LABORATORIO DE URGENCIAS

 

 

Precision Nutrition: A Systematic Review

 

                                                                                                             Daniel Kirk; Cagatay Catal; Bedir Tekinerdogan

                                                                                                             Computers in Biology and Medicine; Jun 2021; vol 133 (104365)

 

 

Remarkable progress has been made over the last few decades in understanding how nutrition interacts with health. However, despite this abundance of knowledge, health conditions related to nutrition are rampant and, in some cases, increasing. Statistics from the World Health Organization show that obesity has almost tripled since 1975, diabetes has almost quadrupled since 1980 and raised blood pressure has almost doubled since 1975. The multifactorial nature of these conditions makes pinpointing their exact etiology difficult, although one idea that has emerged in recent years is that current approaches to managing these conditions and others do not take into account interindividual variability. Evidence for recommendations for healthy eating guidelines is often obtained from epidemiological or large clinical studies, wherein averages or generic cut-off points are made in an attempt to supply nutritional advice on a population level. However, such generalisation, although practical, fails to capture the individualized nature of the biological effects of nutrition. Such variability is known to exist in bodyweight in response to the same dietary intervention, postprandial glycaemia, physiolgical response to salt, caffeine metabolism, vitamin metabolism, and likely many other areas. Such variability can be attributed to factors such as sex, ethnic origin, genetics, metabolic traits, environment, microbiome composition, and probably other yet to be discovered factors. Hence, the concept of precision nutrition (used synonymously here with personalized nutrition; both abbreviated PN) on an individual or stratified level has been put forward as an answer to this problem.

Aside from the management of chronic diseases, nutrition personalization is also of use conditions requiring specific dietary considerations. Phenylketonuria(commonly known as PKU) is such an instance and is also one of the earliest examples of nutrition personalization. Patiens with PKU have mutations in the gene coding for the enzyme responsible for converting pheylalanine to tyrosine. A diet restrictive of phenylalanine and tyrosine supplementation are the only ways to avoid grave complications]. The case of PKU represents a fundamental example of how personal information about an individual (in this case, genetics) can shape dietary requirements.
Personalized approaches to nutrition would have applicability in the maintenance of general health and for athletes maximising sports performance. It is already the case that genetic testing to supply nutrition advice (among other information) is becoming commercially available and gaining interest. Some studies have also shown increased adherence or more effective behavior change in response to personalized approaches. For example, the Food4Me was a large randomised controlled trial investigating personalized versus generic nutrition advice for inducing dietary behavior change. After 6 months, it was clear that the personalized advice groups implemented and sustained more dietary changes thought to be better for their health than the generic group. Results such as these suggest promise that nutrition personalization can improve the health of individuals to a greater degree than generic, population-level advice.

Whilst conceptually PN may be appealing, PN approaches can involve the processing of lots of data of different kinds in a way that has not been possible in the past. However, the development of big data analytics, cloud computing, artificial intelligence, and machine learning (ML) has facilitated such data processing in a way and on a scale unmatched by humans. For PN, this means that complex arrays of factors can be integrated to provide precise nutritional advice on an individual or stratified level, facilitating prediction of postprandial glycemia], triglycerides] and the prediction of cancer. In these scenarios, the use of sophisticated techniques such as ML and deep learning (DL) to interpret multiple factors is of great utility.
Aside from the final output of PN, ML is also helpful in the data collection stages required to obtain the data used as features (i.e., the input) for the model. The number and type of features for PN models highly depend on the desired outcome but they can contain in themselves large amounts of data. Common features that are currently being investigated for ML application to acquire data include energy (food and drink) intake, physical and sedentary activity across the day, glycemia and sleep tracking. It is likely more are to come as new features are further identified or methods are developed that facilitate ML-orientated data extraction and data processing.

This integration of ML into both prediction models for PN and data extraction for PN is exciting for the prospect of deriving more accurate PN models. For this reason, knowing how and in which situations ML can be applied would help facilitate future PN work. However, until now this has not been explored in detail. This forms the basis for the motivation of the current review. The literature across multiple databases including Web of Science, Scopus, PubMed, and Science Direct was systematically searched to find all literature that was related to PN and used ML in their methodology.
The objectives were to provide an overview of where and how ML has been used in PN from various aspects, what such ML models use as input features, what the availability status of the data used in the literature is, and how the models are evaluated.
To the best of our knowledge, this is the first Systematic Literature Review (SLR) study that synthesizes the research performed in Precision Nutrition.

The contribution of this review to the literature is it provides a base for all information relevant to PN-related research utilizing ML, which is currently lacking. Researchers and practitioners can use this review as a reference to gain an understanding of the application of ML in PN-related research areas, inspiring future work, and progressing the research area.

Precision nutrition

Precision nutrition is a relatively new discipline, and this is reflected in its nomenclature. There is no universally agreed-upon definition for the terms precision nutrition or personalized nutrition. In some cases, the terms are used with close overlap, whilst elsewhere a distinction between the two is attempted.
Since there is currently no consensus, the present review makes no distinction between the two. One thing that can be said about these types of approaches, however, is that they aim to use personal information about individuals or groups of individuals to deliver nutritional advice that, theoretically, would be more suitable than generic advice. Note that PN can occur on a group level and still be considered personalized as long as the groups are made based on key characteristics that make the nutritional advice the same for all members within the same group. This is known as stratification and can be considered as a level above PN on an individual level. According to Zeisel, personalization on a stratified level is the real goal of PN since personalization on an individual level will never be possible]. Whilst it is certainly true that stratified approaches will be suitable enough in the vast majority of cases, the concept of individualzation does not seem unachievable in some circumstances. Predicting postprandial glycemia seems to be one instance where an individual approach could be applicable and suitable. In this regard, Zeevi et al., in 2015 published one of the most prominent papers in PN research. Prediction of glycemia for each individual was attempted based on meal content, meal timing features (e.g., time of consumption, time since prior meal, etc.), activity, blood features, continuous glucose monitoring (CGM) data, and data about the microbiome. Although the methods of assessment in this research may currently be infeasible on a large scale, that may change in the future as data gathering methods become more affordable, and this would certainly be an example on an individualized level. However, it could also be that groups of individuals within the data can be identified that respond in the same way to the same meal. The level of detail that PN reaches to will ultimately depend on how much the differences within the same stratified group make to the final prediction outcome; how well these differences can be detected by the technology in use; and the cost-effectiveness trade-off between these two. Indeed, taking these points into account, stratification seems likely to be the dominant choice.

PN is founded upon the concept of biological variability between individuals in response to nutrition. Thus, if the variables responsible for causing this variation and their effect on a desired outcome variable can be known, the outcome variable can be predicted, and this can be translated into nutrition advice. What, then, are these variables? The answer to this question depends on the desired outcome variable. There is no set of fixed variables that will provide any given output. Instead, features thought to be of importance to predicting the outcome are selected on a per situation basis. In some cases, this can reach to large numbers of individual features. However, they can be separated into groups, here referred to as PN elements. One common PN element is genetics. Genetics is understood as a reason for many obvious examples of variation, such as eye color and hair color, and this is extended to response to nutrition]. Indeed, in some circumstances such as PKU, genetics is an extremely relevant feature for PN approaches. There is also some known relationships between genetics and weight management], lactose (as in the case of lactose intolerance), metabolic syndrome, and more.
However, unlike with eye and hair color, what has become clear is that genetics can rarely explain nutritional response entirely. In some cases, the genetic contribution is virtually absent, as Berry et al. witnessed when predicting postprandial triglycerides. Another relevant factor is not only genes alone but also their interaction with nutritional intake, termed nutrigenomics. Genetic variation impacts metabolism of dietary components, but also dietary components regulate gene expression and signaling. Failing to account for this interaction will naturally lead to compromised accuracy of PN models, meaning dietary information is often collected in PN approaches. Gene-diet interactions for various chronic conditions are known and as more continue to be discovered, PN approaches considering nutrigenomics can be improved.

Dietary information is also collected independent of genetics, as a feature in its own right. Information on diet is particularly important in PN approaches to bodyweight management. In some cases, not only dietary features in the long term but also the content of an individual meal and the timing features of the meal (e.g., timing of the meal, time elapsed since the previous meal, etc.) are required to be known. This is the case in research investigating postprandial meal responses, where the composition of an individual meal in relation to its postprandial effect is relevant to know. Meal timing features are relevant due to their impact on health. Metabolomics is an increasingly popular field that quantifies the presence of small molecules in a sample with high accuracy using sophisticated techniques such as nuclear magnetic resonance and mass spectrometry. As the field of metabolomics develops further and these techniques become more frequently used, metabolomics will have a role to play in PN such as by investigating how different individuals metabolize foods and by establishing phenotypes. However, the measurement of clinically relevant biochemical parameters measured with traditional methods (i.e., not assessed from a metabolomics perspective) is currently more commonly seen and represents features in the group of clinical biochemical parameters. Included here are common clinical measures such as blood-sugar, hormonal levels, blood counts, and other parameters deemed to be relevant for a given PN intervention. Other PN features are the microbiome, due to its emerging role in health and relationship with nutritional intake; activity parameters (PA amount and intensity, sedentary behavior, and energy expenditure (EE)), due to their established interaction with health and disease; anthropometric features, such as height, weight, body mass index (BMI), etc.; and personal features, which includes information about individuals that can have an impact on model outcome such age, medical information and disease status, medication use, socioeconomic status, stress, and sleep.
Indeed, it is likely that certain components of these elements will be separated out to become elements in their own right as their perceived importance changes. It is also true that this will differ between studies, as feature importance differs greatly between research topics.

Machine learning

ML can aid in multiple stages of PN including data extraction, such as gathering dietary and PA data, and in integrating the features of the model to provide the output.
The algorithm learns patterns within the dataset(s) and uses these patterns to make a maximum likelihood prediction about the outcome.Some common ML algorithms include random forest, decision trees, support vector machines, k-means clustering, Multi Layer Perceptron (MLP),and Baye classifiers.
Four types of ML can be considered:

• Supervised Learning.The data used to train the algorithm has labels (i.e., the output variable is known). Once the task has been completed by the algorithm, the labels allow a way to check how well the task was performed by comparing the predicted values to the actual values (i.e., the labels on the data). Human intervention has a large role to play in supervised learning and can thus be considered time-consuming and expensive. This is true not only for data labelling but also processing of the data, such as algorithm feature selection (the features the algorithm uses to generate the output) and parameter selection (modifiable constraints inherent to the model).
• Unsupervised Learning In contrast to supervised learning, labels are not present for the data in unsupervised learning. Hence, the algorithm looks for patterns within the data in order to complete the task. Unsupervised approaches may also be used for feature selection as a preprocedding step so that only features of relevance are used in a subsequent main ML task in order to reduce the correlation. Although accuracy cannot be assessed, evaluation methods do exist for unsupervised approaches. For example, a well-known unsupervised learning task is clustering, which consists of grouping data together based on similar features. Here, measures such as cluster purity (i.e., the extent to which each cluster contains a single class) can be used.
• Semi-supervised Learning.As the name implies, this contains a portion of both supervised and unsupervised. Labelling occurs on only a very small portion of the data (e.g., 10%–20%) whilst the rest remains unlabelled. This tries to capitalise on the benefits that each offers, i.e. higher accuracy, and lower time and cost of operation for supervised and unsupervised, respectively.
• Reinforcement Learning. Actions are taken by an agent in a virtual environment to achieve an outcome. Depending on this outcome, the action is either rewarded or punished. The algorithm updates itself in response to this in order to maximise reward. Complex tasks in a dynamic environment are suitable for reinforcement learning application. Algorithms based on this learning type are applied in online games and autonomous vehicles.

ML algorithms can have their work divided into tasks. Six common ML tasks are listed below:

• A supervised approach to assigning unseen data values to a given class based on the properties it has. Binary classification is common, where the data can be categorized into one of two classes (i.e., 1 or 0, yes or no). Multiple classification is also possible, wherein class number is greater than two. An example of classification could be predicting presence or absence of disease from medical variables.
• A supervised task that takes a collection of input variables and uses them to predict a real numerical value as an outcome variable. Predicting blood cholesterol from relevant physiological variables is an example of regression.
• An unsupervised method of grouping portions of data together based on similar characteristics. Because it is unsupervised, the logical underpinning that ultimately drives the grouping process may not be apparent beforehand. Hence, patterns can be identified in the data that humans would be unable to notice. Grouping subjects together based on shared characteristics such as metabolic phenotype is an example of clustering.
• Recommendation systems ultimately use the information available to it to predict the preference a user will have for an output variable. Historical data about the user is used to predict preference, although this differs depending on type of recommender system (i.e., collaborative filtering or content-based systems).
• Dimensionality Reduction.Dimensionality reduction refers to transforming high—dimensional data to low-dimensional data, typically as a preprocessing step before performing a task. This means that only input variables that contribute to the model output are maintained for model input. Reducing input variables in this way improves model performance. It is also possible to reduce the number of data points (i.e., rows) in addition to the features (i.e., columns).
• Anomaly detection. This refers to the process of identifying results that deviate largely from what could be expected [AD]. However, because anomalies, by definition, occur only rarely, having sufficient samples for training data can be an issue. Hence, such anomaly detection algorithms attempt to respond to this issue. Anomaly detection has its most common application in detection of fraudulent bank transactions.

Deep learning (DL) is a sub-branch of ML. It is based on artificial neural networks (ANNs), which are networks designed based on the neuronal connections in the human brain. The term “deep” is added to reflect the number of hidden layers the network has, and this extra depth allows the network to deal with a greater level of complexity than shallow learning (i.e., traditional machine learning) approaches. In this way, DL can deal with certain complex tasks that shallow learning would not perform adequately. However, to do this they require a great deal more data and computational power. Although, if this data can be used for the algorithm, performance will increase, unlike with shallow learning algorithms, which tend to plateau. Note that ANN is not a DL technique per se; its categorization as deep or shallow depends on its depth, namely the number of hidden layers. A shallow ANN with a single hidden layer can be referred to as a Multi-Layer Perceptron (MLP). Examples of DL techniques include

Deep Belief Networks, Restricted Boltzmann Machines, Recurrent Neural Networks (RNN), Long Short Term Memory (LSTM), and Convolutional Neural Networks (CNN)..

Deep Belief Networks are structured the same as MLPs but are trained differently. Restricted Boltzmann Machines (a type of ANN) are stacked upon one another, and patterns recognized from the previous layer is used to train the next. This is repeated across all layers until the output is generated. This can be done unsupervised, where features are detected, or by providing a small set of labelled samples to be associated to the patterns. Either way, this saves largely on labelling time. RNNs differ from traditional ANNs in that they deal with sequential data, which means the input order of the data also has meaning. This is the case in sentences of words, for example, where the word order is relevant to convey information. This is achieved by using both new data and previously processed data as input, instead of only forward propagation as in traditional ANNs. Hence, instead of being propagated once only, the network is propagated a number of times equal to the number of sequential steps in the sample. Long Short-Term Memory is the most common RNN algorithm used. Text generation, such as in chatbots, translation, and speech recognition can be attempted using RNNs.

CNNs are also based on neural networks and are specialized in pattern recognition, making them suitable for the task of image recognition. Since CNNs are largely used for image recognition generally and were mostly used for image recognition in the current literature, they will be explained from an image recognition point of view. CNNs contain convolutional, pooling layers, and fully connected layers stacked on top of one another. In the convolutional layer, filters take an array of pixels (i.e., a small portion of the entire set of pixels) as an input to generate features, which fundamentally represent pieces of information that are distinctive for the image or objects in the image. This array convolves across all the pixels in the whole image, calculating scaler products, and generates features at all positions to form a feature map (or activation map). This is then transformed by a rectified linear unit, making negative values in the scaler product zero, and used as input to the next layer. Pooling layers take each of these filtered arrays in the feature map and make a much smaller image by taking the highest number from each scaler product (i.e., down-sampling), and this again acts as the input for the next layer. Doing this allows the most distinctive features within the image to be retained whilst making the overall size much smaller, reducing computational power. It is often the case that convolutional and pooling layers are stacked multiple times before reaching the final fully connected layer. The fully connected layer forms the final output as in a standard ANN via classification. To avoid overfitting in CNN-based models, dropout and batch normalization layers are also utilized.

 Dietary Recommendation.

Nutritional Management of Chronic Disease is a group composed of two articles that recommended healthier food options to users of the systems according to their chronic disease status. J. C. Kim & Chung developed a system that recommends healthy foods based on the user's physical and mental health through dietary nutrition, food preference, and healthcare personal data in an ANN (J. C. In the proposed system, data regarding the user's body and mental status is collected through online services and smart devices. A hybrid approach is used to overcome the shortcomings of each individual recommender system used. For example, collaborative filtering is used to predict preference based on correlation with the preference of other users but leads to what is known as the cold-start problem, wherein an insufficient amount of data is present to generate any outcomes; however, the use of a neural works is able to overcome this. Testing was performed on 100 participants. The model performed suitably against other conventional methods, though also came with the benefit of solving the cold-start problem. In terms of user satisfaction, it scored 3.92/5.

Baek et al. aim to provide nutritional support to individuals with chronic disease in the form of recommendation of suitable dietary alternatives. They outline how different chronic diseases come with different nutritional requirements and that this should be taken into account when aiming to improve dietary habits. Korean National Health and Nutrition Survey data is used, from which chronic disease data, personal features and biochemical and physiological features become features for clustering to identify groups from the data through hybrid clustering. Food products are recommended to each cluster in a stratified approach, although an individual's food preference is also considered when making recommendations.
Foods recommended to the user groups are also clustered via k-means in order to recommend similar products. This clustering is done on the basis of calories, macronutrients, sodium, cholesterol, saturated fat and trans-fat. The service ontology has the relations between the health data and the food data. Collaborative filtering is also used to predict universal preference (preference of society in general for the food). Thus, the current system is a hybrid model that combines these factors to ultimately provided a food recommendation on a stratified level with regards to chronic disease status and on a personal level in terms of food preference. Upon evaluation, the hybrid model performs best by allowing both health information and preference data to be integrated. The concept of using applications like this is attractive because individuals may be unsure how their dietary choices affect their chronic conditions and having constant access to recommendations on food choices is a convenient and efficient way to help this.
Furthermore, digital platforms for nutritional advice delivery pose some other advantages such as scalability, more effective behavior change and, in the future, reduced costs. The model could be improved by also incorporating other components of food that affect health, such as fiber, micronutrients, vitamins, etc. Although this would complicate the model, such components can have profound impacts on health and failing to capture this may mean their intake is neglected in users of the service, leading to other health issues.

Nutritional Management of Chronic Disease presents two recommendation systems that aim to provide nutritional support for those with chronic diseases.The development  of AI is allowing specialized and personalized information to be delivered to individuals or disease groups at all times and represents a promising avenue for chronic disease management.

Cancer - Prediction of colorecta cancer

Generally speaking, cancer would be considered in the domain of health and not nutrition. However, certain cancer types have a link to nutrition and of these the link between nutrition and colorectal cancer (CRC) is particularly strong. Knowing this, Shiao et al. set out to investigate how diet, genes, the interaction between the two, and other factors could be used to predict CRC in 53 multi-ethnic CRC patients and 53 paired family members. The genes were specific to folate metabolism due to a pre-existing link between this and CRC occurrence. After collecting demographic data, information on dietary intake, anthropometric data, and total number of gene plymorphism mutation in the five genes assessed, the most influential predictors were selected. These were, in descending order of importance, age (under or over 56), gender, total polymorphisms, a total vegetable intake of 10 ounces, folate intake of 100% the recommended daily intake (RDI), a healthy eating index score (HEI) of 77, overweight BMI, 150% RDI of vitamin B12, 100% of thiamine intake, and MTHFR mutations at position 677 (MTHFR 677). Interaction profiles were also assessed, where it was found that HEI and thiamine intake, BMI status and gender, and BMI status and MTHFR 677 polymorphism. Generalized regression models were generated on these interaction factors; four individual parameters associated with these interactions (BMI overweight, thiamine, gender, overweight) and four other individual parameters (age, total polymorphisms, vegetable intake, MTHFR 677 SNP). The best performance saw an area under the curve (AUC) of 0.86 and a misclassification rate of 0.21 using generalized regression with Elastic Net LOO cross-validation as an evaluation approach.

HEI score, folate intake, vegetable intake, thiamine intake and vitamin B12 intake are observed as modifiable risk factors for CRC. Although the identification of these itself is not in the realm of PN, it can give those with a family history of CRC specific dietary intake targets beyond generic advice to “eat healthy”. Since total gene polymorphisms in the genes of the pathway investigated here already enhances CRC risk, eating in a way to reduce the risk of these other, modifiable risk factors can minimize this risk further. This is similar for BMI and MTHFR 677. Whilst BMI should be appropriately managed for many health reasons, advice can be provided specifically to those harboring the MTHFR 677 because there is an interaction between the two risk factors, as seen here. These results are interesting because it shows that PN can have applications beyond classic domains of nutrition and has the potential to show promise elsewhere, in this case cancer. The genes analyzed were chosen based on the results of existing research, but it could also be that this selection is expanded in the future as more correlations between genetic variants and dietary intake are found, providing further specialized advice.

In summary, the only paper identified with a PN theme in the domain of Cancer showed that total genetic polymorphisms of the folate metabolism pathway and modifiable dietary factors are predictors of CRC. Individuals known to harbor such SNPs in these genes can look to adjust their diet based on the dietary factors identified by Shiao et al. in order to prevent additional CRC development risk.

General discussion

The current review represents the first study to systematically review the literature of applications of ML in research areas related to PN. Sixty papers were identified across four extensive databases using search terms designed to be as comprehensive as possible to obtain research relevant to PN. Furthermore, a quality assessment scheme ensured the papers were of a given standard. Both the disciplines of ML and PN are relatively new, as is highlighted by the fact that none of the papers found in the final literature dated to before 2014. It is highly likely that in the coming years the numbers of papers utilizing ML in PN will greatly increase, which is why providing a summary of the current state of the literature as presented here can be helpful for researchers in developing the field of PN. Both PN and ML are complex and have many individual considerations. Hence, a reference that provides all of this information available in one place makes this process less troublesome. The current review considers not only ML application in the final stage of PN (i.e., the generation of a nutrition recommendation outcome) but also in the data collection stages for various elements of PN. A model is only as good as the data it uses for input, and so utilizing ML to enhance the accuracy of data collection will consequently lead to improvements in PN model accuracy. In many cases, this was available for research, if not publicly available. Furthermore, it is possible that many of the papers that do not mention the status of their data could be available from the author upon request. The ability to access data in this way allows researchers to develop their own models on the same data, which means better models can be generated. All these points highlight the strength of the current review.

From the findings, some observations can be made. Despite a total of 60 papers being present in the final literature, only seven domains of nutrition and health were present, showing that the use of ML in PN is currently being concentrated in a small number of nutrition and health areas. In the case of obesity and metabolic health, there is clear motivation to invest more time and resources in solving these crises, given their prominence across the world. Personalized approaches look promising to reducing the burden of these conditions.
However, PN also demonstrated application in some other domains such as that of cancer and in the prevention of orofacial cleft development. This can be taken as a sign to suggest that PN may have a broad application. Indeed, as the fields of nutrigenomics, metabolomics, the microbiome, and PN in general develop further, situations where PN can be applied will become more apparent. Research areas known to have a nutritional link should consider combining ML and PN for treatment, prevention, or maintenance of optimal health.

It should be noted that papers representing prominent research in the field of PN usually utilise multiple features of groups of features. Whilst this not a requisite for PN models, it is certainly in line with the idea that multifactorial diseases such as obesity, diabetes, and cancer will probably not be solved with PN approaches without the use of a lot of data across multiple PN elements. This calls for the need for adequate technology and data processing for effective PN. No papers were identified that included metabolomics in their approach. Metabolomics is concerned with the identification of small molecules in a sample. Whilst the human metabolome is still being characterised, estimates of size are in the degree of tens of thousands, incorporating molecules of many different types. For this reason, papers that used measurements of small molecules as features in models were only considered from a metabolomics perspective if they explicitly stated they took a metabolomics approach or used the sophisticated analysis techniques that is seen in metabolomics research. Otherwise, these features were grouped as “Clinical Biochemical Data”. This only occurred on two occasions, and in both studies they authors use similar group names for such features. One reason no papers were found utilizing metabolomics could be that this discipline uses other methods of analysis, rather than ML. It can also be that advances in metabolomics that make it suitable for application in PN have occurred relatively recently. Only recently have attempts been made to categorize reference values for components of the human metabolome, and a recent paper used metabolite profiles to characterise interindividual response to diet, showing that metabolomics is more and more being incorporated into PN.

There is a fairly clear separation between ML and DL use in the final literature; that is, DL is largely used for imaging for dietary intake assessment, and if ML is used here then DL shows superior performance. Conversely, shallow learning is preferred in other domains. The reason for this is that DL techniques  show particularly good performance in computer vision. In order to perform so well, however, they require lots of data and computational power, making them unsuitable in circumstances without these prerequisites. Despite this, if these requirements are met, they can be expected to perform better than shallow learning techniques, as facilitated by the complexity of their learning architecture. As data increases in abundance and computational power increases whilst its price decreases, DL approaches may be employed more so and in other domains of PN.

Conclusion

The current work used a robust search methodology to review the literature on research related to PN that uses ML. To the best of our knowledge, this is the first systematic literature review to do so. Nine research questions were designed to facilitate the extraction as much relevant information as possible to provide an overview of ML in the field of PN. This included PN-orientated questions, such as the domain of the work and the specific problems tackled; ML-orientated questions, such as ML types, tasks, algorithms, features, evaluation, and data availability status; and a combination of the two, as specific problems of PN were linked to ML tasks and algorithms. We offer a contribution to the literature by summarising this information, providing a reference for future PN work looking to utilise ML to go by. To progress the field of PN further, researchers should consider other areas of health known to have a relationship with nutrition. PN and ML in such areas may allow progress due to the fusion of two promising and powerful avenues in disease prevention and treatment. Future work may also benefit from developing systems to integrate various information in PN approaches to deliver to the general population or patients, as was done in a few cases. Whilst currently PN research is happening in controlled experimental conditions, PN will play a role in every life of the general population. Having interfaces such as smartphone apps that can allow user interaction and regular dietary support or delivery of nutritional advice will be more convenient than having to discuss such matters during appointments with experts, such as doctors or nutritionists.
Finally, research to investigate the actual efficacy of PN, the ability of PN to alter behavior, and cost-benefit analyses is required before full confidence that PN can solve the problems that nutrition and health currently faces can be achieved.

NOTE: it is part of the article. Full text, tables, figures and complete bibliography in the original cited article

References (127)

1. World Health Organisation - Fact Sheets (2020)
   https://www.who.int/news-room/fact-sheets Google Scholar
2. E. Cecil, K.L. Barton.Inter-individual differences in the nutrition response: from research to recommendations. Proc. Nutr. Soc., 79 (2) (2020), pp. 171-173, 10.1017/S0029665119001198
3. D. Gardner, J.F. Trepanowski, L.C.D. Gobbo, M.E. Hauser, J. Rigdon, J.P.A. Ioannidis, M. Desai, A.C. King. Effect of low-fat VS low-carbohydrate diet on 12-month weight loss in overweight adults and the association with genotype pattern or insulin secretion the DIETFITS randomized clinical trial JAMA - Journal of the American Medical Association, 319 (7) (2018), 667 679,  10.1001/jama.2018.0245
4. E. Berry, A.M. Valdes, D.A. Drew, F. Asnicar, M. Mazidi, J. Wolf, J. Capdevila, G. Hadjigeorgiou, RDavies, H. AlKhatib,S. Ganesh, E. Bakker et al. Human postprandial responses to food and potential for precision nutrition.Nat. Med., 26 (6) (2020), pp. 964-973, 10.1038/s41591-020-0934-0
5. Zeevi, T. Korem, N. Zmora, D. Israeli, D. Rothschild, A. Weinberger, O. BenYacov, D. Lador, T. Avnit-Sagi, M. Lotan et al. Personalized nutrition by prediction of glycemic responses. Cell, 163 (5) (2015),  pp. 1079-1094, 10.1016/j.cell.2015.11.001
6. Grillo, L. Salvi, P. Coruzzi, P. Salvi, G. Parati. Sodium intake and hypertension. Nutrients, 11 (9) (2019), 10.3390/nu11091970
7. C. Cornelis, A. El-Sohemy, H. Campos.Genetic polymorphism of the adenosine A2A receptor is associated with habitual caffeine consumption. Am. J. Clin. Nutr., 86 (1) (2007), pp. 240-244, 10.1093/ajcn/86.1.240
8. Galmés, F. Serra, A. Palou. Vitamin E metabolic effects and genetic variants: a challenge for precision nutrition in obesity and associated disturbances. Nutrients, 10 (12) (2018), 10.3390/nu10121919
9. Bashiardes, A. Godneva, E. Elinav, E. Segal. Towards utilization of the human genome and microbiome for personalized nutrition. Curr. Opin. Biotechnol., 51 (2018), pp. 57-63,  10.1016/j.copbio.2017.11.013
10. S. Guest, J. Horne, S.M. Vanderhout, A. El-Sohemy. Sport nutrigenomics: personalized nutrition for athletic performance. Frontiers in Nutrition, 6 (8) (2019), 10.3389/fnut.2019.00008
11. I. Aruoma, S. HausmanCohen, J. Pizano, M.A. Schmidt, D.M. Minich, Y. Joffe et al. Personalized nutrition: translating the science of NutriGenomics into practice: proceedings from the 2018 American college of nutrition meeting. J. Am. Coll. Nutr., 38 (4) (2019), pp. 287-301 , 
12. CelisMorales, K.M. Livingstone, C.F.M. Marsaux, A.L. Macready, R. Fallaize et al. Effect of personalized nutrition on health-related behaviour change: evidence from the Food4Me European randomized controlled trial. Int. J. Epidemiol., 46 (2) (2017), pp. 578-588, 10.1093/ije/dyw186
13. Celis-Morales, J. Lara, J.C. Mathers. Personalising nutritional guidance for more effective behaviour change. Proc. Nutr. Soc., 74 (2) (2015), pp. 130-138, 10.1017/S0029665114001633
14. Mendes-Soares, T. Raveh-Sadka, S. Azulay, Y. BenShlomo, Y. Cohen, T. Ofek, J. Stevens, D. Bachrach et al.Model of personalized postprandial glycemic response to food developed for an Israeli cohort predicts responses in Midwestern American individuals. Am. J. Clin. Nutr., 110 (1) (2019), pp. 63-75, 10.1093/ajcn/nqz028
15. Shiao, J. Grayson, A. Lie, C. Yu. Personalized nutrition—genes, diet, and related interactive parameters as predictors of cancer in multiethnic colorectal cancer families. Nutrients, 10 (6) (2018), p. 795, 10.3390/nu10060795
16. L. Bush, J.B. Blumberg, A. El-Sohemy, D.M. Minich, J.M. Ordovás, D.G. Reed, V.A.Y. Behm. Toward the definition of personalized nutrition: a proposal by the American nutrition association. J. Am. Coll. Nutr., 39 (1) (2020), pp. 5-15, 10.1080/07315724.2019.1685332
17. H. Zeisel. Precision (personalized) nutrition: understanding metabolic heterogeneity. Ann. Rev. Food Sci. Technol, 11 (1) (2020), pp. 71-92, 10.1146/annurev-food-032519-051736
18. M. Ordovas, L.R. Ferguson, E.S. Tai, J.C. Mathers. Personalised Nutrition and Health. BMJ (Online)(2018), p. 361, 10.1136/bmj.k2173
19. de Toro-Martín, B.J. Arsenault, J.P. Després, M.C. Vohl. Precision nutrition: a review of personalized nutritional approaches for the prevention and management of metabolic syndrome. Nutrients, 9 (8) (2017), p. 913, 10.3390/nu9080913
20. Goni, M. Cuervo, F.I. Milagro, J.A. Martínez. Future perspectives of personalized weight loss interventions based on nutrigenetic, epigenetic, and metagenomic data J. Nutr., 146 (4) (2016), pp. 905S-912S, 10.3945/jn.115.218354
21. Mattar, D.F. de C. Mazo, F.J. Carrilho. Lactose intolerance: diagnosis, genetic, and clinical factors. Clin. Exp. Gastroenterol., 5 (1) (2012), pp. 113-121, 10.2147/CEG.S32368
22. Perez-Martinez, A. Garcia-Rios, J. Delgado-Lista, F. Perez-Jimenez, J. Lopez-Miranda. Metabolic syndrome: evidences for a personalized nutrition. Mol. Nutr. Food Res., 56 (1) (2012), pp. 67-76, 10.1002/mnfr.201100531
23. Bordoni, R. Gabbianelli. Primers on nutrigenetics and nutri(epi)genomics: origins and development of precision nutrition. Biochimie, 160 (2019), pp. 156-171, 10.1016/j.biochi.2019.03.006
24. Chaudhary, V. Kumar, P. Sangwan, N.C. Pant, A. Saxena, S. Joshi, A.N. Yadav. Personalized nutrition and –omics. Reference Module in Food Science, Elsevier (2020), p. 495, 10.1016/b978-0-08 -100596-5.22880-1
25. Kim, S. Lin, G. Ferrara, J. Hua, E. Seto. Identifying people based on machine learning classification of foods consumed in order to offer tailored healthier food options. Adv. Intell. Sys. Comput., 1131 (2020), pp. 190-194, 10.1007/978-3-030-39512-4_30
26. Ramyaa, O. Hosseini, G.P. Krishnan, S. Krishnan. Phenotyping women based on dietary macronutrients, physical activity, and body weight using machine learning tools
27. Zellerbach, C. Ruiz. Machine learning to predict overeating from macronutrient composition. Proceedings - 2019 IEEE International Conference on Bioinformatics and Biomedicine, BIBM (2019), pp. 1268-1270, 10.1109/BIBM47256.2019.8983166
28. M. Colmenar, S.M. Winkler, G. Kronberger, E. Maqueda, M. Botella, J.I. Hidalgo. Predicting glycemia in diabetic patients by evolutionary computation and continuous glucose monitoring. GECCO 2016 Companion - Proceedings of the 2016 Genetic and Evolutionary Computation Conference (2016), pp. 1393-1400, 10.1145/2908961.2931734
29. Paoli, G. Tinsley, A. Bianco, T. Moro. The influence of meal frequency and timing on health in humans: the role of fasting. Nutrients, 11 (4) (2019), 10.3390/nu11040719
30. Brignardello, E. Holmes, I. Garcia-Perez. Metabolic phenotyping of diet and dietary intake. Advances in Food and Nutrition Research, vol. 81, Academic Press Inc (2017), pp. 231-270  10.1016/bs.afnr.2016.12.002
31. A. Kolodziejczyk, D. Zheng, E. Elinav. Diet–microbiota interactions and personalized nutrition. Nat. Rev. Microbiol., 17 (12) (2019), pp. 742-753, 10.1038/s41579-019-0256-8
32. Rowe. An introduction to machine learning for clinicians. Acad. Med., 94 (10) (2019), pp. 1433-1436, 10.1097/ACM.0000000000002792
33. I. Webb, J. Fürnkranz, J. Fürnkranz, J. Fürnkranz, G. Hinton, C. Sammut, J. Sander, M. Vlachos, Y.W. Teh et al Decision tree. Encyclopedia of Machine Learning, Springer US (2011), pp. 263-267, 10.1007/978-0-387-30164-8_204
34. Cristianini, E. Ricci. Support vector machines. Encyclopedia of Algorithms, Springer US (2008), pp. 928-932, 10.1007/978-0-387-30162-4_415
35. Mannor, X. Jin, J. Han, X. Jin, J. Han, X. Jin, J. Han, X. Zhang. K-means clustering. Encyclopedia of Machine Learning, Springer US (2011), pp. 563-564, 10.1007/978-0-387-30164-8_425
36. Murtagh. Multilayer perceptrons for classification and regression. Neurocomputing, 2 (5–6) (1991), pp. 183-197, 10.1016/0925-2312(91)90023-5
37. I. Webb, E. Keogh, R. Miikkulainen, R. Miikkulainen, M. Sebag. Naïve Bayes. Encyclopedia of Machine Learning, Springer US (2011), pp. 713-714, 10.1007/978-0-387-30164-8_576
38. I. Webb, J. Fürnkranz, J. Fürnkranz, J. Fürnkranz, G. Hinton, C. Sammut, J. Sander, M. Vlachos, Y.W. Teh et al. Deep Belief nets. Encyclopedia of Machine Learning, Springer US (2011), pp. 267-269, 10.1007/978-0-387-30164-8_208
39. Munro, H. Toivonen, G.I. Webb, W. Buntine, P. Orbanz, Y.W. Teh, P. Poupart, et al. Boltzmann machines. Encyclopedia of Machine Learning, Springer US (2011), pp. 132-136, 10.1007/978-0-387-30164-8_83
40. Sherstinsky. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. Phys. Nonlinear Phenom., 404 (2020), p. 132306, 10.1016/j.physd.2019.132306
41. O'Shea, R. Nash. An Introduction to Convolutional Neural Networks. Neural and Evolutionary Computing (2015)
42. Lecun, Y. Bengio, G. Hinton. Deep learning. Nature, 521 (7553) (2015), pp. 436-444,  10.1038/nature14539
43. Schmidhuber. Deep Learning in Neural Networks: an Overview(2014) htt p://ww w.idsia.ch/˜juergen/DeepLearning8Oct2014.texCompleteBIBTEXfile
44. Kitchenham, O. Pearl Brereton, D. Budgen, M. Turner, J. Bailey, S. Linkman. Systematic literature reviews in software engineering - a systematic literature review.Inf. Software Technol., 51 (1) (2009), pp. 7-15, 10.1016/j.infsof.2008.09.009
45. Hall, D. Perelman, A. Breschi, P. Limcaoco, R. Kellogg, T. McLaughlin, M. Snyder. Glucotypes reveal new patterns of glucose dysregulation. PLoS Biol., 16 (7) (2018), Article e2005143, 10.1371/journal.pbio.2005143
46. Korem, D. Zeevi, N. Zmora, O. Weissbrod, N. Bar, M. Lotan-Pompan, et al. Bread affects clinical parameters and induces gut microbiome-associated personal glycemic responses. Cell Metabol., 25 (6) (2017), pp. 1243-1253, 10.1016/j.cmet.2017.05.002
47. López, F. Torrent-Fontbona, R. Viñas, J.M. Fernández-Real. Single nucleotide polymorphism relevance learning with random forests for type 2 diabetes risk prediction. Artif. Intell. Med., 85 (2018), pp. 43-49, 10.1016/j.artmed.2017.09.005
48. Mendes-Soares, T. Raveh-Sadka, S. Azulay, K. Edens, Y. BenShlomo,et al. Assessment of a personalized approach to predicting postprandial glycemic responses to food among individuals without diabetes. JAMA Network Open, 2 (2) (2019),Article e188102,  10.1001/jamanetworkopen.2018.8102
49. A. Sowah, A.A. Bampoe-Addo, S.K. Armoo, F.K. Saalia, F. Gatsi, B. Sarkodie-Mensah. Design and development of diabetes management system using machine learning. Int. J. Telemed. Appl. (2020), 10.1155/2020/8870141. 2020
50. Wu, V. Tremaroli, C. Schmidt, A. Lundqvist, L.M. Olsson, M. Krämer,et al. The gut microbiota in prediabetes and diabetes: a population-based cross-sectional study. Cell Metabol., 32 (3) (2020), pp. 379-390, 10.1016/j.cmet.2020.06.011