Table 4.

Summary of studies.

Lead Author, Year & Country	Aim & Topic	Study Characteristics	Findings	Limitations	Type of AI	Area in Nutrition	Observations
Maharjan, B. (2019) USA [25]	The use and development of an AI tool that is culturally adapted for Native Americans with low computer/technological skills to support diabetes management using people’s voice.	Feasibility study. AI voice-based success range based on 150 conversations = 76% to 87% Accuracy regarding recommendations considering physical and socioeconomic status = 100%.	Preliminary results showed that Alexa was able to accurately count calories and provide nutrition education. Having the potential to improve health among the target population by recommending meals using ADA’s guidelines.	AI has to learn the needs of the specific target populations to culturally adapt it to other specific groups.	NLP	Dietary assessment for patients with T2DM.	Health improvements, validation and patient’s satisfaction using this technology is yet to be determined.
Davis, C.R. (2020) Australia [26]	To evaluate the performance of AI health assistance, and to verify participants’ adherence to physical activity, diet and their engagement	Pilot study Single-arm repeated measures for 12 weeks. n = 28 Adults 45–75 y/o Accuracy of answers provided to the users = 97% Accuracy of answers for which it was not trained = 20% Adherence (step and food serving goals) = 91%	Paola (the virtual health-assistant) was successful in behavior change. However, she could not answer questions beyond what she was trained for.	The sample size of the study was too small. Due to data loss, other questions (not related to what AI was trained in) could not be evaluated. Men are underrepresented; thus, the results cannot be generalized. The platform used to launch Paola had issues with 10 min time-outs	NLP	Lifestyle intervention	Paola also provided educational videos and different recipes. Users had weekly exchanges with her for data entry and to obtain her feedback based on their entry
Maher, C.A. (2020) Australia [27]	To test the recruitment and retention of a physical activity program that was also based on a Mediterranean diet. The program was delivered by an AI virtual health coach.	Interventional n = 31 Adults 45–75 y/o Participants lost 1.3 kg and waist circumference was reduced by 2.1 cm in 12 weeks	The virtual-health assistant (Paola) successfully delivered a lifestyle intervention program helping to lose weight and to increase participant’s physical activity. AI virtual coach has room for improvement regarding its connection to people in terms of emotions.	The study was not randomized and the follow up was limited	NLP	Lifestyle intervention	Participants showed enthusiasm using a virtual-health coach. This technology may be used in other nutrition areas, such as weight loss or diabetes management
Oh, Y. (2021) USA [28]	To assess the characteristics of chatbots in terms of conversation and function, and to investigate if chatbots interventions were successful in lifestyle changes (healthy eating, exercise, weight control) and health-related outcomes.	Systematic review. n = 9. 5 studies found chatbots had positive outcomes in physical activity. 1 study showed the intervention group reported the intention to eat less meat. Chatbots’ communication was text-based. Persuasion and relational strategies were used by chatbots. Interventions between 1 week–12 weeks with an age range = 15.2–56.2 y/o	Chatbots have the potential to change lifestyle and improve access and effectiveness to personalized nutrition.	Sample sizes used by the studies were too small; thus, it is difficult to draw conclusions on employment of chatbots to deliver lifestyle changing programs.	NLP	Lifestyle intervention	Studies did not evaluate side-effects nor the possible harm that users may encounter when using them. Chatbots should be used with caution, and conversations should be monitored to avoid harmful effects.
Beyeler, M. (2023) Switzerland [29]	To evaluate the usage of a health bot (HB) and how it is perceived by patients receiving bariatric treatment.	Mixed methods approach. Observational. AI usability = 87/100 Usefulness 5.28/7 Satisfaction 5.75/7 Learnability 6.26/7 Reliable nutrition information 5.5/7.	The health bot (HB) was assessed by its response to nutrition-related questions. HB was well accepted among patients, and they found it easy to use and understand. Participants had access to useful information through the HB. However, concerns about the replacements of dietitians, personal information and the privacy of the questions were brought up by the participants.	The sample size of the study was too small. AI used for dietary assessment should not be used without supervision of a healthcare professional due to the potential misinterpretation of the HB answers. HB may exclude people with no or limited access to digital resources, and limited literacy. Making an HB easy to use should be considered.	NLP	Dietary assessment	The HB was not meant to replace consultation; instead, it was meant to be used between consultations with a dietitian.
Limketkai, B.N. (2021) USA [30]	A review of new technologies (apps, wearable devices, and AI remote nutrition assessment) and their integration in clinical nutrition and patient care.	Descriptive study Wearable devices help users to engage. Currently there are ~165,000 apps related to health and wellness and 10,000 are for weight loss and dietary purposes. 58% of the population in the U.S have downloaded an app related to health. 83% of dietitians use mobile apps in clinic.	AI-based apps and wearables devices are used by clinicians since they can be used for diet optimization and to find eating patterns, given their real-time data collection. Smartwatches (e.g., Apple watch, Kardia band) have been approved by the FDA for some health uses, shifting from wellness devices to a more medical focus.	Wearable devices are still being developed, as algorithms cannot fully differentiate between different type of foods, portions, and backgrounds. Some technologies that measure body composition have not been tested in clinical trials; thus, the accuracy of the results needs to be assessed.	ML	Dietary assessment	Some apps that offer measurements such as sleeping patterns and heart rate require a monthly subscription and a smartphone. These emerging technologies in clinical nutrition are still in their infancy and need further investigation. There is concern about the information generated and its use in medical decision making.
Morgenstern, J.D. (2022) Canada [31]	To create a machine learning prediction model, and to evaluate its efficacy in examining the connection between food intake and CVD risk.	Retrospective cohort Observational n = 12,130 with a 14-year follow-up duration. The model’s accuracy has an AUROC = 0.821	The most significant nutritional variables linked to CVD were caffeine, alcohol, supplements and sodium. Without the need of lab tests and anthropometric measurements.	Nutritional variables were used, employing one-time 24 h recall. A larger data set with more frequent dietary assessment is needed. A separate model for dietary variables vs. non-dietary variables is needed to confirm dietary information for CVD prediction.	ML	Nutrition epidemiology	No lab tests and anthropometric variables were used in ML models.
Murumkar, A. (2023) India [32]	To develop an AI-based dietician that acts like a real dietitian. It offers diets and diet plans focused on the individual.	Prospective descriptive study The expected outcomes include BMI calculation and diet recommendations based on anthropometric and demographic data. Alternative diets will be displayed if the user rejects the original diet.	Feeding AI with appropriate information, such as BMI, allergies, food preferences, physical activity, and type of job; AI has the potential to suggest eating plans according to the user’s need without having to pay for it.	The user will be uploading information (height, weight, allergies, etc.) which is self-reported.	ML	Dietary assessment	No dietitian intervention is encouraged.
Fujihara, K. (2023) Japan [33]	To build, develop and evaluate the ability of a ML model to predict variations in body weight over a 3-year period from medical examinations.	Observational. n = 55,000, with a mean age = 48 y/o and 67% males The accuracy of the model was evaluated with the root mean square error = 1.914, similar to the 1.890 of the multiple regression The model successfully predicted body weight change among adults over 3 years using existing data.	The system was able to develop 5 different formulas for body weight change prediction over a 3-year span. It successfully identified lifestyle factors that modified body weight. It has the potential to be used in weight management.	The model may not be generalizable because it was developed using a particular ethnic group. Diet and physical activity information used to build the system was self-reported. Environmental and socioeconomic factors were not considered.	ML	Weight management	5-year data were used to develop the model. Data for 50,000 individuals were used to train the model and 5000 to test it.
Yang, Z. (2021) China [34]	To mimic a dietician’s mental process using AI for food size estimation.	Observational n = 15,000 pictures (for training) The accuracy of the model for volume estimation = 86.7%	This technology can be applied to wearable devices for real food volume estimation.	It was assumed that the food on a plate can easily be detected from a real-world image that also contains other things (e.g., table, background of the picture). It relies on high-quality object detection to crop the food plate from image. The volume estimation was limited with the plate having only one type of food; whereas in real life, a plate of food has more than one item.	DL	Food estimation.	Current data sets are designed for food recognition, but not for food volume (portion size) estimation.
Taylor, S. (2021) China [35]	To develop an AI-based app to map foods on national (U.S) databases, for calories counting vs. a recommended method.	Method comparison study Observational Males and females ≥ 18 y/o n = 35 with 5 days of food intake record. COCO nutritionist was compared to the 24HR with no significant difference, showing similar results. Protein% = 16 vs. 17 Fat% = 36 vs. 36 Carbs% = 50 vs. 50 Energy (kcal/d) = 2092 vs. 2030	National databases combined with an intelligent app using NLP; can estimate energy intake with no significant difference when compared to the 24HR, which is considered the gold standard for dietary intake. This may be used for weight management. Although participants had the option to speak to the COCO nutritionist to enter their dietary intake; they preferred to type their entries.	COCO nutritionist has limited features (it does not include food photography). The 24 h recall was used; food intake may not reflect complete dietary consumption because it is self-reported. The sample size for preliminary data is small.	DL	Weight management and dietary assessment.	MIT reviewed COCO nutritionist’s data without having access to the 24 h recall. 24 h recall was analyzed using a food processor.
Papathanail, I. (2021) Switzerland [36]	To develop and evaluate an AI system that uses input images for energy and macronutrient intake before and after patient’s consumption.	Observational, n = 28 332 pictures were captured over the course of 32 days. AI’s error was less for macronutrient (<15%) and energy intake (11.64%) vs. the control (>30% for macronutrients and 31.45% for energy intake)	The system’s estimation of macronutrients intake performed better than the control (nursing staff and a medical student) in the hospital. The system provides better estimation for individual meal components. AI provides results almost instantly	Meals were not weighted; dietitians and the medical student visually estimated food percentage.	DL	Dietary assessment	This system may be used to prevent malnutrition by monitoring diet among hospitalized older patients
Chen, X. (2021) USA [50]	To assess restaurant nutrition at a big scale by using crowdsourcing food images and to develop a restaurant nutrition index (quality of food offered by the restaurant based on calories).	Pilot study (Hartford area) n = 75 restaurant pictures AI accuracy for image recognition = 75.1% vs. 94.7% (trained raters).	DL used the restaurants’ pictures to determine the quality (calorie-based) of their food. Restaurants offering foods with higher calories were found in areas with limited food access and less healthy food retailers. These results may be used in food environment inequality assessment.	The model could not identify pictures with several food items on the same plate—it was able to estimate only one; it could not identify portion sizes derived from the images. Some foods were not accurately identified. Results cannot be generalized to other geographical areas because crowdsourcing images were from food review websites from a particular area. Restaurants with no online presence were excluded. Young adults were most of the raters, and this might have influenced the type of foods that were reviewed.	DL	Dietary assessment and food environment detection	This tool is not meant to replace current dietary assessment methods, it should be used as a complementary tool only.
Van Wymelbeke-Delannoy, V. (2022) France [38]	To assess food consumption using an AI system that does not need human interaction to determine food leftovers in a hospital setting.	n = 149 dishes, with 22,544 different scenarios (pictures with different amounts of food on the plates) Observational. Food intake estimation accuracy = 57.8%	The FoodIntech project was demonstrated to be useful in picture gathering and estimating patient’s food intake by analyzing food leftovers in a hospital setting, providing instant results. With enough pictures the system can learn to recognize new foods.	The camera vision is limited; thus a 100% performance will not be achieved. AI struggles with certain food containers. Thus, food segmentation is hard to achieve. Trained staff are needed to take pictures with good resolution, lighting and clarity, for adequate dietary assessment.	DL	Dietary assessment	Food Intech was evaluated in a hospital setting, but it was not tested with the hospital’s patients. This might allow to determine whether the patient’s food intake and other factors, such as age, gender, and weight, are related to food intake.
Jin, B.T. (2022) USA [51]	To evaluate the ability of a malnutrition prediction model using longitudinal patient records.	n = 5.9 million Observational. AUROC (model’s performance evaluation) = 0.854–0.869 The model successfully identified between malnourished and non-malnourished patients using their saved records.	DL is accurate in malnutrition prediction by using patients’ longitudinal data. AI used 3 visits instead of using only the patient’s last visit information for prediction. Neither lab tests nor anthropometric measures are used in this model; (less data collection) was needed, relying only on its capacity of predictive diagnosis.	Patients with minimal records or no records were excluded. This may represent bias towards populations at higher risk.	DL	Malnutrition prediction	This model may be incorporated into current healthcare using demographic and diagnostic data. However, this model still needs to undergo clinical validation.
Sefa-Yeboah, S.M. (2021) Ghana [39]	To develop a mobile app for obesity management working both on mobile and on the web; providing personalized meal plans to meet people’s macronutrients and calories needs.	Observational, n = 30 (potential solutions) with different population for 40 days. The simulations were tested using different values (kcal) 1000, 1600, 2000, 2400, 2800 and 3000. Number of generations (rounds) = 500.	AI engine can be used for meals recommendation, and prediction to meet calorie intake for obesity management. It estimates energy intake by selecting the foods from the food record. It also shows how many calories are left to meet the calorie goal.	The system’s overall effectiveness is impacted by the limited method used to assess physical activity, which does not allow to estimate energy expenditure. Additionally, the system is limited to food selection for dietary intake.	GA	Obesity management	This system can also be useful for training those who are in the dietetics field.
Niszczota, P. (2023) Poland [40]	To assess the performance of ChatGPT on diet generation by investigating the precision and safety of 56 diets generated by ChatGPT.	Validation study n = 56 generated diets designed for 1 fictitious allergic woman. 4 out of 56 diets provided diets with allergens. No supplementation was suggested with a low D-level diet. No warning was displayed with calorie restricted diet.	ChatGPT can generate menus, but it is not always safe; it included allergens for a fictitious allergic woman. It also provided wrong calculations for portion sizes, and it could not provide varied menus, repeating the same food items. This study was carried out in Europe and sometimes the measurements were provided in American units. ChatGPT may mislead people with its dietary suggestions. However, it followed recommendations from different dietary guidelines.	This study used only one prompt (one interaction) instead of a series of interactions. Big language models cannot identify when they are providing wrong information.	GPT	Dietary assessment	Results show that AI can also be misused and needs human interaction to verify that the information provided is correct. In some countries (e.g., Italy) ChatGPT had limited access.
Arslan, S. (2023) Turkey [41]	ChatGPT’s potential for treating and managing obesity. Based on the patient’s progress and records, ChatGPT may modify its recommendations.	Letter to the editor Description of potential uses of ChatGPT in obesity management: personalized diets, it can track the user’s progress, and diets can be adjusted from that progress.	ChatGPT can offer personalized advice such as weight control, physical activity and nutrition and meet individual’s needs. Based on the patient’s progress, recommendations for weight management can be adjusted.	AI’s information might be biased, depending on the type of data that were used to train it. AI-systems do not have emotional intelligence like a human and do not offer emotional support. When GPT provides harmful and inaccurate information, it is not clear who is to blame and who is responsible.	GPT	Obesity management	AI in healthcare must be used with caution and ethical issues must be addressed, since AI systems operate without ethical and professional standards.
Sun, H. (2023) China [42]	To develop and validate an AI-nutritionist focused on T2DM.	Validation study ChatGPT (60.5%) and ChatGPT4 (74.5%) proved to be accurate in nutritional knowledge. Ketogenic diet knowledge = 80.7 with 48.81%, classified as excellent, 47.62%, classified as acceptable and 3.57%, classified as unacceptable An overlap of 94.87% between GPT and experts in recommended foods Answers were validated by expert dietitians. 23% of endocrinologists categorized pork as high glycemic food.	ChatGPT and GPT4 are competent to answer the Chinese Register Dietitian Exam and medical nutrition-related questions. It also identified food using pictures. Endocrinologists’ knowledge regarding nutrition might not be reliable. AI has potential to provide dietary assessment and meet the lack of dietitians in China.	The model was presented only with a limited set of questions that a patient may ask. One of AI’s limitations in the training process is that it can provide several answers for the same question; hence, focusing on specific questions might help to obtain more trustworthy responses.	GPT	Dietary assessment for people with T2DM	The model was not tested, nor has a pilot study been conducted. When testing is performed, the authors recommend reviewing the AI-nutritionist’s answers by a human within 48 h span to ensure no harmful/wrong information is provided to the patient, allowing this to be fixed.
Chatelan, A. (2023)Not specified [24]	To provide a guide regarding the potential hazards and benefits of using ChatGPT in clinical, academic and public health contexts.	Descriptive study ChatGPT was able to provide diets according to users’ need (e.g., a patient with T2DM and a patient undergoing dialysis). For a patient with T2DM the diet provided ADA’s recommendations.	Using ChatGPT has both opportunities and risks. It might be beneficial for people to obtain educational material for free (healthy eating, nutrition). However, ChatGPT is not always accurate and might provide harmful responses. Therefore, it should be supervised. Chatbots do not have soft skills, making it harder to replace RDs.	ChatGPT might provide nutritional advice and diets, nonetheless. It cannot provide emotional and psychological support. ChatGPT does not cite the information sources it uses to provide answers; making it hard to determine whether the sources are factual or not.	GPT	Dietary assessment	ChatGPT trainings are limited. It is not aware of information that happens thereafter (it was last trained in January 2022). Given the quick evolution of chatbots, its potential uses are hard to define.
Nunes-Galbez, N. (2022) Brazil [43]	To evaluate AI tools for conducting systematic reviews in the nutrition field	Systematic review, n = 4 75% of the studies were related to dietary assessments and the cause of diseases. The remaining study (25%) was focused on sports nutrition.	The publication dates range from 2015 to 2021. All the retrieved publications are from developed countries. The small number of studies shows that AI is still novel in systematic reviews in nutrition.	The studies that did not address any challenges could be useful when considering the use of these technologies. The review was limited by the number of studies that were included.	LDA	Use of AI tools in systematic reviews in nutrition.	Big data has resulted in an exponential growth in scientific papers. It is hard for scientists to conduct a systematic review without losing data. In consequence, the use of AI has been proposed.
Bond, A. (2023) UK [44]	To identify areas and applications in nutrition where AI might play a role.	Descriptive study DL may deliver dietary assessment that is comparable to or exceed that of a certified dietitian.	AI can be used to enhance healthcare by interpreting images, making prescriptions, and to provide nutritional advice. In a hospital setting, patients can benefit from these technologies instead of waiting for the dietitian. Other uses in healthcare are expected to be developed.	AI might be biased since it is trained by humans. Thus, training AI should be performed with caution. If not trained properly, AI might face opposition within healthcare. Ethical concerns and how AI deals with personal information are still complex.	DL, ML, NLP	Dietary assessment	Not losing control of AI and how it is used in healthcare is of extreme importance. Healthcare personnel should understand how to use AI

Abbreviations: AI: artificial intelligence; cm: centimeters; T2DM: Type 2 Diabetes Mellitus; apps: applications; FDA: Food and Drug Administration; 24HR: 24 h recall; y/o: years old; kcal: kilocalories; MIT: Massachusetts Institute of Technology; CVD: cardiovascular disease; AUROC: area under the receiver-operating characteristic curve; BMI: body mass index; ADA: American Diabetes Association; DL: deep learning; ML: machine learning; NLP: natural language processing; GA: genetic algorithm; GPT: generative pre-trained transformer; LDA: latent Dirichlet allocation.