Although most dietary recommendations today are based on population averages, it is now possible to identify population sub-groups, and even individuals, who may benefit from modified nutrition guidance. As technology continues to advance, it is becoming easier for individuals to determine when they may be at increased risk for disease or a nutritional deficiency. They can access commercially available tools that provide precise guidance relevant to their genetics, phenotype and preferences(Reference Kanter and Desrosiers1). However, personalised nutrition approaches, such as consumer nutrigenetics testing, have come under scrutiny(Reference Pavlidis, Lanara and Balasopoulou2). Such systems require robust scientific substantiation, as recently described in a set of guiding principles(Reference Adams, Anthony and Carvajal3) and outlined by the Academy of Nutrition and Dietetics for application to dietetic practice(Reference Rozga, Latulippe and Steiber4). Over the past several decades, it has become increasingly clear that a complex set of interactions exist among the genetic and environmental variables affecting individual responses to diet and lifestyle behaviours. Since the early 1990s, candidate gene studies have identified several genetic variations which have led to the understanding of the pathophysiological mechanisms and pathways that underlie complex diseases(Reference Patnala, Clements and Batra5). In the last 15 years, significant advances have been made through genome-wide association studies, with the discovery of several novel loci for complex traits(Reference Witte6). However, these studies have shown that only a small amount of individual variability can be explained by genetics alone(Reference Frazier-Wood7). More recently, a holistic view has emerged regarding the interaction of genetics and other individual factors(Reference Ordovas, Ferguson and Tai8). Such interactions can be behavioural, like the relationships between individuals and their food preferences and dietary intake, or physiological, such as genotype or microbiome. These interactions, and the underlying individual variations that drive them, play a significant role in personal health and well-being over time. However, they also challenge our existing approaches to research and complicate the application of research findings for the development of personalised nutrition guidance.
In this paper, we review four recent advances in the field of personalised nutrition that can lead to an improved understanding of individual health: (1) the influence of epigenetic changes and gene expression on cognitive function in early life, (2) the association genotype with nutrient status and corresponding metabolic responses in the context of obesity, (3) the relationship of gut microbial composition (enterotype) with the efficacy of fibre-rich dietary interventions for body weight management and (4) advanced technologies such as deep learning to improve the accuracy of dietary intake measurements. By examining these sources of inter-individual variability, and considering their possible interactions, we can better understand the nuances of individualised guidance and better define future research approaches to advance both personal and public health.
Folate during pregnancy, epigenetic changes and child health outcomes
Maternal nutrition during pregnancy is essential for optimal offspring development, reducing lifelong disease burden and optimising health throughout life(Reference McDonald, Thorne-Lyman, Karakochuk, Whitfield and Green9). In particular, folate is required for one-carbon metabolism, a network of metabolic pathways involved in nucleotide synthesis, DNA repair and numerous methylation reactions(Reference Bailey, Stover and McNulty10). In early pregnancy, there is conclusive evidence that periconceptional folic acid supplementation prevents neural tube defects(11,Reference Czeizel and Dudás12) . Maternal folate supplementation may also affect neurocognitive development in early life(Reference Caffrey, McNulty and Irwin13). This section focuses on the evidence linking folate nutrition in pregnancy, DNA methylation changes and health outcomes in the child.
Folate, DNA methylation and brain development during pregnancy
Inadequate folate intake can interfere with early brain development and function, the consequences of which can vary depending on the timing of the deficiency relative to the development of the neurological structures(Reference Roffman14). The third trimester of pregnancy until 2 years after birth is a critical period of rapid growth and development of some regions of the brain such as cortical and subcortical grey matter(Reference Gilmore, Shi and Woolson15). Therefore, the continuation of folic acid supplementation beyond the first trimester (after the recommended period for the prevention of neural tube defects) may optimise folate status for prenatal brain development(Reference Roffman14,Reference Irwin, Pentieva and Cassidy16) . As such, maternal folate insufficiency appears to influence the developing brain(Reference Georgieff, Brunette and Tran17), which may result in lasting changes in child brain function(Reference Caffrey, McNulty and Irwin13).
The effect of folate nutrition during pregnancy on health outcomes in the child is thought to involve the essential role of folate in one-carbon metabolism, which may impact neurodevelopment(Reference Bailey, Stover and McNulty10,Reference Gabbianelli and Damiani18) . Folate-related epigenetic changes, and specifically DNA methylation, have been proposed as a plausible mechanism underpinning associations between maternal folate nutrition and various offspring health outcomes(Reference Kok, Steegenga and Mckay19). Epigenetic processes, including histone modifications, RNA interference and DNA methylation, involve changes to the genome that can alter gene expression without changing the underlying DNA sequence(Reference Armstrong20). Early life development, ranging from preconception to childhood, is considered to be a critical window of the lifecycle, characterised by rapid DNA methylation changes, susceptibility to environmental factors and programming of epigenetic marks that may have lasting health effects in infants born to mothers with inadequate nutritional status(Reference Numata, Ye and Hyde21).
Observational studies have reported that folic acid supplement usage by women during pregnancy was associated with changes in DNA methylation patterns of candidate genes in cord blood(Reference Joubert, den Dekker and Felix22), with decreased methylation of LINE-1 and PEG3 and increased methylation of IGF2 in cord blood(Reference Haggarty, Hoad and Campbell23) and offspring(Reference Steegers-Theunissen, Obermann-Borst and Kremer24). However, associations between maternal folate exposure and the offspring methylome have been inconsistent(Reference James, Sajjadi and Tomar25).
The Folic Acid Supplementation in the Second and Third Trimester (FASST) trial is thus far the only randomised trial of folic acid supplementation in pregnancy examining DNA methylation and maternal folate and homocysteine responses and related effects in the newborn(Reference McNulty, McNulty and Marshall26). In an analysis of biobanked samples from this trial, changes were found in DNA methylation of LINE-1 and candidate genes related to brain development such as IGF2 and BDNF, in the newborns of mothers who received folic acid(Reference Caffrey, Irwin and McNulty27). Using an epigenome-wide (EWAS) approach, changes in newborn DNA methylation at the imprint regulator ZFP57 were also shown(Reference Irwin, Thursby and Ondičová28).
Cognitive performance in offspring related to folate nutrition in pregnancy
Several observational studies have shown positive associations between self-reported folic acid supplement use in early pregnancy and childhood neurodevelopmental outcomes such as language delay, cognitive function score and verbal and motor function(Reference Julvez, Fortuny and Mendez29–Reference Villamor, Rifas-Shiman and Gillman31). Likewise, clinical studies have reported reduced cognitive ability in the offspring of mothers with suboptimal folate status(Reference Schlotz, Jones and Phillips32,Reference Murphy, Fernandez-Ballart and Molloy33) . Over 40 years ago, Gross et al. (Reference Gross, Newberne and Reid34) showed that children born to mothers with diagnosed folate-related megaloblastic anaemia in the third trimester of pregnancy had abnormal neurodevelopment and lower intellectual abilities(Reference Gross, Newberne and Reid34). Several decades later, a longitudinal study of 256 mother–child pairs linked maternal folate deficiency in later pregnancy with reduced brain volume in the children aged 6–8 years, as measured using MRI(Reference Ars, Nijs and Marroun35). However, the evidence is inconsistent, as other observational studies have found no significant associations of folate status in pregnancy with cognitive performance(Reference Tamura, Goldenberg and Chapman36) or infant neurodevelopment(Reference Wu, Dyer and King37).
The very few randomised trials conducted to date in this area have investigated the effect of multiple micronutrient supplements, not folate supplementation alone(Reference Caffrey, McNulty and Irwin13). The FASSTT Offspring trial studied the effect of folic acid supplementation during the 2nd and 3rd trimesters on the subsequent cognitive performance of the child using validated assessment tools(Reference McNulty, Rollins and Cassidy38). The children of folic acid-treated mothers scored significantly higher than the placebo group in several cognitive domains at 3 and 7 years. When compared with nationally representative samples of British children at 7 years, test scores were significantly higher in children from folic acid-treated mothers for Verbal IQ, Performance IQ, General Language and Full-Scale IQ(Reference McNulty, Rollins and Cassidy38), suggesting a role for folate-mediated epigenetic changes in genes related to brain development(Reference Caffrey, Irwin and McNulty27,Reference Irwin, Thursby and Ondičová28) .
Whilst folate-mediated epigenetic changes in genes related to brain development and function offer a biological basis to link maternal folate with offspring cognitive effects, this area of research is still in its infancy. Further carefully designed randomised trials with validated cognitive tests and ideally incorporating objective brain imaging techniques are warranted. Such studies could shed further light on the links between folate nutrition during pregnancy and offspring health outcomes and the underpinning epigenetic mechanisms. Personalised approaches, such as accurate dietary assessment of folate intake, or minimally invasive testing for folate status, are needed to identify individuals at risk of folate insufficiency or deficiency. More research is required in the case of 10 % of the worldwide population who are homozygous for the methylenetetrahydrofolate reductase C677T polymorphism as they have impaired folate metabolism and may thus have higher dietary requirements for folate(Reference Shelnutt, Kauwell and Chapman39).
Genes affecting vitamin B12 metabolism and potential role in cardiometabolic disease
Cardiometabolic diseases such as obesity, type 2 diabetes and CVD are worldwide health problems and are now increasingly prevalent(Reference Virani, Alonso and Benjamin40). The role of vitamins in cardiometabolic disease risk has more recently been the subject of research in this area. While most of the research has concentrated on vitamin D(Reference Gouni-Berthold and Berthold41), interest in vitamin B12 as a modulator of metabolic disease risk has been increasing(Reference Wiebe, Field and Tonelli42).
Vitamin B12 deficiency and cardiometabolic diseases
The relationship between low plasma or serum vitamin B12 concentrations and cardiometabolic phenotypes could be the result of several mechanisms(Reference Surendran, Adaikalakoteswari and Saravanan43). Vitamin B12 is a co-enzyme which converts methylmalonyl-CoA to succinyl-CoA, a critical step in the metabolism of odd-chain fatty acids. If this reaction cannot occur, methylmalonyl-CoA levels elevate, inhibiting the rate-limiting enzyme of fatty acid oxidation (carnitine palmitoyltransferase), leading to lipogenesis and insulin resistance(Reference Rush, Katre and Yajnik44). On the other hand, reduced vitamin B12 concentrations in obese individuals could result from a nutrient-poor diet and increased nutrient requirements related to increased body size(Reference Pinhas-Hamiel, Doron-Panush and Reichman45,Reference MacFarlane, Greene-Finestone and Shi46) . Additionally, deficiency of vitamin B12 can impair the remethylation of homocysteine in the methionine cycle resulting in raised homocysteine levels(Reference Selhub47), which is associated with an increased risk of CVD(Reference Wald, Law and Morris48).
Although vitamin B12 deficiency is associated with a wide range of chronic diseases and conditions, the relationship between low vitamin B12 status and cardiometabolic-related traits has remained inconsistent in numerous studies(Reference Wiebe, Field and Tonelli42,Reference Rafnsson, Saravanan and Bhopal49) . Polymorphisms in the key genes coding for proteins involved in the absorption, cellular uptake and intracellular metabolism of vitamin B12 may emerge as a feasible explanation for the vitamin B12 variability observed within these studies(Reference Quadros50). Genetic studies have identified several genetic variants related to vitamin B12 status, during the last few years(Reference Surendran, Adaikalakoteswari and Saravanan43). At present, only three studies in European populations have investigated the effect of genetically instrumented vitamin B12 concentrations on cardiometabolic traits such as BMI(Reference Allin, Friedrich and Pietzner51), blood pressure(Reference Husemoen, Skaaby and Thuesen52) and cardiometabolic risk(Reference Moen, Qvigstad and Birkeland53), indicating the need to study more diverse ethnic groups.
Vitamin B12 concentrations, which vary widely among individuals, are responsive to changes in diet and are dependent on the quality and consumption of animal protein(Reference Watanabe, Yabuta and Tanioka54). Therefore, controlling diet is recommended in preventing vitamin B12 deficiency(Reference Watanabe55). As mentioned previously, deficiencies of vitamin B12 and folate during pregnancy can affect DNA methylation(Reference Yajnik and Deshmukh56), suggesting that interactions between genes and nutrients may also play a role in the development of cardiometabolic disease. Given that the genetic make-up varies from individual to individual, and there is variation in genetic heritage and food consumed worldwide, it is vital to examine the interactive effects between dietary factors and genetics on vitamin B12 concentrations and metabolic traits (nutrigenetics)(Reference Vimaleswaran57).
Gene–nutrient interactions with vitamin B12
Whilst only a few nutrigenetics studies have been carried out in lower-middle-income countries, a large-scale collaborative project called the ‘gene–nutrient interactions (GeNuIne) Collaboration’ was established(Reference Vimaleswaran57,Reference Vimaleswaran58) . One of the objectives of the overall project was to investigate the effect of gene–nutrient interactions on vitamin B12 concentrations and cardiometabolic traits using population-based studies from various ethnic groups. The first vitamin B12 pilot study of the GeNuIne Collaboration was the Genetics of Obesity and Diabetes study. This study explored the relationship of vitamin B12 status and metabolic traits in 109 healthy Sinhalese adults in Colombo, Sri Lanka(Reference Surendran, Alsulami and Lankeshwara59). Genetic risk scores (GRS) were derived using ten vitamin B12-associated SNP (B12-GRS). While there was a significant association between the B12-GRS and B12 concentrations, there was no significant impact of genetically instrumented B12 concentrations on any of the metabolic traits. However, there was a significant interaction between the B12-GRS and protein energy (%) on waist circumference, suggesting that a genetically lowered vitamin B12 concentration may have an impact on central obesity in the presence of lower dietary protein intakes.
The Minangkabau Indonesia Study is another nutrigenetic study in South East Asians conducted in Indonesia as part of the GeNuIne Collaboration(Reference Surendran, Aji and Ariyasra60). The B12-GRS was constructed based on nine vitamin B12 SNP, and a metabolic disease-related GRS (metabolic-GRS) was developed based on nine metabolic disease-related SNP. The study examined the relationship of these risk scores with vitamin B12 levels and different metabolic traits in 117 healthy Minangkabau Indonesian women. This study demonstrated the impact of genetically instrumented B12 concentrations on HbA1C levels, a marker of glycaemic control(Reference Borg, Persson and Siersma61), through the influence of dietary fibre intake.
Given the increased prevalence of type 2 diabetes among Asian Indians, the Chennai Urban Rural Epidemiology Study from South India examined the association between two commonly studied fat mass and obesity-associated gene (FTO) SNP on metabolic traits and vitamin B12 concentrations in 548 Asian Indians (GeNuIne Collaboration). The study identified significant associations between the two FTO SNP not only with a higher risk of obesity but also with a lower vitamin B12 concentration(Reference Surendran, Jayashri and Drysdale62). These results suggest that increases in BMI could potentially contribute to the adverse health effects associated with vitamin B12 deficiency. A more recent study has shown that long-term supplementation with vitamin B12 influences the regulation of several type 2 diabetes-associated genes by methylation of miRNA-coding gene, miR21, and could thus epigenetically regulate the risk of obesity, insulin resistance and type 2 diabetes(Reference Yadav, Shrestha and Lillycrop63).
While these findings are fascinating and highlight novel possibilities of gene–nutrient interactions, they need to be replicated in an independent cohort utilising a larger number of samples. Further understanding of the role of these gene–diet interactions at the molecular level is necessary before diets can be personalised according to varying ethnicity or genotype.
Importance of microbial enterotypes in personalised obesity management
As the prevalence of obesity has reached epidemic proportions globally, the search for effective management continues. Through multiple dietary intervention studies to reduce energy content, it has become evident that there is considerable weight loss variation among participants(Reference Poulsen, Due and Jordy64,Reference Hess, Benítez-Páez and Blædel65) , indicating that no ‘one diet fits all’. Accordingly, dietary weight loss success is likely dependent on several individual characteristics, including host genetics and gut microbiota(Reference Zeevi, Korem and Zmora66). While gut microbiota-modulating interventions can alleviate obesity in mice(Reference Sonnenburg and Bäckhed67), similar causal findings appear absent in human clinical trials. However, this has also revealed the enormous variability of the human gut microbiome within and across populations.
To reduce complexity, researchers have proposed stratification of individuals according to distinct gut microbiota composition types termed ‘enterotypes’(Reference Arumugam, Raes and Pelletier68). Primarily two enterotypes have consistently been found across populations; one type dominated by Prevotella species, the other by Bacteroides species(Reference Costea, Hildebrand and Arumugam69). The establishment of these types seems to occur during early childhood and to be highly influenced by long-term dietary habits(Reference Wu, Chen and Hoffmann70), with the Prevotella enterotype associated with a carbohydrate- and fibre-rich diet. Conversely, the Bacteroides enterotype is associated with a ‘Western diet’ low in fibre, and high in fats and refined sugars(Reference Vangay, Johnson and Ward71). Factors such as age, sex, cultural background and geography have been found to have little influence on the establishment of enterotypes(Reference Costea, Hildebrand and Arumugam69).
Enterotype predicts success in weight loss when consuming high-fibre diets
Following the discovery of enterotypes, epidemiological studies have found conflicting associations between enterotypes and metabolic health(Reference Pedersen, Gudmundsdottir and Nielsen72,Reference Ley73) . As the enterotypes seem to differ in their ability to degrade dietary fibre(Reference Costea, Hildebrand and Arumugam69), it is adjacent to speculate that metabolic responses of the enterotypes depend on dietary compositions(Reference Christensen, Roager and Astrup74). Recent dietary intervention studies with healthy European subjects point to a more beneficial role of a high-fibre diet for individuals with a Prevotella enterotype than individuals with the Bacteroides enterotype.
In 2015, Kovatcheva and co-workers convincingly linked enterotypes to glucose metabolism after consuming a high-fibre, barley-rich diet for 3 d(Reference Kovatcheva-Datchary, Nilsson and Akrami75). Specifically, they found that individuals with the Prevotella enterotype improved their enzymatic capacity for fibre degradation and glucose metabolism as a result of the diet, an effect not seen among Bacteroides enterotype subjects.
Since then, Hjorth and co-workers reanalysed a 26-week ad libitum study with a fibre-rich New Nordic Diet (NND) and an average Danish diet, lower in fibre (43·3 v. 28·6 g/10 MJ), and predicted dietary success based on pre-treatment faecal Prevotella:Bacteroides (P:B) ratio(Reference Hjorth, Roager and Larsen76). While the high P:B group lost weight on NND compared with the control diet, the low P:B group did not. Interestingly, during the 1-year follow-up period, subjects with the high P:B ratio who changed from the control diet to the recommended NND managed to maintain their weight loss, whereas subjects with the low P:B ratio who changed from the control diet to the NND regained weight.
To validate these findings, two other dietary intervention studies aimed at weight loss were stratified by enterotype(Reference Hjorth, Blædel and Bendtsen77,Reference Hjorth, Christensen and Kjølbæk78) . Similarly, these studies observed that subjects who consumed a high-fibre diet for 6 months lost more weight if they had a high P:B ratio v. a low P:B ratio. Furthermore, in a 6-week ad libitum study comparing whole grains with refined wheat (fibre intake: 33 g/d v. 23 g/d), a 1·8 kg weight loss was observed among participants with high Prevotella abundance, and no difference in the low Prevotella group(Reference Christensen, Vuholm and Roager79). The finding suggests that specific whole-grain fibres, such as arabinoxylans, benefit the Prevotella enterotype, likely due to a beneficial match between species enzymatic degradation capacity of especially the prevalent Prevotella species, P. copri (Reference De Filippis, Pasolli and Tett80). In comparison, the colonic functional and ecological properties seem less uniform among subjects with a Bacteroides enterotype(Reference De Paepe, Verspreet and Courtin81).
Enterotypes at lower taxonomical levels and host digestive capacity
Until recently, the majority of dietary intervention studies have been limited to genus level taxonomy by 16S rRNA gene sequencing. Although with continuous advances in the field of sequencing, species-level taxonomy is now widespread, revealing a large inter-individual variation in Bacteroides and Prevotella species(Reference Christensen, Roager and Astrup74). Moreover, at the strain level, recent pioneering microbiome studies demonstrated that P. copri is not a monotypic species, but encompasses four distinct clades with substantial functional diversity differences including the ability to degrade whole-grain fibres(Reference Tett, Huang and Asnicar82). Consequently, if weight loss is affected by the colonic bacterial enzymatic capacity, future enterotype studies need deep-level sequencing to explain potential weight loss variability further.
Another intriguing factor in obesity management is the ability of the host to degrade the major polysaccharide component of diets, starch(Reference Falchi, El-Sayed Moustafa and Takousis83). If dietary starches escape host digestion in the small intestine, it will primarily feed Prevotella and Bacteroides species located distally in the gut. Amylase secreted from the salivary glands and pancreas determines the amount of starch escaping host digestion, and genetically the salivary amylase gene (AMY-1) exhibits some of the greatest copy numbers (CN) of any human gene(Reference Elder, Ramsden and Burnett84). Stratification of participants of the 26-week NND study (discussed above) according to AMY-1 CN (i.e. low and high AMY-1 CN groups, respectively) further explained weight loss variability(Reference Hjorth, Christensen and Larsen85). Here, the majority of weight loss difference between enterotypes was observed in the low AMY-1 CN group with a strong linear relationship between P:B ratio and weight loss, which was not found for the high AMY-1 CN group. This indicates that not only dietary fibre but also the total pool of fermentable substrate reaching the colon play a role in the weight loss difference between the enterotypes.
Moving this field forward, there is furthermore a need to understand the underlying mechanisms linking microbial fibre fermentation to host metabolic alterations. To elucidate this, microbial metabolites produced in the colon that enter circulation and interact with host cells need to be investigated (e.g. SCFA)(Reference Roager and Dragsted86). In conclusion, microbial enterotypes are currently highly interesting biomarkers for explaining variability in weight loss upon intake of dietary fibre and whole-grain-rich diets(Reference Ortega-Santos and Whisner87). Soon, it might be possible to complement personalised dietary weight-loss strategies with microbial enterotype tests, thus ensuring dietary preferences of both the individual and the colonic microbes. Such a personalised approach may, in fact, support dietary adherence and long-term weight maintenance.
Advanced technologies for food image recognition in nutrient intake assessment
Dietary assessment is a crucial step in the real-world deployment of any personalised nutrition programme. The ability of an individual to track their food intake plays a role in self-monitoring as a critical aspect of behaviour change(Reference Peterson, Middleton and Nackers88). It can also provide a professional dietitian with information on how a client is adhering to their individualised meal plan. However, assessing dietary intake with traditional methods carries considerable costs and burden to the individual. Such methods are also prone to errors as they often rely on self-reporting(Reference Burrows, Ho and Rollo89).
Advanced solutions are needed to objectively quantify food and beverage intake(Reference Mezgec, Eftimov and Bucher90). Food image recognition is a promising strategy because most individuals own a smartphone with a camera, so the barrier to entry is low, and it can reach a large population. However, automatically recognising food items from images is a challenging computer vision problem due to a variety of issues: (1) foods are typically deformable objects, (2) foods can lose their visual information during preparation, (3) different foods can appear visually similar, (4) the same food can appear differently depending on the lighting or angle and (5) limited amount of visual information for beverages(Reference Mezgec and Koroušić Seljak91). Only after foods are visually recognised, can they be reliably linked to a food composition database(Reference Eftimov, Korošec and Koroušić Seljak92).
The introduction of the Pittsburgh Fast-Food Image Dataset in 2009 facilitated early research in this area based on manual recognition methods, but these approaches mostly achieved only 10–40 % classification accuracy(Reference Chen, Dhingra and Wu93,Reference Yang, Chen and Pomerleau94) . In 2014, deep learning was first used to recognise food images. Deep learning, or deep neural networks, allows computational models composed of multiple processing layers to learn relevant image features through training on a set of input images(Reference LeCun, Bengio and Hinton95,Reference Deng and Yu96) . Deep convolutional neural networks are inspired by the visual system of animals, where individual neurons assess the visual input by reacting to overlapping regions in the visual field(Reference Hubel and Wiesel97). Because they can classify each pixel of the image, they can recognise any number of items, along with their location and size, allowing for food volume and food weight estimation. This approach has achieved substantially better results than other methods, resulting in an increased focus on deep learning in recent research(Reference Zhou, Zhang and Liu98,Reference Knez and Šajn99) .
A novel deep learning architecture for food image recognition, called NutriNet, has been developed by Mezgec and Koroušić Seljak(Reference Mezgec and Koroušić Seljak91). It is a modification of the well-known AlexNet architecture(Reference Krizhevsky, Sutskever, Hinton, Pereira, Burges and Bottou100), with increased image size and an additional convolutional layer at the beginning of the neural network(Reference Mezgec and Koroušić Seljak91). NutriNet was first trained on 225 953 freely available images that were downloaded from the Internet and organised into appropriate food classes (520 unique food items). When tested against three popular deep learning architectures of the time (AlexNet(Reference Krizhevsky, Sutskever, Hinton, Pereira, Burges and Bottou100), GoogLeNet(Reference Szegedy, Wei and Yangqing101) and ResNet(Reference He, Zhang and Ren102)), NutriNet was found to be superior to AlexNet and GoogLeNet and faster to train than all three of the other architectures. The deep learning approach was then used to recognise any number of items in a single food image using a training set obtained from the ‘fake food buffet’(Reference Bucher, van der Horst and Siegrist103), which is visually similar to real food. Fully convolutional networks, introduced by Long et al.(Reference Long, Shelhamer and Darrell104), were applied to perform semantic segmentation, partitioning the image into logical parts and classifying each part on a pixel level. Due to the complexity of food images, an fully convolutional network variant that can segment images at the finest grain (fully convolutional network-8s)(Reference Long, Shelhamer and Darrell104) was used to train a model on the fake food buffet image data set. Output predictions of the trained model were compared with the ground-truth labels using the pixel accuracy measure(Reference Long, Shelhamer and Darrell104), and the final accuracy of the trained fully convolutional network-8s model was 92·18 %.
In recent years, deep learning has been validated numerous times as a suitable solution for recognising food images(Reference Zhou, Zhang and Liu98). Availability of food image data sets has been improving(Reference Ciocca, Napoletano and Schettini105–Reference Cai, Li and Li107), although there is a need for validation against data sets from different regions across the world. Future work will focus on real-world food images, which exhibit more variance compared with the test images used in the research environment. In the future, such technology could be used to improve dietary assessment in clinical trials. For example, it can play a role in human studies in the areas mentioned above, where accurate quantification of folate, vitamin B12 and energetic intake is critical. Consumer wellness apps can also apply this solution in the future, improving the dietary assessment of individuals and facilitating self-monitoring towards positive behaviour change.
Looking towards the future: considerations to facilitate individualised nutrition guidance
From the four examples discussed here, it is clear that scientific evidence about individualised responses to nutrient intakes continues to emerge. By recognising these sources of individual variation, we have the potential to improve the conduct and interpretation of clinical trials. For example, in studies of folate status and epigenetics, dietary intake assessment may be enhanced by the use of an advanced technique such as the one described here. In obesity research, understanding genes related to salivary amylase and vitamin B12 metabolism, and gut microbiota composition, may enable better prediction of responders and non-responders and dietary adherence to an intervention could be assessed by photo recognition.
Such approaches can also play a role in the development of personalised nutrition programmes in the public health arena. For maternal health, such an app could help to identify women at risk of folate deficiency, providing them with guidance on how to identify foods with adequate folate, or supplementation options that meet their individual preferences. A weight management app could integrate automated dietary intake assessment with information on genotype and microbiome composition to provide a user with more pertinent and individualised guidance.
As new information comes to light, such as that described in the four examples in this paper, confidence in individualised nutrition guidance is likely to increase. In the meantime, companies and organisations developing personalised nutrition programmes must take science-based, ethical and rigorous approaches in developing their guidance(Reference Adams, Anthony and Carvajal3,Reference Bush, Blumberg and El-Sohemy108) . In this way, personalised nutrition approaches can maintain their credibility, provide maximum benefit to the individual and advance public health.
Acknowledgements
S. M. and B. K. S. would like to thank Tamara Bucher from the University of Newcastle, Australia, for providing the fake-food image data set. A. C., H. M. and K. P. would like to acknowledge the researchers on the ‘EpiFASSTT’ and ‘EpiBrain’ projects. S. S. and K. S. V. acknowledge support from the British Council, Newton Fund, British Nutrition Foundation and the authors of the GeNuIne Collaboration(Reference Watanabe, Yabuta and Tanioka54).
S. M. and B. K. S. were supported by the European Union’s Horizon 2020 research and innovation programmes (grant numbers 863059 – FNS-Cloud, 769661 – SAAM); and the Slovenian Research Agency (grant number P2-0098). The European Union and Slovenian Research Agency had no role in the design, analysis or writing of this article. A. C., H. M. and K. P. were supported in part by the HSC Research and Development Division of the Public Health Agency, Northern Ireland (Enabling Research Award STL/5043/14), the Biology and Biological Sciences Research Council and the Economic and Social Research Council (Grant Ref: ES/N000323/1 ‘EpiFASSTT’) and the European JPI ERA-HDHL “Nutrition & the Epigenome” scheme jointly funded by the Biology and Biological Sciences Research Council and the Medical Research Council (Grant Ref: BB/S020330/1 ‘EpiBrain’). The Northern Ireland Department for Economy (DfE) funded the PhD studentship for Aoife Caffrey. The funders had no role in the design, analysis or writing of this article. S. S. and K. S. V. were supported in part by GeNuIne Collaboration: British Nutrition Foundation; Sri Lankan (Genetics of Obesity and Diabetes) study: Farnborough College of Technology, UK; Indonesian (MINANG) study: British Council Newton Fund Researcher Links Travel Grant: 2016-RLTG7-10215; and Indian (Chennai Urban Rural Epidemiology Study) study: Research Society for the Study of Diabetes in India (RSSDI) (Project No: RSSDI/HQ/Grants/2014/250).
N. V. M., A. C., L. C., S. M. and S. S. drafted the manuscript; M. H., H. M., K. P., H. R., B. K. S., K. S. V., M. R. and S. P. contributed important intellectual content and critically revised and edited the manuscript. All authors read and approved the final manuscript.
N. V. M., M. R. and S. P. are employees of DSM Nutritional Products, a provider of personalised nutrition products and services. M. F. H. and L. C. are co-inventors on a pending provisional patent application for the use of biomarkers to predict responses to weight loss diets.
Open access
