Predictive microbiology and machine learning by optimization productive process: Metanalysis

Introduction

The “traditional” approach to food quality and safety management is based on end-product testing. However, end-product testing provides only very limited information about the safety status of a food. The food safety crisis of the 1990s (Listeria, Salmonella, Escherichia coli, Campylobacter, dioxins, antibiotics, acrylamide, etc.) highlighted the failure of this traditional approach. If a dangerous organism is found, it means something; but its absence in a limited number of samples is no guarantee of the safety of an entire production batch. Testing of the final product is often done too late. Today, effective food safety and quality management systems must be based on prerequisite programs (good manufacturing practices and good hygienic practices), on the Hazard Analysis and Critical Control Point (HACCP) plan, as well as on quantitative tools, namely predictive microbiology and risk assessment approaches (Committee on the Review of the Use of Scientific Criteria and Performance Standards for Safe Food. & Institute of Medicine (U.S.), 2003; Garcia et al., 2019; Liu et al., 2019)

Accurate and appropriate testing for potential hazards in the food supply chain is key to ensuring consumer safety and food quality. Periodic monitoring and testing of ingredients can reveal fluctuations within the supply chain that may be an indicator of an ingredient’s quality or potential hazard. Such quality is assessed through standardized chemical and microbiological composition testing to meet the requirements and specifications of government agencies around the world. For raw materials or finished products to meet these safety and quality limits, their composition generally must have a low microbiological load and be chemically identical in macro components such as carbohydrates, proteins, and fats. Methods in this segment must avoid false negative results that can endanger consumers (Beck et al., 2021).

A worrying aspect concerning food safety management is the continuous and indiscriminate recourse to microbiological analysis of final products, the production logic assumes that sampling only the final product is enough. It is common to find the following statements: “if I sample all the batches of food I have verified the process”, “if I have a deviation in the process, I take more samples”, “quality is an expense and not an investment and microbiological test take up to 8 days, since traditional methodologies are used or whoever performs the analysis has no knowledge of how microorganisms work individually and within an biological system” and “microbiology is a Petri dish with some colonies” (Pérez-Lavalle et al., 2020). As a consequence of the above thoughts, “death certificates”, are issued because the product analyzed at the end of its process has no option to be adjusted or modified to meet the necessary requirements for its distribution; tons of food go to a return cellar or enter as “waste” to other links in the chain with millionaire losses. How to make processes more efficient? The answer lies in the combination of several worlds: microbial ecology, understood as the study of the interactions of microorganisms with their environment, with each other and with plant and animal species. Data mining or data exploration is a field of statistics and computer science concerned with the processing of data sets. It uses the methods of artificial intelligence, machine learning and database systems. And statistics the science of collecting, analyzing, presenting and interpreting data (Sarker, 2021). It sounds complicated, expensive and not very feasible, but no; when we understand the way in which microorganisms develop and the conditions that favor or not their development, we could predict how they can behave within a system and make it scalable using data mining to have patterns over time with significant levels of accuracy that complement the work done in laboratories with traditional, alternative and/or molecular tests. Think about something, to reproduce any living being requires: nutrients, time, temperature and variables such as pH, water activity (aw) and oxygen, if you know these requirements you can design production lines where they are enhanced or controlled to favor or avoid the emergence of certain microbial strains (Hibbing et al., 2010) . In the production of yogurt with probiotics of the Bifidobacterium type, it is required that after the maturation tanks, the passage of the yogurt base to flavoring is done without the inclusion of oxygen, since this lactic acid bacteria are strictly anaerobic and the entry of oxygen into the system causes its death (Fratamico et al., 2005). When meat dairy is elaborated, it is natural that the raw material contains Salmonella or Listeria monocytogenes if it is poultry or bovine origin respectively (Gonçalves-Tenório et al., 2018). The preparation of meat by-products implies the validation of cooking times and temperature as well as cooling times and temperatures for packaging in order to eliminate these two pathogenic bacteria. Generally, the systems are designed so that the heat treatment reaches 75ºC in the chamber and is maintained. This is because Salmonella achieves a logarithmic reduction to 70ºC every one minute and this genus has a thermal resistance of 4.5ºC and the temperature must be reduced to 4ºC in less than 2 hours since it has optimal growth between 30 to 45ºC (Ma et al., 2007). Pathogenic bacteria are not always measured in the processes due to cost issues and index microorganisms are used. An increase in the Escherichia coli process line correlates with an increase in the probability of Salmonella appearance, evolutionarily Escherichia coli has a common ancestor with Salmonella and conserves pathogenicity genes with this genus, which makes it an excellent indicator for this pathogen (Ma et al., 2007; Vega et al., 2013; Wessels et al., 2021).

Although the concept of predictive modeling was introduced starting in the 1920s with Bigelow`s model describing the inactivation kinetics of microorganisms or, alternatively, the heat resistance of microorganisms, it took about 60 years to develop models of microbial growth and death. Since then, predictive microbiology has focused on the development of primary deterministic models to predict microbial behavior in foods as a function of storage time (growth and survival) and treatment time (inactivation) with the help of secondary models describing the effect of intrinsic, extrinsic or processing factors on kinetic variables. Despite advances in this area, most models developed in the 1990s were based on deterministic approaches without taking into account the variability of factors affecting microbial response. Deterministic growth or inactivation models that produce “optimal” estimates are usually not optimal for successfully managing food safety. If, for instance, the downstream effects of unsatisfactory levels of a surviving pathogen in food after processing design, or if the growth limits of a foodborne pathogen at the single-cell level are not known, product formulation for safer food production would be a difficult issue. In addition, the use of a worst-case scenario approach to food processing leads to unrealistic estimates with a negative impact on food quality (Koseki, 2016; Mcmeekin et al., 2007; Stavropoulou & Bezirtzoglou, 2019).

The current challenge is how to implement these technologies in the “farm-to-table” industrial chain, since the main limitation of culture-depend methods is that they only allow the study of culturable microorganisms, therefore, there is a risk of not having a view of the biodiversity of the microbial population present in the food (Sabater et al., 2021). However, the use of open data obtained through different omics and analyzed under models, algorithms or predictive technologies or machine learning allows the extraction of information encrypted from the microbiome and identify if there are pathogenic and/or altering microorganisms. This information provides new opportunities to explore the diversity and functionality of microorganisms in food, as well as in food processing and production environments (Melnyk et al., 2020), (Cammarota et al., 2020) (Sabater et al., 2021) (Whon et al., 2021).

If we assume that microbiome research is the analysis of the composition, function and biotic and abiotic interactions of microorganisms in complex communities (Sakowski et al., 2019). Current microbiome analysis generally consists of two important components: upstream community profiles (what is the abundance of all microorganisms in each sample) and high-level downstream analysis (alpha/beta diversity analysis, differential abundance analysis). In recent years, evolving data analysis, visualization and machine learning methods have been gradually applied to the development of many software tools and web servers for microbiome data analysis covering these two components (Zhao et al., 2021).

By taking all of the above what we get is the observation of environmental factors, integration of the data into statistical models and prediction of bacterial behavior in food that allow us the identification of hidden patterns, leading to implementation of data-driven interventions. The food industry needs appropriate quality and safety testing technologies to monito and evaluate indicators in real time before the supply chain reaches the consumer, so that personnel at each link can make timely adjustments to ensure food safety and reduce economic losses (Kang et al., 2022).

Predictive microbiology

Predictive microbiology has been defined as the knowledge of microbial responses to environmental conditions, synthesized in mathematical models, which allows an objective assessment of processing, distribution and storage operations on microbiological safety and food quality. Predictive microbiology, which was initially based on three fundamental principles: first, the growth, survival and inactivation of microorganisms are considered reproducible responses; second, microbial behavior is dictated primarily by a limited number of environmental factors; and finally, by quantifying the combined effect of these factors, it is possible to predict the response of microorganisms (Chaix et al., 2015; González et al., 2019; Koseki et al., 2021).

Predictive models are the most effective mathematical tools for estimating changes in microbial levels in food as the product moves through the farm-to-table chain, and have been recognized as important components in the food safety and quality risk assessment process (Koutsoumanis et al., 2016; Possas et al., 2022). Databases o microbial responses in food as affected by intrinsic and extrinsic food factors, such as temperature, pH, and water activity (aw), are very useful for the development and validation of predictive models and risk analysis. Between 2016 and 2021, the following predictive microbiology software resources and tools have been released (Table 1) (Possas et al., 2022).

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 1

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 2

Table 1: Prediction models between 2016 and 2021.

The rapid progress of microbiology is inseparable from the simultaneous updating of observation methods or technologies. With the advent of the era of big data, the pressing problem for researchers has gradually evolved into how to quickly and efficiently filter/condense this exponentially increasing information to obtain general high-quality data, and how to transform massive microbiome data into easy-to-understand and easy-to-understand data. - to understand the data. visualize the data. Compared to traditional research, where insufficient data or technology leads to cognitive bias, low repeatability and long error times, the modern microbiology research process is likely to incorporate new technologies and big data methods to make it better and more accurate (Jiang et al., 2022).

Machine learning

Machine learning (ML), first proposed by Arthur Samuel in 1959, a core subfield of artificial intelligence (AI), is a proven technique that generates momentum in the domain of microbiology. It has been used so far in computationally intensive problems, such as predicting drug targets and candidate vaccines, diagnosing microorganisms that cause infectious diseases, classifying drug resistance, predicting disease outbreaks and exploring microbial interactions. ML has been applied to all areas of microbiology research, including virology and parasitology (Golden et al., 2019; Goodswen et al., 2021; Jiang et al., 2022).

Prior to ML, attempts to perform complex tasks such as letter and image recognition involved programming computers with rule-based systems. These systems tended to fail outside of test scenarios with unseen data. ML has shifted the paradigm from programming computers on how to perform a task to one in which computers learn how to perform the task without being programmed and has gained popularity in the era of big data in the scientific community. While mechanistic modeling relies primarily on simplified mathematical formulations to solve problems involving complex data sets, ML algorithms recognize patterns and/or underlying data structures and make predictions based on the patterns learned in those data sets (Tanui, Karanth, et al., 2022) ML is an extensive field of study that overlaps and inherits ideas from related fields such as statistics, computer science and AI, dedicated to explore to simulate or implement human learning behaviors, reorganize existing knowledge structures and continuously improve their performance to predict or classify other unknown data (Goodswen et al., 2021; Batool et al., 2022; Kang et al., 2022).

ML has two main learning modes, supervised or predictive to make future predictions from training data and unsupervised or descriptive, which is exploratory in nature with no training data, defined target or output. The goal of supervised ML is to find a model that best matches the input and output data. For instance, an image of a Petri dish containing a culture medium could be considered as “growth present” or “ growth absent”. A computer develops an algorithm to classify each image and compares the accuracy of the algorithm to the “supervised truth”; typically, the initial results do not provide adequate performance and the algorithm will iteratively adjust its parameters by trial and error to optimize accuracy (Mateo et al., 2021). Another example of supervised ML is its effectiveness in attributing food sources to clinical cases of foodborne diseases such as listeriosis from sequencing data. Inferring genetic information from pathogen genotypes is often crucial biological inference. Attribution of the origin of Listeria monocytogenes infections allows food industry professionals, data managers, epidemiologists, microbiologist, and bioinformaticians to adapt their practices to prevent the spread of foodborne pathogens. It also enables more efficient use of resources to contain the survival and proliferation of pathogens at their source (Tanui, Benefo, et al., 2022).

The use of ML algorithms in food safety since 2020 has increased in response to the complexity of quantifying some features and their effects on microbial population dynamics (Hiura et al., 2021). Table 2 discriminates the studies between 2020 and 2022 published in the food safety domain using ML.

Tanui, Karanth, et al., 2022. used ML algorithms to predict the presence of Salmonella in ground chicken meat. Using known genotypes as input features, the semi-supervised Random Forest showed the highest overall accuracy of 94% (95% confidence interval: 85%-99%) and a Kappa value of 82% and 87% prediction. The model predicted virulence-associated genes that could be used as features in predictive modeling efforts in the future, demonstrating its utility in different areas of food safety, including pathogen source identification, antibiotic resistance prediction and foodborne pathogen assessment.

Im et al., 2021 developed several ML models to predict the pathogenicity of Shiga toxin-producing Escherichia coli (STEC) using whole genome sequencing data. The input dataset for the ML models was revealed using distinct gene repertoires of positive (pathogenic) and negative (non-pathogenic) control groups in which each STEC isolate was designated based on source attribution, risk potential relative to isolate sources. Among the ML models examined, the support vector machine (SVM) model discriminated between the two groups most accurately. The SVM model successfully predicted the pathogenicity of isolates from major sources of STEC outbreaks, suggesting that the SVM model is a reliable and widely applicable method for assessing the pathogenic potential of STEC isolates compared to conventional methods.

Huang et al., 2022 conducted a study to define the growth and non-growth limit of Listeria monocytogenes under the use of different preservatives used in the meat industry and pH as control factors. They developed a multifactorial logistic regression model to calculate the probability of growth and then define the growth limit using 2 thresholds. With threshold 1 (p= 0.0104), the classification accuracy for growth events was 68,6% with a true positive rate of 77.6% and a true negative rate of 45.5% for growths obtained on PALCAM agar. At threshold 2 ( p= 0.04) the accuracy becomes 82.6% with a true positive rate of 95,5% and a true negative rate of 80% . This information can be used to formulate ready-to-eat- products that can prevent the growth of Listeria monocytogenes if stored at the optimum temperature.

Different algorithms have their own advantages and disadvantages, and there is no distinction between advantages and disadvantages. What needs to be done is to fully interpret the input data according to different demand scenarios, then build an appropriate model and constantly adjust it to achieve the best performance. Moreover, the belief that scientific problems will be solved only by using the most sophisticated and sophisticated models is not objective. Basically, computer technology simply helps people make decisions or automates the decision-making process and improves efficiency. Therefore, the choice of model should be the most suitable, not the most complex. There are four criteria to evaluate the advantages and disadvantages of machine learning algorithms (Greener et al., 2022) (1) Correctness, the most important criterion to evaluate the advantages and disadvantages of an algorithm (2) Robustness, that is, error tolerance. , which represents the algorithm's response and handling of unreadable data input. Chance. (3) Readable, easy-to-understand algorithms mean less time-consuming debugging, modification and extension processes. (4) Actuality, that is, time complexity and space complexity, indicate the amount of computation and memory space required to run the algorithm, respectively (Jiang et al., 2022)

Advances and applications of new experimental and computational methods drive the integration of microbial ecology research with cutting-edge technologies in computational biology and other fields. Simplify the scientific problem by predicting relationships between species or their environment (natural diversity, life cycles, interactions, and co-evolution) by altering prediction strategies based on species characteristics.

ML research is evolving rapidly, with rapidly changing architectures, algorithm combinations and computational strategies. The ultimate goal is not only to predict task accuracy, but also to discover the biological processes underlying scientific problems. The development of effective analytical tools, including ML, ensures data validity, proper annotation and open sharing and that most studies born from the intersection of microbiology and ML show results.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 3

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 4

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 5

Table 2: Studies between 2020 and 2022 published in the food safety domain using ML.

Materials and Methods

As a statistical analysis of a large collection of published study results, meta-analysis aims to integrate and interpret findings to achieve comprehensive conclusions that individual studies alone would not clearly demonstrate. An electronic literature search was conducted in Web of Science and PubMed databases to find articles and official reports published since 2018 that summarized the prediction of pathogenic microorganisms, mainly Listeria monocytogenes and Salmonella spp through supervised machine learning algorithms at any stage of food production globally. The search was performed systematically and aimed to find quality studies validated by the scientific community.

The literature reviews were performed using a syntax combining terms on the existence (prevalence, incidence, occurrence, quality, safety, food safety) of pathogens (Salmonella spp, Listeria monocytogenes, Campylobacter, Escherichia coli STEC) in food and their prediction through artificial intelligence (machine learning, data mining, models, algorithms, prediction), while other studies in the clinical domain, on insects and/or parasites, where products were artificially inoculated (inoculo, enriched) and published before 2018 were excluded. All terms were combined by correctly applying AND, OR and AND NOT logical connectors.

In the initial assessment, 323 publications were retrieved in the Web of Science database and 151 in the PubMed database, after the title review it was identified that the 151 PubMed articles were included in the 323 retrievals made in Web of Science, so we continued with this bibliography.

After assessing all the information in the retrieved publications, seventeen (17) primary studies published from 2019 to 2022 were deemed appropriate for inclusion as they met the following criteria: (1) presence of pathogenic microorganisms of public health importance; (2) present extractable data; (3) results on supervised machine learning algorithm scoring metrics: accuracy, F1 score, area under the curve (ROC-AUC), sensitivity, specificity and/or confusion matrix; and (4) results at any stage of food production globally.

The initial treatment of the 17 articles for the execution of the meta-analysis was performed with the creation of a matrix in Microsoft Excel 2010 whose main variables were: Reference, algorithms, model qualification metrics accuracy, F1 score, sensitivity, specificity, area under the curve and Cohen's Kappa coefficient, in addition to the type of data and application in food safety

Results and Discussion

A review of 17 articles associated with the use of machine learning with application in food safety was carried out in order to compare the information obtained in the literature, using the metafor program of the R software version 4.2.1 to perform the metanalysis.

-Metanalysis

The synthesis of the results of different previous studies (17) on the use of automated learning algorithms on the prediction of pathogenic microorganisms as a biological risk within food safety for the metrics: accuracy, sensitivity, specificity, ROC-AUC, Cohen`s Kappa coefficient and F1 score.

Under a restricted maximum likelihood estimation from the “n” of each item and the mean of each metric for the different machine learning algorithms, the accuracy obtained for predicting the presence of a pathogenic of 78.67% with an interval between 73.51-83.84%. For specificity 71.95% with an interval between 61.19-82.72%. An estimate for the area under the curve a value of 76.41% (interval 69.99-82.83% and for the F1 score 72.35% (interval 61.0-83.66%)

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 6

Figure 1: Results accuracy’s metanalysis.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 7

Figure 2: Sensitivity’s metanalysis.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 8

Figure 3: Specificity’s metanalysis.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 9

Figure 4: F1 score’s metanalysis.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 10

Figure 5: ROC-AUC metanalysis.

Predictive microbiology and machine learning by optimization productive process: Metanalysis - Image 11

Figure 6: Cohen´s Kappa coefficient’s metanalysis.

Conclusion

With these preliminary results, it was evident that machine learning algorithms are high impact tools in food safety and with greater data inclusion can deliver the best model to predict microbiological behaviors of pathogens of public health importance without the limitations of conventional culture-dependent microbiology. Since the coarse, overfitting occurs in the machine learning model.

We have an initial robust base to continue applying these models with more data to validate if the classification performed is consistent with these preliminary data and in the future have a model that can be presented to the food industry for its application in situ.

Supervised machine learning offers the opportunity to improve production processes, it does not eliminate conventional microbiology tests for process verification, it becomes an ally of food safety to predict behaviors at specific stages and obtain solutions in advance.

This article was originally published in “Predictive Microbiology and Machine Learning by Optimization Productive Process: Metanalysis". Acta Scientific Microbiology 6.2 (2023): 54-66. This is an Open Access article distributed under the terms of the Creative Commons Attribution License.