Conventional machine learning (ML): supervised and unsupervised

ML is a cornerstone of AI applications in PN, offering well established techniques for analysing complex, well-structured data (see Figure 1).

Figure 1.

AI models and concepts applied to PN applications. Abbreviations provided in Table 1.

Table 1.

Table of common abbreviation used in the field of AI and within this review

Supervised learning involves training algorithms on labelled datasets to predict specific outcomes, such as linking dietary intake patterns to metabolic responses or disease risks. Tree-based methods such as XGB or Random Forest, Support Vector Machines, Naïve Bayes, and K-Nearest Neighbors belong to this class of models. This review does not mention linear and logistic regression as they are traditional statistical models used for data analysis and modelling. Tree-based methods proved to be effective in capturing complex patterns^{(Reference Uddin and Lu17)} yet being less complex and less computationally demanding than advanced Artificial Neural Network (ANN) models for plenty of applications. These models are particularly appealing when dealing with static tabular data to identify patterns between input variables and targeted outcomes, for example, nutrient intake^{(Reference Kirk, Catal and Tekinerdogan18)}, disease prediction^{(Reference Zou, Qu and Luo19)}, or health intervention engagement^{(Reference Brankovic, Hendrie and Baird20)}.

Unsupervised learning, in contrast, identifies hidden patterns in unlabelled data. To that end, the algorithm searches for patterns present in the data based on similarities, and groups data that share similar characteristics. This approach has been used in metabolic phenotyping^{(Reference Kirk, Catal and Tekinerdogan21)}, to identify profiles or biomarker signatures of health and disease risks^{(Reference Shomorony, Cirulli and Huang22)}, or clustering microbiome samples into groups that share similar microbial compositional patterns^{(Reference Subedi, Neish and Bak23)}.

Deep learning

DL is a subset of ML that leverages Artificial Neural Networks (ANN) models – so called Deep Neural Networks (DNN) – to analyse and process data. With the ability to automatically extract and learn features from raw data, deal with complex, high-dimensional, or unstructured data, ANN have great capabilities to identify complex relationships and patterns between inputs data (e.g. genomic, microbiome, medical, lifestyle data, or any other) and targeted outcomes such as nutrient deficiencies, food intolerances, engagement interventions etc. Depending on the data format and application, some model types may be more appropriate than others. Unlike ML, it requires large datasets to train on and better computational resources. Thus, if data is limited traditional ML models are recommended over ANN.

Convolutional Neural Networks (CNN) is a type of DNN which are particularly effective for image analysis and as such have been extensively used for image-based dietary assessments such as identifying food items from photographs^{(Reference Uddin and Lu17,Reference Brankovic, Hendrie and Baird20,Reference Kirk, Catal and Tekinerdogan21, Reference Tahir and Loo24)}. It performs convolution on input data using filters to produce feature maps and extract positional relationships in the image data. CNN can be used to automate food recognition tasks, enhancing the accuracy of dietary intake monitoring.

Recurrent Neural Networks (RNN) is a type of DNN specialised for sequence modelling. It is designed to capture historical information in time series data and is capable of learning order dependence in sequence prediction problems. For instance, to understand the impact of food intake in predicting BMI^{(Reference Cunha, Duarte and Oliveira25)} or to predict the risk of chronic diseases based on dietary habits to inform early warnings for at-risk individuals. Long short-term memory (LSTM) is a type of advanced RNN designed to overcome the so-called vanishing gradient problem, the problem when the learning signals (called gradients) become very small as they move through the model. This problem makes it difficult for traditional RNN to learn and retain information from earlier steps in long sequences. LSTM networks use a specialised structure to preserve important information over longer time periods, improving their ability to handle long-term dependencies.

Reinforcement Learning (RL) is suitable for complex problems in dynamic environments for which an optimal solution is unknown, and labelled data is unavailable. In RL, ‘learner’ which refers to the AI system or model can act autonomously to reach its goals without the need for constant human guidance, and as such it is called an agent. The agent makes choices and does things in the environment to see what works best. Each choice can earn a reward, or a penalty based on a user-defined set of rules. The algorithm then adjusts its future choices to get more rewards and fewer penalties. This allows it to master a task without being told exactly how to do every step.

RL is emerging in PN by enabling systems to adapt dietary recommendations over time based on feedback and evolving health data. These advanced DL techniques allow for dynamic and personalised dietary modelling as well as an opportunity to consider different data as an input – multi modal systems – to capture interactions between diet, genetics, and lifestyle. It is suitable for analysis of big data from different sources to generate adaptive models over time for personalised health recommendations^{(Reference Mulani, Heda, Tumdi, Agarwal, Acharya, Mittal, Abraham and Kelemen26–Reference Forman, Kerrigan and Butryn28)}.

Transformers are a specific type of DL model originally developed for text, that is, natural language processing (NLP) but has successfully expanded to other data and applications. It is based on self-attention mechanisms^{(Reference Vaswani, Shazeer and Parmar29)}, which enable the model to consider the entire context of the input to better understand its meaning. Transformers are suitable for complex and unstructured data of any type - text, images, time series, tabular data or their combination) – so their application is widespread.

Natural language processing

NLP and Large Language Models (LLM), as the most advanced technology in NLP, are instrumental in extracting insights from textual data such as those related to dietary behaviours. NLP algorithms are useful to analyse food diaries, survey responses, and social media content to identify eating habits, preferences, and patterns in words. These insights can be used to customise dietary interventions and understand broader trends in dietary behaviour. NLP also supports sentiment analysis, helping to capture emotional responses to food and dietary changes, or interventions, which is crucial for designing user-friendly and sustainable nutrition programmes^{(Reference Hassanien and Soliman13)}. Advanced Large Language Models (LLMs) are mostly built on Transformer ANN architecture due to their ability to analyse the entire context at once. For example, ChatGPT is powered by an LLM in combination with transformers and RL from human feedback. It and similar technologies could be used for PN advice by creating meal plans^{(Reference Papastratis, Konstantinidis and Daras30)}.

Explainable AI (XAI)

Advanced DNN models are quite complex in structure, and therefore often referred to as a ‘black box’ because its internal decision-making processes are not interpretable or transparent to humans. As AI systems become more complex and influential in providing dietary recommendations, there is a growing need for more transparency and interpretability in their decision-making processes. Also, having limited input from nutritional experts to guide the system to generate safe and accurate recommendations has raised questions about the trustworthiness of the nutritional advice provided by such AI systems^{(Reference Papastratis, Konstantinidis and Daras30)}.

It has been argued that XAI could build greater trust with the healthcare workforce by providing transparency into the AI decision-making process^{(Reference Ghassemi, Oakden-Rayner and Beam31)}. This would provide insight into contributors that drive an algorithm’s decision-making process to generate an output, for example, dietary advice. A greater understanding of key contributing factors in an algorithm’s decision process may boost trust in AI-generated outcomes. For instance, XAI can elucidate why a specific nutrient is recommended based on a user’s genetic data or explain the relationship between microbiome composition and suggested dietary changes. However, recent studies report the dangers of overreliance and potentially being misled as a result of XAI^{(Reference Cabitza, Fregosi, Campagner, Longo, Lapuschkin and Seifert32,Reference Brankovic, Cook and Rahman33)} , especially if variables are correlated and/or if deployed XAI does not capture casual relationships. Instead, using interpretable models, which are models that trace and comprehend the reasoning mechanisms behind the outcomes, has been recommended for high-stakes decisions^{(Reference Rudin34)}. However, interpretable models may be less accurate in handling complex, highly non-linear and/or high-dimensional data, which is a limitation of using these models. In multidisciplinary literature, explainable and interpretable are often used interchangeably. However, one should be aware they have completely different meanings as explained above and should be used appropriately.

The advantage and potential of XAI is its ability to help researchers and practitioners uncover potential biases or inaccuracies in AI models, ensuring that recommendations are both fair and scientifically reliable^{(Reference Straková27,Reference Vaswani, Shazeer and Parmar29)} . SHapley Additive exPlanations (SHAP)^{(Reference Lundberg and Lee35)} and Local Interpretable Model-agnostic Explanations (LIME)^{(Reference Caruana, Lou and Gehrke36)} are the most popular explanation methods applicable to any ML model, and LLM. They show which parts of the input (like words in a sentence or features in data) has the biggest impact on the model’s prediction. SHAP does this by fairly assigning credit to each factor that influences a prediction, similar to how rewards might be shared among team members based on their contributions. LIME, on the other hand, simplifies the model’s decision just around the example being examined, helping us understand why the model made that specific choice.

There is more literature using SHAP and in nutrition research it is used for explaining predictions and identifying relative importance of nutrition-related factors for different outcomes including mortalities^{(Reference Park, Baik and Lee37,Reference Martin-Morales, Yamamoto and Inoue38)} , cardiovascular diseases^{(Reference Turnbull, Nghiep and Butsorn39,Reference Wang, Kang and Chen40)} , depression^{(Reference Hosseinzadeh Kasani, Lee and Park41,Reference Huang and Huang42)} , feeding intolerances in preterm newborns^{(Reference Xu, Peng and Peng43)}, disengagement in online weight loss applications^{(Reference Brankovic, Hendrie and Baird20)}, overweight and obesity^{(Reference Sun, Yuan and Farrahi44)}, glucose monitoring^{(Reference Zignoli, Skroce and Lipman45,Reference Cao and Hu46)} .

Unlike SHAP and LIME, which are model-agnostic, that is, methods that can be used with any kind of AI model, some XAI algorithms are specifically designed to explain only neural network-based models. Their taxonomy has been reviewed by Ibrahim et al.^{(Reference Ibrahim and Shafiq47)} and Samek et al.^{(Reference Samek, Montavon and Lapuschkin48)}. However, there is lack of evaluation of XAI methodologies^{(Reference Caterson, Lewin and Williamson49)}.

Role of Data in AI Models

AI’s potential to revolutionise PN lies in its ability to integrate diverse and complex datasets (Figure 2). In the following, the types of data that could be part of a PN framework^{(Reference Özdemir and Kolker50)} and the information they contain are described.

Figure 2.

Data for AI-powered holistic PN.

Medical Data allows nutrition recommendations to be tailored to an individual’s unique health profile. Dietary recommendations could be aligned with an individual’s health status, their risk factors and health needs to prevent or manage chronic diseases through more specific dietary strategies.

Genomic data provides insights into individual genetic predispositions, such as sensitivities and intolerances to specific nutrients or responses to treatment options.

Metabolomic data captures real-time snapshots of metabolic activity, offering a detailed understanding of an individual’s physiological responses to dietary intake.

Microbiome data sheds light on the gut’s role in health and nutrition, highlighting how gut bacteria interact with diet to influence overall well-being.

Lifestyle data, including physical activity, sleep habits, and stress patterns, meal timing and composition can be used to personalise recommendations within the context of their broader lifestyle routine.

Other data, including demographic information and other relevant health and wellbeing data that can be collected via variable devices, mobile and online application may also be informative in modelling complex process and interactions.

By synthesising these data, AI models generate holistic, data-driven insights that guide more precise and PN strategies. This integration is essential for understanding the dynamic interplay of factors influencing dietary behaviour and health, paving the way for innovative and effective interventions in PN.

Dietary assessment and monitoring

Along with recommendation systems, this is the most mature application of AI in PN. Currently there are at least 11 AI-enabled food image recognition apps available from Australia’s Apple App and Google Play^{(Reference Li, Yin and Choi51)}. They are transforming conventional dietary assessment and monitoring.

Traditional methods of dietary assessment are time consuming, requiring completion of food diaries or surveys. AI-driven solutions, integrated into mobile applications, offer more efficient, real-time alternatives for food recognition and tracking of food intake. The fundamental operating principles of AI-based food segmentation and classification systems for supporting nutrition monitoring are detailed in the work of Freitas et al.^{(Reference Freitas, Cordeiro and Macario52)}.

Segmentation performed with CNN is followed by estimation of energy and nutrient intake by combining CNN and NLP^{(Reference Ruede, Heusser and Frank53)}. Tahir and Loo provide comprehensive summary of AI methodologies leveraged for automatic image-based food recognition and volume estimation methods for dietary assessment^{(Reference Tahir and Loo24)}. The authors identified several limitations, including the lack of comprehensive datasets for benchmarking and performance evaluation, which they regard as one of the biggest challenges. Additionally, they highlighted the issue of so-called open-ended learning - the inability of models to continuously learn and adapt to new information over time, without forgetting previously learned information. Furthermore, challenges remain in the accurate classification of ingredients in prepared and mixed food items.

Mobile apps integrated with AI algorithms enable users to scan food items, automatically recognise ingredients, and calculate nutritional values. Image analysis, another promising tool, utilises computer vision to identify food items from photographs, helping users track meals without manual input. Current accuracy is relatively high in identifying pre-categorised foods (87 % or more), however further improvements especially related to diversity of foods, mixed and culturally varied dishes^{(Reference Li, Yin and Choi51)} would support more widespread use of such tools in PN.

Initial steps towards fully automated food monitoring have been made by leveraging advanced AI models. LSTM-based system for food consumption identification to improve diabetes management, utilising dynamic temporal data representing full-day record of wrist movement, has demonstrated great performance^{(Reference Dénes-Fazakas, Simon and Hartvég54)}. This approach enables automatic meal detection and reduces reliance on patient’s input, however, has constraints in relation to diabetes management due to the artificial pancreas systems – delayed insulin action and the need for pre-meal administration.

Lastly, engagement with digital tools is key to achieving successful behaviour change and improvements in digital health interventions. Although technology is increasingly utilised in health interventions, a notable challenge continues to be the decline in user engagement and non-use attrition^{(Reference Eysenbach55)}. ML-algorithms have been shown to be effective in predicting and explaining factors leading to disengagement, paving the way for improved and more supported delivery of personalised weight loss programmes or other personalised digital interventions reliant on engagement^{(Reference Brankovic, Hendrie and Baird20)}.

Personalised diet recommendations

Leveraging AI, in particular DNN models and large datasets, to generate tailored meal plans based on individual profiles has significantly increased^{(Reference Tsolakidis, Gymnopoulos and Dimitropoulos56)}. Different data sources are exploited and combined with traditional surveys and questionnaires. These include data collected via wearable devices such as watches and activity trackers, demographics and personal data, preferences and motivation, health and clinical data (e.g. insulin levels), microbiota and genetic data. Physiological data is used most while use of genetics and microbiota data is less common. Data can be analysed with the range of AI technologies discussed in Section 2. However, an approach integrating all data relevant to generate personalised advice is still unavailable. Tsolakidis et al.^{(Reference Tsolakidis, Gymnopoulos and Dimitropoulos56)} provided an overview of advancements in data-driven technologies for personalised nutrition, focusing on data collection methods employed in modern -ML technologies and their challenges.

Prediction models for health outcomes

AI predictive models, in particular those suitable for modelling time sequences, are promising for the prevention of chronic disease such as CVD, diabetes, and obesity due to their ability to analyse complex temporal data, capability to capture temporal relationships and identify correlations and patterns. AI-powered prediction models can estimate the likelihood of developing specific health conditions, enabling early intervention strategies and personalised dietary recommendations.

For instance, a person with pre-diabetes can receive AI-driven dietary recommendations obtained based on their health, dietary, lifestyle and genetic data which could help to avoid progression to Type 2 diabetes. Another example is leveraging AI models to predict how certain foods or nutrients may interact with a person’s unique microbiome or identifying genetic predispositions to nutrient deficiencies or sensitivities that would allow for personalised dietary interventions to optimise gut health. Some initial steps have been made for the later example with two studies developing recommender system based on individuals’ genetic data^{(Reference Chen, Toumazou and Galanakis57,Reference Chen, Karvela and Sohbati58)} .

Few studies have deployed complex models or time series data in PN^{(Reference Kirk, Catal and Tekinerdogan21,Reference Theodore Armand, Nfor and Kim59)} , and no studies have explored multimodal predictive models for chronic disease or other health outcomes directly related to PN. Combining diverse data sources, for example, clinical data, nutrition intake, health data, genetic or microbiome, has potential to improve the precision and effectiveness of health predictions with AI technologies but there is limited evidence of its application in PN at present.

(i) Scalability and accuracy

The era of ‘big data’ means that information can come from multiple sources such as wearable devices, mobile apps, websites, online journals, electronic medical reports, wearable devices. AI systems have capability to process vast amounts of complex data rapidly and precisely^{(Reference Bajaj and Lata60)}, enabling efficient analysis of large datasets and provision of real-time feedback. This capability allows for scalable nutrition solutions that can reach a wide population^{(Reference Zheng, Wang and Shen61)}. AI-driven tools, such as mobile applications and wearable devices equipped with conventional and DL algorithms, can provide dietary recommendations, enable real-time accurate nutrient estimation, monitor intake in real-time and personalised nutrition plans, making personalised nutrition accessible to users across different regions and socioeconomic backgrounds^{(Reference Abdallah, Godwins and Ijiga62)}. Leveraging ML enables predictive analytics for early intervention and better support^{(Reference Brankovic, Hendrie and Baird20)}. A scoping review^{(Reference Zheng, Wang and Shen63)} has outlined the advantages of AI in improving accuracy of dietary assessment using a DL. ML and DL solutions are already part of various commercial programmes and apps for weight loss.

(ii) Personalisation

One of the benefits of integrating AI into PN research is its ability to enhance precision in identifying individual dietary needs. Unlike traditional methods that rely on generalised dietary guidelines^{(Reference Ge, Sadeghirad and Ball1)}, AI algorithms can analyse complex datasets to account for individual variation related to phenotype, genotype, lifestyle behaviour (diet, activity, etc.), goals, and preferences^{(Reference Adams, Anthony and Carvajal64)} to create tailored nutritional strategies. The ability of AI to harness large, multidimensional datasets provides an opportunity to improve the accuracy of identifying individual dietary needs and tailoring intervention accordingly.

ML models excel at recognising subtle patterns in these data, enabling the identification of specific dietary components that align with an individual’s unique health requirements and goals. For instance, AI can predict how a person might be engaged during a weight loss programme^{(Reference Brankovic, Hendrie and Baird20)}, respond to certain foods based on their genetic predispositions or gut microbiome^{(Reference Miyazawa, Hiratsuka and Toda8,Reference Zeevi, Korem and Zmora65–Reference Carpio, Sanz and Gozalbes67)} , reducing the trial-and-error approach often associated with dietary guidance. By leveraging this level of precision, AI ensures that dietary recommendations are more accurate and possibly also more effective in preventing disease, optimising health outcomes, and improving overall quality of life.

(iii) Dynamic and adaptiveness

Improved patient-provider communication can be made through the real-time monitoring and evaluation of patient progress via apps^{(Reference Chen, Gemming and Hanning68)}. One of the key advantages of AI in PN is its ability to generate recommendations based on longitudinal data. Unlike traditional periodic dietary assessments that rely on self-reporting, AI models offer real-time monitoring capabilities^{(Reference Zheng, Wang and Shen61)}, providing timely feedback and intervention. Adaptive nature allows for dynamic guidance^{(Reference Brankovic, Hendrie and Baird20)}, where recommendations are continuously refined as users log their behaviours such as their meals and snacks, exercise routines, and health metrics. Such real-time adjustment ensures that dietary advice remains relevant and effective, adapting to changes in the user’s health status, activity levels, or even seasonal variations in food availability. By leveraging AI’s capabilities in data analysis, real-time feedback, and adaptive learning, PN strategies can be more effectively implemented, potentially leading to better adherence to dietary guidelines and improved health outcomes. Longitudinal studies powered by AI provide an opportunity to reveal long-term effects of certain dietary patterns for instance the risk of chronic diseases based on dietary habits using RNN, unlike current approaches that consider only one time point^{(Reference Maurya, Wable and Shinde69)}.

Data-related challenges

Data-related challenges play a key role in driving data-driven innovation within the field of PN^{(Reference Tsolakidis, Gymnopoulos and Dimitropoulos56)}. To that end one of the main technical challenges in integrating AI into PN research is data heterogeneity and lack of interoperability among various data sources. PN relies on different and complementary data^{(Reference Özdemir and Kolker50,Reference Adams, Anthony and Carvajal64)} , each with its own format, structure, and collection methods. There is a lack of interoperable infrastructure and data standardisation that would enable data access and data harmonisation^{(Reference Cabitza and Campagner70)}, which is a key challenge for developing effective multimodal AI systems. Moreover, differences in data standards, tools, and platforms across healthcare systems complicate the seamless exchange and analysis of information, limiting the potential for broader applications and scaling of PN interventions. Overcoming these challenges requires standardised protocols, improved data-sharing frameworks, and better tools for data integration.

AI-related challenges

Transparency and Trust: AI models, particularly DL algorithms, have the biggest potential in terms of their application in PN but at the same time are considered ‘black boxes’, meaning their decision-making processes are not interpretable by humans. The lack of transparency reduces their trustworthiness^{(Reference Vayena, Blasimme and Cohen71)}. As a result, healthcare professionals may be hesitant to adopt AI technologies due to a lack of familiarity with AI tools, concerns about the reliability of AI-driven recommendations, or scepticism about the technology’s clinical validity. This poses a significant challenge in PN, where understanding the rationale behind AI-generated decisions and recommendations is critical for uptake by healthcare professionals and patients. Different explainable AI (XAI) methods are introduced to mitigate this problem. Though explainability and interpretability are often used interchangeably, they refer to distinct concepts and should be used carefully. Interpretability refers to the intrinsic transparency of a model’s internal structure, permitting direct inspection of how inputs mathematically or logically generate outputs. Explainability encompasses post hoc techniques, that is, methods applied after model training, applied to any model (including ‘black boxes’) to generate human-understandable rationales behind the model outputs. However, explanations obtained from different XAI methods are frequently non-identical and as such less trustworthy to guide clinical decision making^{(Reference Brankovic, Cook and Rahman33)}. They can be useful and provide insight into key factors that drive algorithms, but not the whole decision-making process. To be more trustworthy, explanations should be able to capture causality, context and match reality^{(Reference Brankovic, Cook and Rahman33)}, however this still remains a challenge in the application of AI. These reasons are in support of the view that current XAI approaches will induce trust with the health-care workforce represent a false hope^{(Reference Ghassemi, Oakden-Rayner and Beam31)}, though they may be very useful especially for system audit and troubleshooting^{(Reference Ghassemi, Oakden-Rayner and Beam31,Reference Brankovic, Cook and Rahman33)} . As an alternative, redirecting focus from XAI to interpretable AI has been suggested^{(Reference Rudin34)}.

Fairness: Lack of model opacity can complicate the identification of systematic errors or biases in the models. The Fairlearn package^{(Reference Weerts, Dudík and Edgar72)} and Fairness Indicators TensorFlow tool kit^{(Reference Weerts, Dudík and Edgar72)} can help in assessing models and mitigating unfairness.

Over-reliance: When it comes to AI use in applications, there is risk of over-reliance on AI^{(Reference Kostick-Quenet and Gerke73)}. Some studies have already reported how over-reliance on AI dialogue systems affects critical cognitive capabilities, including decision-making, critical thinking, and analytical reasoning of clinicians^{(Reference Zhai, Wibowo and Li74)}. Educating on AI limitations, using AI only as an adjunct and domain expertise to judge and evaluate AI systems can prevent risks associated with over reliance while maximising the benefits of its application.

Generalisation: State-of-the-art models lack strong empirical evidence supporting model performance across diverse and heterogeneous datasets. This limitation raises concerns about the generalisability and reliability of AI-driven dietary advice, particularly for underrepresented populations. High-quality datasets and rigorous validation are essential for ensuring AI systems provide equitable, effective, and trustworthy personalised nutrition guidance.

Data privacy and security

AI models in PN may require access to sensitive health data, such as genetic information, personal health history, and real-time biometric measurements. This raises significant concerns about data privacy and security. Unauthorised access, breaches, or misuse of health data can lead to identity theft, discrimination, or other adverse consequences^{(Reference Choi, Johnson and Lehmann75)}. Moreover, ethical and legal risks due to gaps and inconsistencies in legal protection, remain a concern^{(Reference Gostin, Halabi and Wilson76)}. It is crucial to establish robust data protection measures, implement strong encryption methods, and ensure that AI models comply with privacy regulations such as the General Data Protection Regulation^{(Reference Regulation77)} and Health Insurance Portability and Accountability Act⁽⁷⁸⁾ to safeguard personal health information.

Advancements in AI techniques

Future advancements in AI and the application of more sophisticated approaches show promise in advancing PN research.

Federated learning offers a way to train AI models across decentralised datasets, preserving individual privacy while leveraging diverse and extensive data sources. This approach is particularly relevant in healthcare, where sensitive data often cannot be centralised.

Explainable AI (XAI) is another critical development, aiming to make AI models more transparent and interpretable. By providing insights into how decisions are made, XAI will increase the level of trust healthcare professionals and individuals have in the technology, making personalised nutrition interventions more acceptable. Currently, they uptake in PN applications has been limited.

Additionally, multimodal models are gaining prominence for their ability to integrate various types of data – such as genomic, dietary, and lifestyle data – into a unified framework. These advancements will allow AI to generate more accurate, personalised, and actionable dietary recommendations.

Wholistic approach

A more holistic approach to AI in PN is essential to address the interconnected nature of human health. Integrating data from different sources can offer greater insights, supporting personalised dietary recommendations that promote overall health and wellbeing while also considering disease risk stratification and prevention. For instance, models that integrate data on stress, sleep, and physical activity alongside nutrition can provide a better picture of an individual’s health and wellbeing. This information would enable interventions to address causes rather than symptoms.

Interdisciplinary collaboration

The future of AI-driven PN relies heavily on interdisciplinary collaboration. Nutritionists, AI researchers, and policymakers should work together to develop practical, ethical, and scalable solutions. Nutritionists provide domain expertise to guide AI model development and ensure recommendations are evidence-based and clinically relevant. AI researchers bring technical expertise to design robust and innovative models, and under the risk associated with the application of different approaches, while policymakers play a crucial role in setting regulations that protect data privacy and promote equitable access. Collaborative efforts between stakeholders will ensure that AI-driven personalised nutrition benefits a broader population while maintaining ethical integrity. On that note, future advancements in AI, particularly in DL, are expected to enhance PN research. These technologies may allow better data analysis of large volume multimodal data and hence lead to more accurate predictions and more effective personalised health recommendations^{(Reference Mustapha, Mustapha and Mustapha79)} being delivered at scale.

Integration with emerging technologies

The integration of AI with emerging technologies will enhance its effectiveness and accessibility in PN. Wearable sensors and Internet of Things devices are already transforming how health data is collected, offering real-time insights into physical activity, heart rate, and caloric expenditure. Modern continuous glucose monitors, which track blood sugar levels in real time^{(Reference Reddy, Verma, Dungan, Feingold and Blackman80)} and wearable devices, provide invaluable data for dietary management, particularly for individuals with conditions like diabetes. But it is possible for their application to target the promotion of health and wellbeing as well as the prevention of disease progression. Already there are number of recommendation systems that combine data from wearable devices and AI to generate adaptive personalised dietary recommendations^{(Reference Tsolakidis, Gymnopoulos and Dimitropoulos56)}, but broadening the data sources available will support greater personalisation and tailoring to improve an individual’s overall health.