Figure 3. 

      Three fluorescence components peaks (a) A, (c) B, and (e) C in fish fillets and their corresponding excitation-emission wavelength loading diagram peaks (b) A, (d) B, and (f) C.

    • Based on the span of wavelengths of the fluorescence peaks and the position of the maximum fluorescence intensity, Peak A and Peak B could be attributed to tryptophan. Therefore, Peak AB was used to represent Peak A and Peak B. In Fig. 4, two fluorescence peaks (Peak AB and Peak C) were presented in the contour maps of fish fillets stored at different times and temperatures. The fluorescence intensity of Peak AB was significantly higher than that of Peak C of bighead carp heads throughout the preservation period (except for keeping at 16 °C on the first day). The fluorescence intensity of Peak AB gradually rose with prolonged storage time under isothermal conditions and the simulated cold chains, which followed the conclusions drawn by Shi et al.[2]. The fluorescence intensity of tryptophan was enhanced with the increasing pH of bighead carp heads in the final phase of storage due to the increase in electron density and the deprotonation of the amino group in the benzene ring[33,34]. Fluorescent signal intensity of Peak C gradually improved with increasing storage time, which is consistent with the results of Hassoun & Karoui[35]. The high NADH content resulted in the reduction of NAD to NADH during storage of bighead carp heads due to the reduction in the enzyme activity of NADH after fish death[30]. The fluorescence intensity of Peak C decreased at 16 °C on the 3rd day (Fig. 4p), which could be attributed to the phenomenon of fluorescence quenching induced by elevated temperature[36].

      Figure 4. 

      Changes of fluorescent components in fish fillets under (a)−(p) different storage conditions (0, 4, 8, 12, and 16 °C), storage conditions in the (q)−(s) SDS, and (t)−(v) LDT cold chain.

      The increase in temperature accelerated the deterioration of the quality of bighead carp heads and affected the change in the fluorescence intensity of the fluorescence peaks under various storage conditions, which was analogous to the changing trend pertaining to the freshness indicators of the bighead carp heads under various storage conditions. Therefore, fluorescence peaks have great potential in tracing the freshness changes of aquatic products, and fluorescence matrix of excitation and emission spectroscopy technology may be used as a monitoring tool for the rapid assessment of fish freshness.

    • According to the results of the PARAFAC analysis, three types of fluorescent components could be extracted and determined for the construction of freshness prediction of the LSTM network model. The model that emerged was organized into three distinct layers: an input layer, a concealed layer, and an output layer. The input layer was composed of storage time (day), storage temperature (°C), and trait fluorescent parts of EEMs extracted through PARAFAC analysis. The output layer consisted of TVB-N, TBARS, K, and TVC values. To verify the suitability of this simulation of the actual cold chain, experimental data collected at 0, 4, 8, 12, and 16 °C were utilized as the coaching subset, and gathered experimental data in the SDS and LDT cold chains were utilized as the validation subsets. RMSE and losses continued to decrease rapidly after 200 iterations within 7 s and were stable after the 140th cycle during the training progress, which demonstrated that the EEM-LSTM network model could well predict the freshness of bighead carp heads after multiple training cycles.

      The R2, RMSE, MAE, and MAPE values for the effectiveness evaluation of the EEM-LSTM model are given in Table 1. For predicting changes in the TVB-N, TBARS, K, and TVC values, the EEM-LSTM model had RMSE values of 0.8887, 0.0214, 2.1773, and 0.2740, respectively, R2 values of 0.8840, 0.9581, 0.9801, and 0.9224, MRE values of 4.25%, 6.60%, 7.50%, and 3.72%. The MAE values were all below 1.9282, and the MAPE values were all below 0.750. The EEM-LSTM freshness prediction model had high R2 values and low RMSE and MAE values, which indicated that the EEM-LSTM network model had good predictive performance. The prediction of the K value resulted in an R2 value of 0.9801, which was higher than that of the other indicators, while the prediction of the K value resulted in the highest RMSE, MAE, and MAPE values, probably because the number of K value was much larger. The study showed that the MAE, MSE, RMSE, and MAPE values indicate the quality of the regression only relative to other types of regression performance and not in an absolute manner, in contrast to the R2 value[37]. Therefore, the EEM-LSTM model showed the best performance fitting effectiveness for the K value. When integrating LSTM with optimized EEM, it is possible to predict the freshness of bighead carp under non-isothermal storage parameters. As a non-damaging detection technique, fluorescence spectroscopy combined using deep learning models can monitor the freshness of bighead carp in real-time but there is still a lot of research space. Future research can investigate the use of this method for predicting the freshness of other fish. In addition, It would be valuable to investigate whether extreme environmental conditions (such as high or freezing temperatures) impact the accuracy of the model's predictions for bighead carp freshness.

      Table 1.  Performance evaluation of four freshness indexes of bighead carp head predicted based on EEM-LSTM freshness prediction model of fish flesh in the SDS and LDT cold chains.

      Indicators R2 RMSE MAE MAPE MRE (%)
      TVB-N 0.8840 0.8887 0.6241 0.0425 4.25
      TBARS 0.9581 0.0214 0.0188 0.0660 6.60
      K value 0.9801 2.1773 1.9282 0.0750 7.50
      TVC 0.9224 0.2740 0.2233 0.0372 3.72
      R2, coefficient of determination; RMSE, root mean square error; MAE, mean absolute error; MAPE, mean absolute percentage error; MRE, mean relative error.

      For the SDS cold chain, the proportional errors in the TBARS, TVB-N, K, and TVC values were below 8%, 9%, 10%, and 9%, respectively. For the LDT cold chain, the proportional errors in TVB-N value, TBARS value, K value, and TVC value were below 11%, 12%, 17%, and 12%, respectively (Fig. 5). The relative errors of the LDT cold chain were higher than those of the SDS cold chain, which could be related to the frequency of temperature fluctuations and the range of temperature fluctuations. Overall, the findings indicate a high level of agreement between the measured values and the predicted values of the EEM-LSTM model. The EEM-LSTM network model reported in this study could simultaneously predict four indicators, in contrast with the network models reported in other studies[9], indicating that the EEM-LSTM network model can effectively predict the change rules of indicators in simulated cold chains.

      Figure 5. 

      Prediction of TVB-N, TBARS, K, and TVC values by the EM-LSTM model based on fish fillets in the (a)−(d) SDS, and (e)−(h) LDT cold chains. Note: A0, A1, A2, A3, and A4 in the figure correspond to five storage time points in the SDS cold chain respectively; B0, B1, B2, B3, B4, and B5 correspond to six storage time points in the LDT cold chain respectively.

    • In this study, excitation-emission matrix fluorescence spectra of fish fillets under isothermal and simulated cold chains were analyzed to study the relationship between fluorescent components and fish freshness. An EEM-LSTM network model was then established to concurrently predict the TBARS, TVB-N, K, and TVC values. By employing residual analysis, core consistency diagnostic, and half-split analysis of the PARAFAC algorithm, three crucial fluorescent components were successfully extracted. The fluorescence intensity of the fluorescent components of samples kept at different temperatures increased and consistently linked to the time-temperature history, which could be used to indicate the freshness of the heads of bighead carp. The EEM-LSTM freshness prediction model effectively forecasted the condition of bighead carp heads, showing high predictive accuracy with a high R2 value and low MSRE value, and closely aligned with the increasing trends of all four freshness indicators, which were influenced by temperature fluctuations in simulated cold chains, including both the range and frequency of fluctuations. The relative errors of the four freshness indexes predicted by the EEM-LSTM model were less than 10% and 17% in the SDS and LDT cold chains, respectively, and the total average relative error was less than 7.50%. The EEM-LSTM network model could predict the freshness of bighead carp heads more intelligibly and quickly, and it was demonstrated that excitation-emission matrix fluorescence spectroscopy had a certain potential in the freshness detection of bighead carp heads, which provided a theoretical foundation for the freshness prediction of aquatic products in actual cold chain scenarios.

      • This study was supported by the Beijing Academy of Agriculture and Forestry Sciences Outstanding young scientist training program, the Fund of Young Beijing Scholar, Jiangsu Science and Technology Plan (Key Research and Development Plan Modern Agriculture) Project (BE2023315), and the Ministry of Finance and Ministry of Agriculture and Rural Affairs: National System of Modern Agricultural Industry Technology (CARS-45-28).

      • The authors confirm contribution to the paper as follows: study conception and design: Shi C; draft manuscript preparation: You J; conceptualization, visualization and supervision: Sun Z; methodology: Li X, Fan H; writing - revision & editing: Fan H; statistical and mathematical methods to data analysis: Ying X. All authors reviewed the results and approved the final version of the manuscript.

      • The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

      • The authors declare that they have no conflict of interest.

      • Copyright: © 2024 by the author(s). Published by Maximum Academic Press on behalf of China Agricultural University, Zhejiang University and Shenyang Agricultural University. This article is an open access article distributed under Creative Commons Attribution License (CC BY 4.0), visit https://creativecommons.org/licenses/by/4.0/.
    Figure (5)  Table (1) References (37)
  • About this article
    Cite this article
    You J, Sun Z, Li X, Ying X, Shi C, et al. 2024. Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains. Food Innovation and Advances 3(4): 405−415 doi: 10.48130/fia-0024-0037
    You J, Sun Z, Li X, Ying X, Shi C, et al. 2024. Noninvasive freshness evaluation of bighead carp heads based on fluorescence spectroscopy coupled with long short-term memory network: simulation of cold chains. Food Innovation and Advances 3(4): 405−415 doi: 10.48130/fia-0024-0037

Catalog

  • About this article

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return