Machine Learning-Based Analytical Systems: Food Forensics

Abstract

Despite a large amount of money being spent on both food analyses and control measures, various food-borne illnesses associated with pathogens, toxins, pesticides, adulterants, colorants, and other contaminants pose a serious threat to human health, and thus food safety draws considerable attention in the modern pace of the world. The presence of various biogenic amines in processed food have been frequently considered as the primary quality parameter in order to check food freshness and spoilage of protein-rich food. Various conventional detection methods for detecting hazardous analytes including microscopy, nucleic acid, and immunoassay-based techniques have been employed; however, recently, array-based sensing strategies are becoming popular for the development of a highly accurate and precise analytical method. Array-based sensing is majorly facilitated by the advancements in multivariate analytical techniques as well as machine learning-based approaches. These techniques allow one to solve the typical problem associated with the interpretation of the complex response patterns generated in array-based strategies. Consequently, the machine learning-based neural networks enable the fast, robust, and accurate detection of analytes using sensor arrays. Thus, for commercial applications, most of the focus has shifted toward the development of analytical methods based on electrical and chemical sensor arrays. Therefore, herein, we briefly highlight and review the recently reported array-based sensor systems supported by machine learning and multivariate analytics to monitor food safety and quality in the field of food forensics.

1. Introduction

Recently, it is assessed that approximately 420 000 people pass away every year as a result of foodborne ailments due to food poisoning; consequently, nearly $100 billion has been spent every year on its treatment worldwide as per World Health Organization (WHO) records, which may be increased drastically in the future due to rapid population growth.¹ Therefore, the need of the hour is to deal with food safety and monitoring, forestall contamination universally, and to develop novel ways to overcome food contamination to reduce the antagonistic effects of food contamination on the worldwide population. Therefore, food forensics plays a dominant role in the identification of food contaminants to help keep a safe food supply throughout the world. Food forensics is the investigation of food products to prove their authenticity as well as traceability by utilizing analytical techniques.² Typically, food authenticity depends on the detection of nucleic acid, sugar content, different food contaminants, etc. Food contamination is broadly classified into three categories, namely, chemical, biological, and physical contamination. Chemical contamination can be either direct or indirect, which may include different hazardous chemicals such as pesticides, heavy metals, toxins, and fertilizers as well as different types of food additives. Direct contamination includes the replacement of expensive authentic products with low-cost additives, similar counterparts such as artificial flavoring agents, the addition of colorants, false labeling, bogus naming, etc. in order to sell at low cost, whereas indirect contamination may occur from pesticides, fertilizers, insecticides, small toxin molecules, etc., which are being used to protect crops from pests, disease transporters, etc. Statistically, it has been assessed that only 0.1% of pesticides act on the targeted organism. The unconsumed toxins penetrate the climate gratuitously, which may further be found in the same state or derivatized products in food products. These types of contaminations are being analyzed in food forensically by using vigorous analytical methods such as Mass Chromatography,^3,4 UV–visible absorption Spectroscopy,⁵⁻⁹ Fluorescence Spectroscopy,^6,10−12 Surface Plasmon Resonance Spectroscopy,¹³⁻¹⁵ etc (Figure 1). Biological contamination includes the growth of various microorganisms that may cause unsuitable sensory modifications, through the production of metabolites, which results in change in appearance/odor.¹⁶ The primary contaminating sources of spoilage organisms are zoonotic in nature, which can exist in sewage, water, air/dust, insects, workers, live animals, etc., and their complete elimination from food is not possible. For instance, various spoilage organisms such as Escherichia coli, Staphylococcus aureus, and Salmonella causes outbreaks of food poisoning and leads to the death of human beings.¹⁷ Apart from these zoonotic organisms, fungal contamination can easily grow at low pH in high humidity conditions.¹⁸ These molds or yeasts can produce low molecular weight metabolites called mycotoxins such as zearalenone, which is produced by Fusarium fungi, aflatoxin by Aspergillus flavus, etc.¹⁹ These mycotoxins damage the central nervous system, which may lead to paralytic attack and death.²⁰ Similarly, allergic contamination of food arises mainly due to the addition of a limited quantity of food allergen with another food. It is estimated that 1–2% of adults and 5–8% of children of the total population in Western countries are affected by food allergies.²¹ The major effects of an allergic reaction vary from itching through sneezing to gastrointestinal reactions as well as life-threatening diseases such as anaphylactic shock.²² In general, the biological contaminations are analyzed using polymerase chain reaction and immunoassay methods, which are highly specific for the particular analytes. Immunoassay-based methods are broadly classified as chemiluminescence, radioimmunoassay, enzyme-linked immunoassay, or lateral flow immunoassays.²³⁻²⁶ Further, apart from chemical and biological contaminations, physical contamination may occur due to the addition of unfamiliar material such as plastics, hair, cloths, glasses, bones, metals, pest bodies, etc., to food, which can appear at any phase of food preparation and packaging. Many physical food contaminants are highly stable and non-biodegradable, which may cause infertility, neurological disorders, cancer, leukemia, nerve blockage, etc. Moreover, the physical contaminants may also carry spoilage organisms, representing a much more serious gamble. These types of contaminations are analyzed using a metal detector machine or X-ray machine, which generates radio graphing by utilizing an electric X-ray tube as demonstrated in Figure 1.

Figure 1.

Schematic representation showing types of food contaminations with their detection methods.

Literature reveals that most of the analytical methods are based on the detection of a specific analyte for each type of food contaminant. However, food analysis is much more challenging due to the complex matrix of different analytes, their combination in different ratios rather than individual analyte, which results in a false positive signal/report. Thus, it is a well-known fact that the simultaneous analysis for presence of all analytes present in food by utilizing these sophisticated instruments is time-consuming, costly, complicated, and requires trained manpower. Therefore, the development of analytical methods/devices that can provide the on-site detection of hazardous analytes in the field of food forensics and environment monitoring is the need of the hour. Thus, the most effective method to lower the false-positive reports is to establish multiplexed sensors targeting several analytes to provide a series of signal outputs. It is well-established that the machine learning (ML) can precisely discriminate and accurately identify analytes and successfully fit complex data, which makes machine learning widely adapted in food forensic. Nowadays, machine learning techniques, which include supervised, unsupervised, and reinforcement learning, are being utilized in food forensics, environmental analysis, bioanalytical systems, clinical diagnostics, etc. Numerous optical and fluorescence-based machine learning approaches have been reported; the most common methods are principal component regression or analysis (PCR or PCA), multiple linear regression (MLR), partial least-squares discriminant regression and analysis (PLSR or PLSDS), linear discriminant analysis (LDA), hierarchical clustering analysis (HCA), and their combinations. Therefore, herein we describe some basics of machine learning, i.e., how machine learning works along with analytical techniques for detection of food contaminants in a complex matrix. Further, the utilization of numerous machine learning algorithms toward the detection of different types of food contaminants has also been highlighted along with fabrication of portable devices for food forensics, which overcome the limitations of conventional techniques. Additionally, some commercially available integrated hand-held devices based on machine learning have also been discussed.

2. Recent Trends in Food Forensic: Introduction of Machine Learning for Analytical Application

In general, the major disadvantage of conventional analytical methods is the complexity of analysis in the field of food forensics. The detection of a specific analyte is not always feasible due to structural similarities among various analytes. Apparently, fabrication of such highly selective sensors is time-consuming as well as costly, especially for applications in complex matrices such as environmental analysis, food forensics, bioanalytical systems, clinical diagnostics, etc. Consequently, to overcome these problems, nowadays array sensors have emerged as the best computational methods when combined with conventional analytical techniques. Array sensors can detect and quantify multiple analytes simultaneously in a complex matrix system and, thus, can be best applicable for food forensics. These systems work on the segregation of multiple similar analytes by utilizing the various cross-reactive components. Array sensors do not selectively distinguish the target analyte; however, a specific pattern develops, called a fingerprint, of all results that are produced after the interaction of sensors with analytes. To further differentiate the obtained fingerprint responses, various machine learning-based approaches have been utilized. As a discipline of artificial intelligence, machine learning contrasts from conventional problem-solving strategies. A machine learning framework learns from models or trains a model with information to get familiar with its parameters through optimization and makes predictions utilizing new information (model-based learning). A practical requirement in daily living is the quick, precise, and automated identification of food qualities. To identify food qualities, modern methods including electronic noses, computer vision, spectroscopy, and spectral imaging have been extensively applied. These methods can gather a lot of digital data about food qualities. Due to the fact that the vast amount of data contains a lot of redundant and unnecessary information, data examination for these approaches is crucial. It is difficult to put these strategies to use in real-world situations because dealing with such a massive volume of data and extracting usable characteristics from the gathered data is a pressing and crucial issue. Therefore, to cope with the vast amount of data, several data examination or analysis techniques have emerged including unsupervised and supervised techniques as shown in Figure 2. The training of machine learning algorithms can be done with several approaches, each with their own benefits and drawbacks. Unsupervised learning, as its name suggests, is a machine learning technique in which models are not supervised using training data sets. It describes issues in which only inputs are given without any corresponding output or labels.²⁷ Unsupervised learning can be further classified into two categories of problems such as clustering and association. The most popular and simplest clustering algorithm is the k-means clustering algorithm, which groups unlabeled data sets into various clusters.²⁸ The primary goal of this algorithm (k-means) is to minimize the sum of the distance between each data point and each of its relevant clusters.²⁹ The requirement of selecting k (number of clusters) is a major disadvantage of k-means and related algorithms,³⁰ whereas this issue can be solved in hierarchical clustering by creating a tree structure known as a dendrogram that reflects an ensemble of clustering models with all possible values of k.^31,32

Figure 2.

Flowchart of the types of machine learning techniques.

Principal Component Analysis (PCA) falls under the category of an unsupervised descriptive approach and allows the dimension reduction of the data by removing the redundancy in the data set while retaining the discriminatory abilities of the array.³³⁻³⁵ Although a PCA score plot allows us to visualize general trends in the overall data set for analytical applications, for better classification of data and predictive analysis it is preferred to use other tools such as LDA, artificial neural network (ANN) (supervised learning approach), etc., along with PCA. Supervised learning is a kind of machine learning in which machines are trained using “labelled” data, and machines anticipate the result based on the input; for instance, LDA is a supervised technique that retains the information about the class of analyte during the transformation of the data. LDA attempts to increase the separability among different classes while minimizing the intraclass separation, i.e., separation among data points of the same class. The initial data, i.e., training set is utilized to construct a linear classifier for discrimination of analyte into multiple classes. The efficacy of the model is determined using a cross-validation procedure, and finally the optimized model can be utilized for the analysis of an unknown data set. In sensor array applications, LDA has recently become the most often used supervised algorithm,^36,37 whereas another supervised learning approach, i.e., Support Vector Machine (SVM), can be applied to both classification and regression problems.³⁸ The fundamental objective of the support vector machine approach is to define the ideal decision boundary that can divide an n-dimensional space into classes, enabling the quick assignment of new data points to the appropriate categories. SVM can be more advantageous than discriminant analysis since it does not need a lot of training data to do fruitful discrimination.³⁹ SVM and ANN, two different techniques, share a similar idea of utilizing linear learning models for pattern identification. Fundamentally, SVM uses nonlinear mapping to make the information linearly distinguishable; thus, kernel function is the key. However, an artificial neural network utilizes multilayer connection to deal with linear issues.^40,41 In comparison to earlier classification methods such as PCA and LDA, artificial neural networks offer a number of advantages since the architecture can be designed for both classification and descriptive tasks. Although an ANN allows one to solve a large range of problems and, thus, requires a large data set, consequently, there are chances of underfitting or overfitting the model during training of the network model. Therefore, other optimization techniques are often incorporated, which reduces the chances of overfitting or underfitting the model, thereby resulting in more accurate predictions. In this context, Artificial Intelligence (AI) is the process of imitating human intellect in machines that are programmed to think and behave like humans. In contrast to traditional computer programs, which produce outputs based on an explicit set of instructions, Artificial Intelligence systems are intended to make predictions using data-driven models. Thus, machine learning has been widely used in various sectors, being a powerful tool for analyzing data. Since conventional machine learning approaches cannot decipher raw natural data, these approaches typically need to be augmented with human component extraction strategies. A machine can extract characteristics from unstructured data for detection, classification, or regression via representation learning. Therefore, deep learning is a kind of representation learning that utilizes profound ANN, which is made up of many layers of neurons, to improve multilayer representation (nonlinear modules). In research domains, Convolutional Neural Networks (CNN) are as of now recognized as the most widely utilized machine learning approach for enormous data handling. ML is a rapidly developing area in chemical analysis for a variety of reasons. First, the vast volume of data generated by the new generation of instruments necessitates the application of novel methodologies to extract crucial data pertaining to a sample composition or a chemical reaction. Second, advanced information technologies and the Internet enable data mining from a variety of sources, which encourages the development of specialized representative data sets, such as those that describe the toxicity of a particular class of chemicals or typical reaction and biochemical pathways.⁴² The quality of food and agricultural products must meet people’s expectations and standards for food processing. As a result, it has evolved into a time-consuming and labor-intensive operation. Machine learning’s remarkable precision and efficiency can save labor expenses and potentially outperform human performance. For instance, the quality of fish and fruits/vegetables can be determined by a variety of physical factors such as color, weight, shape, ripeness, etc., but machine vision can collect numerous other parameters that cannot be noticed by human vision.⁴³

3. Applications of Machine Learning in Food Forensics

Over the past three decades, research in the area of optical chemical sensors has been flourishing. Researchers in this field have published a variety of research papers and books that highlight the benefits of optical sensing over other transduction techniques.⁴⁴ In today’s world, the need for efficient approaches that enable the quick, sensitive, compact, and affordable detection of hazardous analytes in food is urgent. Array-related sensing approaches, often known as electronic noses/tongues, interact with analytes via physical adsorption, which results in an electrical response. When an analyte interacts with a sensor array, a response pattern is produced that is noticeably different, allowing molecular detection by correlation with a specified library of responses. Dodd and Persaud established the idea of an electronic nose as a tool to imitate the human olfactory system’s ability to distinguish between different odors.⁴⁵ They employed three distinct metal oxide gas sensors, and their steady-state signals identify a number of substances. Due to the active development of chemical sensing approaches that do not rely on electrical responses that are now available, particularly prominent among them are optical sensors. The majority of optical sensors rely on colorimetric/fluorescent modulations that result from binding of the chromophore/fluorophore molecules with analytes. Array-based approaches with digital imaging techniques enable the interaction with each analyte differently and therefore generate a unique fingerprint pattern for different analytes. The rising popularity of array-based sensing led to fabrication of colorimetric and fluorescent sensor arrays for a variety of analytes to perform sensing in solution phase as well as vapor phase. In the beginning phase of colorimetric sensor arrays, identification and detection of multiple toxic analytes in food was done utilizing pattern recognition approaches such as HCA, PCA, and LDA. After seeing the excellent results produced by colorimetric sensor arrays, other machine learning-based approaches such as ANN and SVM have been utilized because training and testing of the developed sensor array models and validation of the results are also important. Furthermore, for a practical requirement in daily lives, namely, the quick and automated identification of food qualities, research has been focused on the development of portable devices that can generate unique fingerprint patterns after the interaction with multiple target analytes or can generate data that can be further analyzed by using various machine learning approaches. In order to regulate the food sector, quality assurance of foods and drinks has become important. It is crucial to identify contaminated foods or beverages from the authentic ones. For this reason, the kind of fingerprinting that sensing arrays can carry out can be quite beneficial. In this regard, Suslick et al. designed the colorimetric sensor array in the beginning for the naked eye recognition of various volatile organic chemicals (VOCs) utilizing a number of different metalloporphyrins,⁴⁶ wherein sharp color changes resulted from the coordination of analytes to metalloporphyrin, and the patterns of these color changes were employed to distinguish analytes.

3.1. For Protein-Rich Foods/Meat Products

One of the primary issues in modern times is to supply healthy food that is free of toxic residues, pesticides, and allergens as well as pathogen contamination (bacteria, virus, or protozoa). The development or growth of microbes can contaminate the food and also affect the texture, flavor, smell, and color of the food. The primary determinants needed to check for meat freshness during storage are sulfur compounds (thiols) and biogenic amines, which are byproducts of microbial decarboxylation of amino acids. Thus, a quick detection of these byproducts is necessary to avert a calamity caused by these. In this regard, two fluorescent cucurbits[n]uril probes were reported for the construction of a supramolecular assembly by Minami et al. that can detect and measure the nitrosamines as shown in Figure 3A,B.⁴⁷ To monitor the aging of chicken flesh, Salinas et al. designed an optoelectronic nose using 16 pigments embedded in porous silica or alumina.⁴⁸ The colorimetric fluctuations of the array-based sensor were a sign of aged chicken meat in a packing environment (30% CO₂-70% N₂), wherein the color responsive data were analyzed using both PCA and Partial Least Square (PLS) methods as shown in Figure 4A,B.⁴⁸ Xiao-Wei et al. fabricated a nanoporous color responsive array-based sensor composed of TiO₂ film that was obtained by a sol–gel method, for the detection of trimethylamine gas (10 ppm–60 ppb).⁴⁹ The advantage of utilizing nanoporous surfaces increases the reactivity rate, and it provides stability to color responsive sensor arrays. Furthermore, this fabricated nanoporous color responsive sensor array was utilized for the detection of trimethylamine in Yao-meat, with correlation coefficient values of 0.896 and 0.837 in the training and prediction sets, respectively, as determined by PLS. Over the past decade, microfluidic paper-based color responsive sensor arrays have gained a lot of interest due to various advantages such as low cost, easy to use, easy fabrication, etc. Using cross-reactive vapor-sensitive dyes enclosed in resin microbeads, Chen et al. fabricated a paper-based color responsive array-based sensor.⁵⁰ These dyes reacted vigorously with the volatile organic compounds (VOCs) that were abruptly released during spoilage of chicken. To quantitatively analyze the aging and eventual rotting of chicken under various temperature circumstances, an affordable smartphone camera was used to collect color data from the sensor array barcode.

Figure 3.

(A) Diagram showing partial quenching of fluorescence of Cucurbit[n]uril Probes after binding with a metal ion. (B) Qualitative LDA results of sensor array probes with biological amines, cancer-associated nitrosamines, and tobacco alkaloids. Reproduced with permission of ref (47). Copyright 2012, American Chemical Society.

Figure 4.

(A) Change in color of the developed sensor array during chicken aging process after interaction with atmosphere. Numbers demonstrating the days after packing. (B) The PCA score plot representing diverse chicken aging days. Reproduced with permission of ref (48), Copyright 2012. Royal Society of Chemistry.

On the other hand, Li et al. have also designed a color responsive sensor array for trimethylamine detection in both aqueous and gaseous phases.⁵¹ With more than 99% accuracy, distinctive color change patterns enable simple trimethylamine discrimination over a wide concentration range. Computed detection values (4 ppb in gaseous phase and 2.3 μM in aqueous phase) are much lower than the diagnostic trimethylaminuria quantity (fish malodor syndrome) (10 ppm).⁵¹ Aldehydes were identified as Warmed-Over Flavor (WOF) markers of cooked meat by Kim et al., who developed a color responsive sensor array to detect warmed-over flavor in cooked food and thiobarbituric acid (TBARS).⁵² HCA and PCA were used to assess every sample and correctly forecast changes in WOF concentrations.⁵² For better classification of data and predictive analysis, other algorithms such as LDA, ANN, etc. including PCA have recently become the most often used. Li et al. constructed an affordable sensor array using 12 chemically sensitive dyes oriented on a silica gel flat plate that exhibits a distinctive fingerprint pattern response when reacted with hazardous volatile chemicals produced from pig samples.⁵³ Pattern recognition techniques such as LDA and Back Propagation Artificial Neural Network (BP-ANN) were utilized for the data analysis. Findings of the experiment revealed that the BP-ANN algorithm gave good results as compared to LDA. Discrimination percentage values for the training and prediction sets were 100% and 97.5%, respectively.⁵³ Nguyen et al. developed an affordable paper-based colorimetric probe to detect gaseous forms of ethanolamine, trimethylamine, dimethylamine.⁵⁴ The color response of the developed probe was examined using a color model (RGB) from a mobile phone application.⁵⁴ Dicyanovinyl-substituted oligothiophene derivatives (Figure 5A) were developed by Liu et al., and the effects of various end-cappers on the photophysical properties were examined.⁵⁵ Additionally, zinc porphyrin derivative NA-3T-CN has been developed for the simultaneous identification of aliphatic amines and hydrazine via a colorimetric array-based sensor, and it demonstrated good results in the case of optical response and with different concentrations of analyte in selective single sensor and in multisensory node as shown in Figure 5B. Further, a readout system for smartphones and data processing based on RGB changes in the sensor array were carried out, and the capacity to distinguish between hydrazine, aliphatic amines, and aromatic amines was satisfactorily high as shown in Figure 5C. Depending upon the fluorescence shifts of the developed sensor, the detection value for hydrazine is 1.22 × 10^–5 mol/L in tetrahydrofuran (THF), which is used as a solvent.⁵⁵ Orouji et al. designed multicolor sensor arrays based on gold nanorods and gold nanospheres in order to discriminate and identify biogenic amines such as spermine, tryptamine, ethylenediamine, tyramine, spermidine, histamine, etc., in meat and fish samples.⁵⁶ Furthermore, the detection limit of each analyte has also been calculated and found to be 24.6, 4.79, 8.58, 14.26, 10.03, and 27.29 mmol L^–1 for spermidine, spermine, tryptamine, histamine, and tyramine, respectively. Various data visualization and pattern recognition techniques (LDA, PLS) were used to visually evaluate and statistically analyze the data.⁵⁶

Figure 5.

(A) Diagram showing structures of Dicyclovinyl-Substituted Oligothiophene derivatives with different end groups. (B) (a) Zn-TPP-based sensor array plate after interaction with different amines. (b) Zn-TPP-coated TLC plate showing colorimetric response against varying concentration of hydrazine media. (c) Mechanistic illustration for detection of hydrazine by utilizing Zn-TPP fluorophore. (C) RGB value-based differentiation of amines. Reproduced with permission of ref (55). Copyright 2020. American Chemical Society.

Ma et al. developed a color responsive sensor array, which is made up of ice-templated Dye@chitosan/UiO-66-Br and optimized by Density Functional Theory (DFT) and Grand Canonical Monte Carlo (GCMC) simulation, which successfully detected the gases released from shrimp (ammonia, methylamine, and trimethylamine with limit of detection of 37.17, 25.90, and 40.65 ppm, respectively) and formed a unique fingerprint.⁵⁷ On the other hand, by synthesizing complexes of pH-sensitive dyes and nanomaterials, Zhong et al. examined the discrimination of eight biogenic amines with more than 92% accuracy using pattern recognition approaches and measurement of trimethylamine with a detection limit up to 1.3 ppb.⁵⁸ Further by taking cognizance of pattern recognition techniques, Du et al. fabricated a sensor array utilizing gold nanoparticles functionalized with molecules that have carboxylate moieties for the detection and discrimination of amines by hydrogen-bonding and electrostatic interactions as shown in Figure 6A,B.⁵⁹ Additionally, a fabricated sensor array was utilized for a real sample investigation of raw fish; it exhibited precise detection of histamine concentrations (7.2 ppm).⁵⁹ Cai et al. devised a cellulose-based fluorescent ink system consisting of curcumin and its derivative using a combination of ethanol and water for the detection of shrimp freshness.⁶⁰ Three detecting units with various dyes were created by screen-printing when the rheological features of the inks were improved. Upon exposure to 0–50 ppm, each sensor label showed different properties on the basis of optical signals, high reactivity of NH₃, and fluorescent shifting paths due to intramolecular charge transfer.⁶⁰ According to the volatile basic nitrogen (TVB-N) and total viable counts (TVC) levels, Xu et al. developed the color responsive sensor array (CSA) coupled with whale optimization algorithm (WOA) and back-propagation neural network (BPNN) for the monitoring or identification of beef freshness.⁶¹

Figure 6.

(A) (a) Diagrammatic representation showing synthesis of functionalized Au nanoparticle. (b) Au nanoparticle sensor showing colorimetric change with 10 different amines. (B) Corresponding LDA plot for the differentiation of 10 amines. Reproduced with permission of ref (59) Copyright 2022, Royal Society of Chemistry.

3.2. For Beverages and Sweeteners

Artificial sweeteners account for 62% of the sweetener production industry and are utilized in a variety of products, from medications to fizzy drinks. Since 1977, the use of many non-nutritive sweeteners, including saccharin, aspartame, acesulfame-K, and sucralose, have been authorized by the U.S. Food and Drug Administration (FDA). Cyclamate, a high-intensity sweetener that is legal in more than 100 countries worldwide, was outlawed in the U.S. because of possible ties to bladder cancer in rats. Using an affordable color responsive sensor array as shown in Figure 7A, Musto et al. were able to recognize 14 different sweeteners, both organic and synthetic in origin, at biological pH (Figure 7B).⁶² A sequence of ormosil-encapsulated pigments was fixed on a hydrophobic, porous membrane to form the array.⁶² Han et al. designed ionic fluorescent dyes to recognize 30 different whiskies based on their fluctuating fluorescence intensities.⁶³

Figure 7.

(A) Color difference maps of 14 natural and artificial sweeteners and one control with a sensor. (B) PCA score plot for discrimination of six natural and artificial sugars. Results showing two patterns: (1) the close conjunction between natural sugars and (2) the proximity of all three artificial sweeteners. Reproduced with permission of ref (62), Copyright 2009. American Chemical Society

Further, to differentiate commercial beers, Qian et al. reported a sensor array made of nanosized polydiacetylene vesicles as shown in Figure 8A,B.⁶⁴ The head groups and the diacetylene composition used for the polydiacetylene (PDA) sensor’s self-assembly had a significant impact on the sensor’s ability to detect different concentrations of ethanol and distinct flavored additives and to separate between beers. One sensor demonstrated a minimal detection value (0.19 mM) with the greatest colorimetric factor (15.9%) for 4-Vinyl Guaiacol, whereas another sensor displayed minimal detection value (0.15 mM) with the largest colorimetric response (12.5%) for diacetyl.

Figure 8.

(A) Different derivatives of diacetylene monomer utilized for the formation of sensor array. (B) Diagrammatic representation of differentiation of beers by utilizing PDA sensor array. Reproduced with permission of ref (64) Copyright 2020. American Chemical Society.

Lyu et al. constructed a paper-based chemosensor array device to concurrently classify 12 saccharides and quantify fructose and glucose among those 12 analytes as shown in Figure 9A,B. The rationale for the saccharide detection is based on an Indicator Displacement Assay (IDA) on the paper-based device utilizing four kinds of catechol dyes, 3-nitrophenylboronic acid, and saccharides.⁶⁵

Figure 9.

(A) (a) Diagrams of the catechol dyes and 3-nitrophenylboronic acid for the sensing of saccharides with (b) detection mechanism and color change behavior by utilizing PCSAD. (c) Image of the developed PCSAD. (B) (a) LDA plot showing discrimination of 12 targets using PCSAD. (b) SVM regression analysis results for fructose detection in a soft drink. Reproduced with permission of ref (65), Copyright 2021, American Chemical Society.

3.3. For Pesticides

It is a well-known fact that, in the fields, farmers have used various pesticides and organophosphates. Because of this, these contaminants may get into the food chain via air, water, or soil, posing a major risk to both human and animal health. Organophosphates and carbamates are the two main types of pesticides that have been the most utilized because of their comparatively weak persistence under normal conditions and strong ability for eradicating insects and pest. Therefore, in order to detect and discriminate these pesticides/organophosphates, various commercially available dyes such as fluorescein, 4-methylumbelliferyl phosphate, and PAMAM dendrimers as shown in Figure 10A were used for the development of a fluorescent turn-on sensor array by Liu et al. (Figure 10B,C).⁶⁶ The rationale behind the detection depends on an indicator displacement assay. The developed sensor array can detect and identify inorganic phosphates, glyphosate, and methylphosphonate simultaneously throughout the concentration range from 10 μM to 2 mM.⁶⁶ Further, Qian et al. have also devised a colorimetric sensor array using five easily accessible, inexpensive thiocholine- and H₂O₂-sensitive indicators for recognition of organophosphates and carbamates.⁶⁷ The basis of detection by a developed array is based upon the ability of organophosphates and carbamates to irreversibly block acetyl cholinesterase (AchE) activity. Apparently, this inhibition prevents the release of thiocholine and H₂O₂ from S-acetylcholine and acetylcholine and consequently results in diminishing the color responses to thiocholine- and H₂O₂-sensitive indicators.⁶⁷ Unsupervised classification methods such as PCA and HCA revealed that the proposed sensor array can recognize and separate different organophosphates (including carbaryl, methomyl, metolcarb, isoprocarb, and fenobucarb) and carbamates in real samples such as apple juice and green tea.

Figure 10.

(A) Diagram showing fluorescence shifts generated by the competitive binding between the quenched G5–dye complexes and the organophosphate. (B) An optical pitcher showing array’s response toward the target, showing its affluent cross-reactive behavior. “A” denotes absorbance, “F” denotes fluorescence emission, and “P” denotes fluorescence anisotropy. (C) LDA plot showing discrimination of five phosphates. Reproduced with permission of ref (66), Copyright 2014, American Chemical Society.

For the purpose of detection and differentiating organophosphate-based insecticides, Kashani et al. constructed a color responsive sensor array utilizing citrate-capped gold nanoparticles (AuNPs) as shown in Figure 11A.⁶⁸ The spectrum modulation as a result of the aggregation of AuNPs upon the addition of organophosphates has been assessed using pattern-based approaches such as HCA and LDA as shown in Figure 11B.⁶⁸ Furthermore, the constructed sensor array has also been evaluated for the identification of analytes in a complex mixture of rice and paddy water, which was analyzed using an LDA-based technique, and results revealed that the different concentrations of organophosphates were segregated in different groups. Gao et al. synthesized N-(aminobutyl)-N-(ethyl isoluminol) functionalized with graphene quantum dots (ABEI-GQDs) with a typical measurement of 7–8 nm.⁶⁹ These prepared ABEI-GQDs displayed outstanding chemiluminescence features upon interacting with H₂O₂ and wavelength-tunable fluorescence emission with increment in excitation wavelength and were further utilized for distinguishing pesticides in orange juice with the help of unsupervised algorithms such as PCA and LDA.⁶⁹ Further, Sun et al. have also developed a fluorescent sensor array that uses an ensemble of carbon dots and metal ions for the determination and differentiation of five phosphate anions.⁷⁰ Additionally, a real sample application of the sensor array is approved by the effective recognition of phosphates in serum and unknown samples.⁷⁰ Zhao et al. constructed a laser-induced fluorescent device for the identification of pesticide residue utilizing a sensor array approach.⁷¹ Four pesticide residues (carbendazim, diazine, fenvalerate, and pentachloronitrobenzene) were easily identified by their ability to produce distinctive fingerprint-like response patterns and can also be distinguished by using an unsupervised pattern recognition technique, i.e., Spectral Recognition Method (SRM). The device showed good response toward concentrations below 10 ppb. Further, the practicability of this device was also demonstrated by investigating a spiked cabbage sample, and the recovery rate was found to be 95.17–105.76%.⁷¹

Figure 11.

(A) Schematic representation for color responsive array-based detection of organophosphates on AuNPs. (B) 2D canonical score plot showing discrimination of pesticides by utilizing nanoparticles. Reproduced with permission of ref (68) Copyright 2016. American Chemical Society.

3.4. For Dairy Products

It is well-known that the utilization of antibiotics in animal husbandry has significantly increased in recent years. Because of the broad spectrum of antimicrobials, aminoglycoside antimicrobials (AAs) are being frequently utilized as feed additives to treat animals and promote animal growth. However, the overuse of antibiotics, in animal husbandry, may result in accumulation of antibiotic residues in the animal bodies and their derived products. Consequently, the consumption of animal-derived meals may put people at high risk. In this regard, Mungkardnee et al. have constructed a fluorescent array-based sensor for the evaluation of milk.⁷² The fluorescence results of four receptors were analyzed after the addition of milk samples utilizing excitation wavelength at λ_ex = 375 nm, and the findings were examined using pattern recognition methods.⁷² Further, the LDA technique has been utilized to distinguish thermal processing-based pasteurized, sterilized, and recombined milk. Yan et al. developed label-free ssDNA-AuNPs conjugates for the discrimination and quantification of five AAs (tobramycin (TOB), streptomycin (STR), gentamycin (GEN), ribostamycin (RSM), and kanamycin (KAN)) at concentration ranges of 120–280 nm as shown in Figure 12A.⁷³ Furthermore, a complex matrix of five AAs in water as well as in milk was analyzed using Fisher linear discriminant and HCA, and results revealed that the sensor array could identify five AAs with 100% accuracy as well as discriminate a mixture of STR and gentamicin as shown in Figure 12B,C.⁷³ Long et al. fabricated a quadruple fluorescent array-based sensor containing label-free quantum dots for the purpose of determining and differentiating a variety of tetracyclines (chlortetracycline, oxytetracycline, tetracycline, and doxycycline).⁷⁴ When tetracyclines (TCs) and carbon dots (CDs) were combined, dampening of the fluorescence was observed. The discrimination of these TCs basically depends on the difference of binding affinity with CDs, which showed a different fluorescence pattern after a PCA analysis. The limit of differentiation of the fabricated sensor array was found to be as low as 1 mM.⁷⁴ To differentiate powdered infant formula based on their species, place, label, etc., Zhao et al. created a three-unit fluorescence array-based sensor made of PPE₂ (poly(p-phenylene ethynylene), PPE-SO₃, and PPE-N₁ that gives nonspecific interactions with compounds in the infant formula and generates a response signal, which was utilized to discriminate 24 distinct infant formula samples using an LDA approach.⁷⁵

Figure 12.

(A) Detection principle for AAs by utilizing ssDNA-AuNPs conjugate. (B) Discrimination of tobramycin, streptomycin, gentamycin, ribostamycin, and kanamycin and the mixtures of gentamycin and streptomycin by utilizing FLD analysis. (C) HCA investigation of AAs in the presence of TOB, KAN, GEN, RSM, STR, and the mixtures of GEN and STR Reproduced with permission of ref (73), Copyright 2018, Elsevier.

4. Machine Learning-Inspired Devices in Food Forensics

Standard techniques used in laboratories for chemical food safety assessment generally depend upon liquid or gas chromatography along with mass spectrometry. Although these techniques are acknowledged as best for quantitative confirmatory investigation, they involve sampling, transporting the samples to laboratory where they are examined by professional staff, and the utilization of costly equipment. In this manner, there is a rising need for mobile, portable hand-held gadgets or devices to give fast, effective, and on-site screening of food contaminants. The essential strategy for assessing the nature of food quality is sensory analysis, which depends upon the utilization of human senses such as smell, taste, vision, etc. In this manner, scientists have developed artificial senses, which may act as the replacement of human senses for fast as well as effective detection and quantification of food contamination, such as electronic nose, electronic tongue, paper-based devices, etc. However, considering the proliferation of such affordable and portable sensors, alternative designs for the related read-out devices must be taken into account. To ensure broad utilization of these sensing devices, features such as field portability, affordability, and network connectivity are all desired. An electronic nose is an array-based device utilized for the quick identification and discrimination of distinct types of contaminants on the basis of their odorant. This device contains chemical sensor units that produce specific odor profiles called fingerprints, after interacting with gaseous mixtures. Subsequently, the obtained odor profiles can be evaluated by comparing with standard odor profiles, and further validation of the obtained responses was pursued using machine learning based approaches. Initially, an electronic nose was used to examine air pollutants that had already been identified by the olfactory system. Currently the use of this instrument is more widespread since it involves the evaluation of fluids and O₂ gas mixtures.⁷⁶ For the purpose of detecting foods with protein content, Li et al. constructed a disposable chemical sensor array joined with a hand-held gadget for the rapid evaluation and recognition of the freshness of five meat items.⁷⁷ This hand-held gadget comprises an on-board diaphragm pump and one-dimensional (1D) complementary metal oxide semiconductor (CMOS) camera to capture on-site color data as shown in Figure 13A. This constructed sensor array is profoundly responsive toward amines and sulfides at the parts per billion level (Figure 13B).⁷⁷

Figure 13.

(A) Gas sampling from a meat sample by utilizing a hand-held analyzer with its 20-element color responsive sensor and its cartridge. (B) Dendrogram representing HCA results of four gases discharged by spoiled meat. Reproduced with permission of ref (77), Copyright 2016, American Chemical Society.

Further, a 36-element color responsive array-based sensor was constructed by Li et al. for the rapid and easy detection of alcoholic beverages including brandy, scotch, bourbon, and vodka.⁷⁸ Aldehydes, ketones, carboxylic acids, polyphenols, sulfides, ethanol, and various other major components of alcoholic beverages as well as a number of other volatile organic chemicals are highly sensitive to this newly developed sensor array.⁷⁸ After a 2 min vapor exposure, each of the 14 alcoholic drinks or the control showed distinct color difference patterns that were simple to detect even with the bare eyes as shown in Figure 14A,B.

Figure 14.

(A) Picture showing color responsive array-based sensor with its peroxidation tube agent and sampling by utilizing a hand-held analyzer. (B) (a) Diagram representation of preoxidation of liquid vapors and (b) its colorimetric array response toward target samples. Reproduced with permission of ref (78), Copyright 2017, American Chemical Society.

Even though electronic noses have undergone substantial development, few of them have reached the market because of sensing or pattern identification problems. For instance, metal oxide-based electronic noses often need a complex power supply and wiring to operate at high temperatures. These metal oxide-based sensors are humidity-sensitive and unable to differentiate between chemically identical substances.⁷⁹ Thus, in order to commercialize electronic noses, we require a system with a strong cross-reactive sensor array and data processing technique that can reliably anticipate fingerprint patterns and extract data from nonlinear data sets. In this regard Guo et al. developed a color responsive barcode that is composed of 20 different kinds of porous nanocomposites composed of dye-loaded chitosan nanoparticles embedded on cellulose acetate that can simultaneously recognize fragrance fingerprints and intelligent fingerprints, and the developed model was trained by a deep learning convolutional neural network as shown in Figure 15.⁸⁰

Figure 15.

(A) Barcode recognition by DCNN with its input, multiple convolutions, full connection (blue circles), and output layers. (B) With increment in epochs, training accuracy also increases. (C) Euclidean distance computations and DCNN training are used to compare the detection accuracy rates for freshness of chicken, fish, and beef. (D) Confusion matrix exhibiting accurate classification of three freshness categories by utilizing DCNN. (e) Image exhibiting a smartphone interface that, after scanning a barcode, indicates whether the meat is fresh (left panel), less fresh (middle panel), or spoiled (right panel). Reproduced with permission of ref (80), Copyright 2020, Wiley.

For the analysis of liquid samples, an electronic tongue is employed, which works on the principle of human sense of taste. Multiple components present in the liquid that is under investigation can be analyzed simultaneously with this type of equipment. This device is used in the sectors of environmental monitoring, food processing, and medical sciences (detection of pathogens in liquid samples). Gutierrez-Captain et al. constructed an electronic tongue made of various electrochemical sensors and a color responsive optofluidic device for the categorization of white grape juices as shown in Figure 16A,B, and the information obtained from the device was analyzed with multivariate approaches such as PCA and Soft Independent Modeling Class Analogy (SIMCA).⁸¹

Figure 16.

(A) Image of the microsensor chips utilized for detection of pathogen and (B) optical MIR system. (C) Electrochemical output for one sensor and reference grape juices. (D) Multivariate analysis-based discrimination of reference genotypes. Reproduced with permission of ref (81), Copyright 2013, American Chemical Society.

Among other applications, paper-based analytical devices (PADs) are increasingly used in clinical and health determination, contamination monitoring, food quality testing, and drug quality evaluation. PAD-based assays are regarded as point-of-need assays since they are less expensive, simple to use, and convenient.⁸²⁻⁸⁴ Digital or Computerized picture-based colorimetric identification is a regularly employed approach that leverages color information captured in the digital image for the interpretation of assay findings due to the widespread utilization of digital cameras, especially smartphone cameras.⁸⁵ A smartphone image-based colorimetric detection is a feasible and convincing field-based approach in contrast to established approaches such as optical techniques. In order to determine which machine learning model could most effectively estimate the quantification of analytes on paper devices, Khanal et al. examined four machine learning models—Logistic regression, SVM, random forest, and ANN with three color models (RGB, HSV, and LAB).⁸⁶ They used pictures of paper-based devices taken at different light conditions, with various camera and enzyme inhibition assays, to make training and test data sets. The expectation of precision was higher for food color than enzyme inhibition assays in the majority of machine learning models and color space combinations.⁸⁶ Pounds et al. designed a rapid, on-site food spoilage identification technique utilizing a smartphone application that can check and examine the color of a novel designed sensor film installed inside a quick response (QR) sticker to recognize features related with food deterioration. This strategy incorporates an inherent food categorization machine learning technique from the developed QR sticker with fabricated pH responsive sensor film in Figure 17A–C.⁸⁷

Figure 17.

(A) Images showing color variations of sensor film during the course of a storage period and at various pH levels. (B) Developed QR sticker with a fabricated pH responsive sensor film placed in the center and three color standards (red, blue, and green) placed at the three corners. (C) Categorization of pork loin samples according to their better (blue dots) or poor (red dots) quality samples utilizing a k-NN algorithm (k = 31). The categorization border is shown by the line. Reproduced with permission of ref (87) Copyright 2022, American Chemical Society.

Despite the existence of quick and accurate analytical techniques for detecting biogenic amines that are present in food products, extensive attempts have been made to develop portable and affordable instruments for discriminating biogenic amines in food items to accomplish on-site detection of food spoilage. Singh et al. developed a field-deployable cross-reactive array-based sensor and mobile plate reader for the identification of biogenic amines as shown in Figure 18A.⁸⁸ The sensor array comprises metal complexes of azo-dye-based ligand that produced effective results upon interaction with distinct target analytes such as spermine, histamine, spermidine, tryptamine, creatine, histidine, cysteamine, etc. After that, color-responsive results and the discrimination viability of the developed array-based sensor were assessed utilizing pattern-based recognition approaches such as PCA and LDA as shown in Figure 18B.⁸⁸

Figure 18.

(A) Experimental setup of portable plate reader. (B) Multivariate analysis (PCA and LDA) of sensor array and reduced sensor array. Reproduced with permission of ref (88). Copyright 2021, Elsevier.

In the field of portable devices for food diagnostics, where the goal is to avoid the need of expensive instrumentation-based investigations, sample preparation is still a barrier. Sampling carried out by an untrained user might result in undesired contamination and defiled outcomes. The commercial devices shown have been used to make an effort to address this issue with a user-friendly design that includes instructions to help customers during the sample preparation and calibration process. In addition, because smartphones are so widely connected, they may use the vast computing power and storage that cloud computing provides as internet of things (IoT) modules.⁸⁹ Coskun et al. constructed a food allergy detection gadget called iTube that runs on a smartphone and captures the image and analyzes the color responsive assays conducted in test tubes for the rapid and accurate identification of allergies in food specimens. This gadget, which is mechanically linked to the smartphone’s existing camera unit, allows access to the test and control tubes from the side and illuminates them vertically using two different light-emitting diodes (Figure 19).⁹⁰

Figure 19.

(A) (a) Image of the food allergen testing device called iTube that works on a smartphone-based digital reader utilizing colorimetric assays. (b) The optomechanical attachment that is attached behind the cell phone. (c) An image of the same iTube platform is also depicted. (B) Diagram showing dose–response curve for peanut allergen detection by utilizing iTube device. Reproduced with permission of ref (90) Copyright 2012, Royal Society of Chemistry.

Nimasensor was created by Nima Laboratories Inc. and allows for the detection of gluten in meals. The Nima device is basically a lateral flow strip placed within the automated testing apparatus that extracts the test sample, does the lateral flow analysis, and identifies any positive result with an optical sensor as shown in Figure 20A,B. A specific app is used to deliver results. A smile is displayed for gluten levels under 20 ppm on the Nima device’s embedded organic light-emitting diode (OLED) display as shown in Figure 20D.⁹¹⁻⁹³

Figure 20.

(A) Representation of Nima strip with labeling. (B) Sample analysis of Nima strips with wheat, barley, and rye-gluten at mentioned concentrations. (C) Developed Nimasensor device and (D) its working for the detection of gluten in meals.⁹³

5. Conclusion

The freshness of food and its safety is one of the important parameters that need to be addressed for selecting food for consumption. Therefore, the development of portable food sensors is an emerging area of interest in analytical applications, which may provide more benefits as compared to other sophisticated instruments such as gas chromatography-mass spectrometry (GC-MS), high-performance liquid chromatography (HPLC), etc. In this context, with the emergence of machine learning and artificial intelligence-based applications, many scientists have focused their research interest toward the development of portable solutions for food forensics that are easy to use, rapid, robust, and do not require skilled manpower to operate. Taking cognizance of this, herein we have discussed a brief overview of the contaminants responsible for food spoilage and highlight various types of developed portable devices based on machine learning approaches to monitor food freshness and food spoilage. We believe that advancements in the field of machine learning and electronics will further facilitate the development of customized solutions for particular problems related with on-site food forensics.

Author Contributions

^⊥ Ranbir, M.K., and G.S. contributed equally to this work. All authors have given approval to the final version of the manuscript.

The authors declare no competing financial interest.