Applications of Artificial Intelligence to Obesity Research: Scoping Review of Methodologies

Introduction

Background

The double burden of malnutrition, characterized by the coexistence of overnutrition (eg, overweight and obesity) and undernutrition (eg, stunting and wasting), is present at all levels of the population: country, city, community, household, and individual [1]. Obesity is a leading cause of preventable death and consumes substantial social resources in many high-income and some low- and middle-income economies [2]. Worldwide, the obesity rate has nearly tripled since 1975 [3]. In 2016, 13% of the global population, or 650 million adults, were obese [4]. More than 340 million children and adolescents aged 5 to 19 years and 39 million children aged <5 years were overweight or obese [4]. By 2025, the global obesity prevalence is projected to reach 18% among men and 21% among women [5].

Health data are now available to researchers and practitioners in ways and quantities that have never existed before, presenting unprecedented opportunities for advancing health sciences through state-of-the-art data analytics [6]. By contrast, dealing with large-scale, complex, unconventional data (eg, text, image, video, and audio) requires innovative analytic tools and computing power only available in recent years [7,8]. Artificial intelligence (AI), characterized by machine learning (ML) and deep learning (DL), has become increasingly recognized as an indispensable tool in health sciences, with relevant applications expanding from disease outbreak prediction to medical imaging and patient communication to behavioral modification [9-14]. Over the past decade, an upsurge of the scientific literature adopting AI in health research has been witnessed [15,16]. These investigations applied a wide range of AI models: from shallow ML algorithms (eg, decision trees (DTs) and k-means clustering) and deep neural networks [17] to various data sources (eg, clinical and observational) and types (eg, tabular, text, and image) [18]. This boom in AI applications raises many questions [19-21]: How do AI-based approaches differ from conventional statistical analyses? Do AI techniques provide additional benefits or advantages over traditional methods? What are the typical AI applications and algorithms applied in obesity research? Is AI a buzzword that will eventually fall out of fashion, or will the upward trend of AI adoption to study obesity continue in the future?

Synthesizing and Disseminating AI Methodologies Adopted in Obesity Research

Three previous studies reviewed the applications of AI in weight loss interventions through diet and exercise [22-24]. They found preliminary but promising evidence regarding the effectiveness of AI-powered tools in decision support and digital health interventions [22-24]. However, to our knowledge, no study has been conducted to summarize AI algorithms, models, and methods applied to obesity research. This study remains the first methodological review on the applications of AI to measure, predict, and treat childhood and adult obesity. It serves 2 purposes: synthesizing and disseminating AI methodologies adopted in obesity research. First, we focused on summarizing and categorizing AI methodologies used in the obesity literature in the hope of identifying synergies, patterns, and trends to inform future scientific investigations. Second, we provided a high-level, beginner-friendly introduction to the core methodologies for interested readers, aiming to facilitate the dissemination and adoption of various AI techniques.

Methods

The scoping review was conducted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines [25].

Study Selection Criteria

Studies that met all of the following criteria were included in the review: (1) study design: experimental or observational studies; (2) analytic approach: use of AI, including ML and DL (ie, deep neural networks), in measuring, predicting, or intervening obesity-related outcomes; (3) study participants: humans of all ages; (4) outcomes: obesity or body weight status (eg, BMI, body fat percentage [BFP], waist circumference [WC], and waist-to-hip ratio [WHR]); (5) article type: original, empirical, and peer-reviewed journal publications; (6) time window of search: from the inception of an electronic bibliographic database to January 1, 2022; and (7) language: articles written in English.

Studies that met any of the following criteria were excluded from the review: (1) studies focusing on outcomes other than obesity (eg, diet, physical activity, energy expenditure, and diabetes); (2) studies that used a rule-based (hard-coded) approach rather than example-based ML or DL; (3) articles not written in English; and (4) letters, editorials, study or review protocols, case reports, and review articles.

Search Strategy

A keyword search was performed in 2 electronic bibliographic databases: PubMed and Web of Science. The search algorithm included all possible combinations of keywords from the following two groups: (1) “artificial intelligence,” “computational intelligence,” “machine intelligence,” “computer reasoning,” “machine learning,” “deep learning,” “neural network,” “neural networks,” or “reinforcement learning” and (2) “obesity,” “obese,” “overweight,” “body mass index,” “BMI,” “adiposity,” “body fat,” “waist circumference,” “waist to hip,” or “waist‐to‐hip.” The Medical Subject Headings terms “artificial intelligence” and “obesity” were included in the PubMed search. Multimedia Appendix 1 documents the search algorithm used in PubMed. Two coauthors of this review independently conducted title and abstract screening on the articles identified from the keyword search, retrieved potentially eligible articles, and evaluated their full texts. The interrater agreement between the 2 coauthors was assessed with Cohen kappa (κ=0.80). Discrepancies were resolved through discussion.

Data Extraction and Synthesis

A standardized data extraction form was used to collect the following methodological and outcome variables from each included study: authors; year of publication; country; data collection period; study design; sample size; training, validation, and test set size; sample characteristics; the proportion of female participants; age range; AI models used; input data source; input data format; input features; outcome data type; outcome measures; unit of analysis; main study findings; and implications for the effectiveness and usefulness of AI in measuring, predicting, or intervening obesity-related outcomes.

Methodological Review

We classified AI methodologies adopted by the included studies into 2 primary categories: ML and DL models. Among the ML models, methods were organized into 2 subcategories: unsupervised and supervised learning. Among the DL models, methods were classified into 3 subcategories: tabular data modeling, computer vision (CV), and natural language processing (NLP). Rather than enumerating every single model performed by the included studies, which is unnecessary and unilluminating, we focused on the popular models used by multiple studies.

Results

Identification of Studies

Figure 1 shows the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram. We identified a total of 3090 articles through the keyword search, and after removing 499 (16.15%) duplicates, 2591 (83.85%) unique articles underwent title and abstract screening. Of these 2591 articles, 2532 (97.72%) were excluded, and the full texts of the remaining 59 (2.28%) were reviewed against the study selection criteria. Of these 59 articles, 13 (22%) were excluded. The reasons for exclusion were as follows: no adoption of AI technologies (1/13, 8%), no obesity-related outcomes (11/13, 85%), and commentary rather than original empirical research (1/13, 8%). Therefore, of the 3090 articles identified initially through the keyword search, 46 (1.49%) were included in the review [26-71].

‎

Study Characteristics

Table 1 summarizes the key characteristics of the 46 included studies. An increasing trend in relevant publications was observed. The earliest study included in the review was published in 1997; others were published in, or after, 2008; for example, 2% (1/46) each in 2008, 2012, and 2017; 4% (2/46) each in 2014 and 2016; 7% (3/46) each in 2009 and 2015; 9% (4/46) in 2018; 15% (7/46) in 2019; 20% (9/46) in 2020; and 26% (12/46) in 2021. Of the 46 studies, 16 (35%) were conducted in the United States [28,32,33,37,42,46,48, 50-53,57,58,60,62,63]; 6 (13%) in China [39,40,45,56,64,65]; 3 (7%) each in the United Kingdom [27,68,69] and Korea [35,43,49]; 2 (4%) each in Italy [36,71], Turkey [41,70], Finland [44,59], Germany [54,55], and India [36,71]; and 1 (2%) each in Saudi Arabia [26], Iran [67], Serbia [66], Portugal [61], Spain [47], Singapore [38], Australia [34], and Indonesia [29]. Of the 46 studies, 32 (70%) adopted a cross-sectional study design [26,27,29-32,37,39-42,46-50,52,55-58,60-63,65-71], 7 (15%) a prospective study design [28,33,38,43,45,54,59], 6 (13%) a retrospective study design [34-36,51,53,64], and 1 (2%) a cotwin control design [44]. Sample sizes varied substantially across the included studies, ranging from 20 to 5,265,265. Of the 46 studies, 7 (15%) had a sample size of between 20 and 82; 11 (24%) between 130 and 600; 19 (41%) between 1061 and 9524; 6 (13%) between 16,553 and 49,805; 2 (4%) between 244,053 and 618,898; and 1 (2%) study had a sample size of 5,265,265. Of the 46 studies, 23 (50%) focused on adults, 14 (30%) on children and adolescents, 1 (2%) on people of all ages, and the remaining 8 (17%) did not report the age range of participants.

Data Sources and Outcome Measures

Table 2 summarizes the data sources and outcome measures of the studies included in the review. Input data were obtained from a variety of sources, including health surveys (eg, National Health and Nutrition Examination Survey), electronic health records, magnetic resonance imaging (MRI) scans, social media data (eg, tweets), and geographically aggregated data sets (eg, InfoUSA and Dun & Bradstreet). Of the 46 studies, 34 (74%) analyzed tabular data (eg, spreadsheet data) [26-29,33-37,39,41,44-47,49-51,53-56,58-68,70], 8 (17%) analyzed digital image data [30,31,38,40,42,43,57,71], and 4 (9%) analyzed text data [32,48,52,69]. Obesity-related measures used across the studies included anthropometrics (eg, body weight, BMI, BFP, WC, and WHR) and biomarkers.

Main Findings

Table 3 summarizes the estimated effects and main findings of the studies included in the review. Four key findings have emerged.

First, the studies found that ML or DL models were generally effective in detecting clinically meaningful patterns of obesity or relationships between covariates and weight outcomes; for example, ML and DL models were found useful in classifying obesity severity [30,47,52], identifying anthropometric [34] and genetic characteristics of obesity [56], and predicting obesity onset in children [28,53,63]. ML algorithms (eg, random forest [RF] and conditional RF) revealed meaningful relationships between school and neighborhood environments and overweight and obesity [45,58,62]. DL algorithms (eg, convolutional neural network [CNN]) effectively extracted built environment features from satellite images to assess their associations with the local obesity rate [57].

Second, most (18/22, 82%) of the studies comparing AI models with conventional statistical methods reported that the AI models achieved higher prediction accuracy on test data, whereas others (4/22, 18%) found similar model performances; for example, ML and DL models were found to explain a larger proportion of variations in county-level obesity prevalence than conventional statistical approaches [50]. ML models showed flexibility in handling various variable types [36,41] and large-scale data sets [32] and producing robust, generalizable inferences [41,54,64,65] with higher prediction accuracy [61,66]. By contrast, Cheng et al [37] reported that ML algorithms and conventional statistical approaches had similar performance.

Third, some (5/46, 11%) of the studies comparing the performances of different AI models yielded mixed results, reflecting the interdependence between model and data or task; for example, logistic regressions were reported to achieve higher prediction accuracy than DTs, naïve Bayes (NB) [29], and DL [35]. By contrast, Heydari et al [67] found that logistic regressions and DL models performed equally well in solving classification problems. Zhang et al [68] and Ergün [70] reported that data mining and DL techniques outperformed logistic regressions in classification accuracy.

Fourth, newer studies increasingly adopted state-of-the-art DL models to address CV and NLP tasks; for example, chatbots built on NLP models were used to support pediatric obesity treatment [48]. CNN-based CV models were used to construct indicators for the built environment using images from Google Street View [42]. DL-based tools were used to efficiently visualize and analyze abdominal visceral adipose tissue and subcutaneous adipose tissue [38].

Methodological Review

AI Overview

AI symbolizes the effort to automate intellectual tasks usually performed by humans [72]. In general, AI consists of 2 domains or developmental periods: symbolic AI and modern AI [73]. Symbolic AI prevailed from the 1950s to the 1980s, characterized by the endeavors to achieve human-level intelligence by having programmers handcraft a sufficiently large set of explicit rules for manipulating knowledge [74]. Although symbolic AI proved suitable for solving well-defined, logical problems, such as a rule-based question-answer system, it became intractable when creating rules to solve more complex, fuzzy issues such as image classification, speech recognition, and language translation [74]. The definition of ML is “the field of study that gives computers the ability to learn without being explicitly programmed” [75]. Instead of hard coding all the rules in the symbolic AI, researchers provide examples (eg, images with labels that identify the objects in them) to train modern ML models to output rules [74]. As a subdomain of ML, DL is based on artificial neural networks in which multiple (deep) layers of artificial neurons are used to progressively extract higher-level features from data [76]. This layered representation enables the modeling of more complex, dynamic patterns compared with traditional ML (which sometimes is called shallow learning in contrast to DL), which finds its utility in analyzing big data: data massive in scale and messy to work with (eg, unstructured texts and images) [77]. The first ML and DL algorithms were developed in the 1950s, attracting initial excitement but then lying dormant for several decades [72]. Since the late 1980s, partly because of the rediscovery of backpropagation algorithms, the invention of CNNs, and the strong growth in computational capacity, ML and DL have regained their popularity vis-à-vis symbolic AI [72].

AI Versus Conventional Statistical Methods

Admittedly, the concept of conventional statistical methods is dubious at best because the development of statistical theories and algorithms is continual in time and intertwines at all levels [78]. Indeed, many conventional models fall into the ML domain, such as linear and logistic regressions. Despite the poorly defined domain and overlapping algorithms, at least 2 distinctions could be made between modern AI (ie, ML and DL) and other statistical methods. In terms of aims, the objective of AI models and their evaluation metrics predominantly concern prediction precision (often at the cost of compromising interpretability as models become complex) [78,79]. By contrast, conventional statistical approaches usually attempt to reveal relationships among variables (statistical inference) and focus on model interpretability [80]. In terms of procedures, it is standard practice to split data into training, validation, and test sets so that an AI model can be trained using the training set with the aim of achieving the optimal performance on some predefined evaluation metrics (eg, accuracy and mean squared error) when testing on the validation set [81,82]. The fine-tuned AI model is subsequently tested on the test set. The utility of the validation set is to prevent model overfitting (ie, too tailored to the training set while losing generalizability to new, unseen data) and fine-tune hyperparameters (ie, parameters external to the model, whose values cannot be automatically learned from data). The test set is preserved to test the final model’s performance on unseen data. By contrast, conventional statistical methods do not usually fit and evaluate models using training, validation, and test sets but use other model selection criteria (eg, adjusted R-squared and Akaike and Bayesian information criteria) to evaluate model performance [83].

ML Subcategories

Overview

ML is classified into 2 subcategories: unsupervised ML and supervised ML [84]. Unsupervised ML analyzes and clusters unlabeled data sets, discovering hidden patterns or data groupings without the need for human intervention [85]. Its capability to reveal similarities and differences in information makes it ideal for exploratory data analysis. Unsupervised ML models are used for 3 main tasks: clustering, association, and dimensionality reduction [86]. Clustering algorithms (eg, k-means clustering, hierarchical clustering, and Gaussian mixture) group unlabeled data based on similarities [86]. Association algorithms (eg, Apriori, Eclat, and FP-Growth) identify rules and relations among variables in large databases [87]. Dimensionality reduction algorithms (eg, principal component analysis [PCA], singular value decomposition, and multidimensional scaling) deal with an excessive number of features during data preprocessing, reducing them to a manageable size while preserving the integrity of the data set as much as possible [88]. Supervised ML uses a training set consisting of input-output pairs to enable the algorithm to learn a function that maps input to output over time [89]. The algorithm measures its accuracy through the loss function, adjusting until the error is minimized sufficiently. The critical difference between supervised ML and unsupervised ML is that the former requires labeled data (ie, input-output pairs), whereas the latter only requires inputs (ie, unlabeled data) [84]. Supervised ML models are used for 2 main tasks: classification and regression [84]. Classification algorithms assign data to specific categories (eg, obese or nonobese). Regression algorithms learn the relationship between input features and continuously distributed outcomes and are commonly used for projections (eg, BMI in 5 years).

Unsupervised ML

K-means Clustering

K-means clustering is an iterative algorithm that tries to partition the data set into a total of k nonoverlapping groups (ie, clusters) [86,90]. Each data point belongs to only 1 group. The algorithm attempts to make the intracluster data points as similar as possible while keeping the clusters apart. In particular, it assigns data points to a cluster such that the sum of the squared distance between the data points and the cluster’s centroid (ie, arithmetic mean of all the data points belonging to that cluster) is minimized. As the number of clusters k needs to be determined before implementing the algorithm, silhouette coefficients are commonly used to identify the optimal k value. Lin et al [34] used k-means clustering to classify patients with obesity into 4 groups based on 3 biomarkers concerning glucose, insulin, and uric acid.

Fuzzy C-means Clustering

In nonfuzzy clustering (also known as hard clustering; for example, k-means clustering), data are divided into distinct clusters, where each data point can only belong to 1 cluster [86]. In fuzzy clustering, data points can potentially belong to multiple clusters [91]. Fuzzy c-means clustering assigns each data point membership from 0% to 100% in each cluster center [92]. The fuzzy partition coefficient is often used to determine the optimal number of clusters with a value ranging from 0 (worst) to 1 (best) [93]. Positano et al [71] used the fuzzy c-means algorithm to classify MRI pixels into clusters to assess abdominal fat.

Group Factor Analysis

Factor analysis describes relationships among the individual variables of a data set [94]. Group factor analysis (GFA) extends this classical formulation into describing relationships among groups of variables, where each group represents either a set of related variables or a data set [95]. GFA is commonly formulated as a latent variable model consisting of 2 hierarchical levels: the higher level models the relationships among the groups, and the lower-level models the observed variables given the higher level [95]. Kibble et al [44] used GFA to jointly analyze 5 large multivariate data sets to understand the multimolecular-level interactions associated with obesity development.

PCA for Large Data Sets

Large data sets are increasingly common nowadays. PCA is a classic, widely adopted method to reduce the dimensionality of a large data set while preserving as much statistical information (ie, variability) as possible [86]. In particular, PCA attempts to find new variables, called principal components, that are linear functions of those in the original data set. The new variables are uncorrelated with each other (ie, orthogonal) and maximize the projected data variance. Rashmi et al [31] used PCA to reduce the feature dimensions of a thermal imaging data set to classify children by their obesity severity level.

Supervised ML

Linear Regression

Linear regression is considered a conventional statistical model and a classical architecture to develop a predictive model [96], but it fulfills all criteria from an ML point of view and is widely used as an ML algorithm to predict continuous outcomes such as BMI or BFP [97]. Trainable weights (ie, coefficients) of linear regression are commonly estimated using ordinary least squares or gradient descent. Compared with many other ML models, linear regression has the advantages of simplicity and interpretability [98]. It is easy to understand how the model reaches its predictions. Wang et al [56] used linear regressions to identify features of single-nucleotide polymorphisms that predict obesity risk. Phan et al [42] used linear regressions to estimate the associations between built environment indicators and state-level obesity prevalence.

Regularized Linear Regression

The bias-variance tradeoff is a fundamental issue faced by all ML models [86,99]. Bias is an error from erroneous assumptions in a learning algorithm. High bias may cause the algorithm to miss the relevant relations between features and outputs (called underfitting). Variance is an error from a learning algorithm’s sensitivity to small fluctuations in the training set. A high variance may result from the algorithm modeling the random noise in the training data, often leading to the algorithm’s poor generalizability to new, unseen data (called overfitting). In general, decreasing variance increases bias and vice versa, and ML algorithms need to be fine-tuned to balance these 2 properties. Regularization is an essential technique to prevent model overfitting and improve generalizability (at the cost of increasing bias) by adding a penalty term of trainable weights to the loss function [86]. Optimization algorithms that minimize the loss function will learn to avoid extreme weight values and thus reduce variance. The penalty term with the sum of squared trainable weights is called L2 regularization, used in Ridge regression. The penalty term with the sum of the absolute values of trainable weights is called L1 regularization, used in the least absolute shrinkage and selection operator (LASSO) regression. Unlike Ridge regression, LASSO regression often shrinks some feature weights to absolute zero, making it useful for feature selection. Finally, ElasticNet regression uses a weighted sum of L1 and L2 regularizations. Gerl et al [54] used LASSO regression to estimate the relationship between human plasma lipidomes and body weight outcomes, including BMI, WC, WHR, and BFP.

Logistic Regression

In its simplest form, logistic regression uses a logistic function, called the sigmoid function, to model a binary outcome [100]. A sigmoid function is a continuous, smooth, differentiable S-shaped mathematical function that maps a real number to a value in the range of 0 and 1, making it ideal for modeling probabilities. The estimated probabilities are converted to predictions (0 or 1, denoting exclusive group membership) based on some predefined threshold (eg, >0.5). In ML, logistic regression often incorporates regularizations (L1, L2, or both) to prevent overfitting. Another common extension of logistic regression in ML is to solve multiclass classification problems when classification tasks involve >2 (exclusive) classes. A typical strategy uses the one-vs-rest method (also called one-vs-all) that fits 1 classifier (eg, a logistic regression) per class against all the other classes [101]. A data point is assigned to the class with the highest confidence score among all classifiers. Thamrin et al [29] used logistic regressions to assess the predictability of various obesity risk factors. Cheng et al [37] used logistic regressions to classify obesity status based on participants’ physical activity levels.

NB Classifier

NB algorithms apply the Bayes theorem with the naïve assumption of conditional independence among each pair of features given the value of the class [102]. Despite this oversimplified assumption, NB classifiers have been widely used and have worked well in solving many real-world problems. The decoupling of conditional feature distributions allows each distribution to be independently estimated as 1D, making the training of NB classifiers much faster than more sophisticated ML models [86]. By contrast, the predicted probabilities of NB classifiers are less trustworthy owing to the algorithm’s naïve assumption. Rashmi et al [31] used NB to classify childhood obesity based on thermogram images. Thamrin et al [29] adopted NB to predict adult obesity using Indonesian health survey data [29].

K-nearest Neighbor

K-nearest neighbor (KNN) is a nonparametric, supervised learning algorithm suitable for classification and regression tasks [103]. The input consists of the k closest training data points based on a prespecified distance measure (eg, Euclidean, Manhattan, or Minkowski distance). For classification tasks, the output is a class membership. A test data point is assigned to the class most common among its k-nearest neighbors (if k=1, the test data point is assigned to the class of the single nearest neighbor). For regression tasks, the output is the average value of its k-nearest neighbors. KNN should not be confused with k-means. The former is a supervised ML algorithm to determine the class or value of a data point based on its k-nearest neighbors, whereas the latter is an unsupervised ML algorithm to classify data points into k clusters that minimize the distances within clusters while maximizing those between clusters [90]. KNN is a memory-based learning algorithm that requires no training (called a lazy learner) but can become significantly slower when the sample size increases. Wang et al [56] used KNN to predict obesity risk based on features of single-nucleotide polymorphisms. Ramyaa et al [51] performed KNN to predict body weight using physical activity and dietary data.

Support Vector Machines

Support vector machines (SVMs), which are supervised learning models that construct a hyperplane in a high-dimensional space, can be used for classification and regression tasks [104]. SVMs attempt to identify the hyperplane separating different classes while maximizing the distance to any class’s nearest training data point (ie, margin). Intuitively, the larger the margin, the more likely the model’s generalizability to new, unseen data. The choice of margin type can be critical for SVMs [86]. Hard-margin SVMs maximize the margin by minimizing the distance from the decision boundary to the training points. However, hard-margin SVMs may lead to overfitting and have no solution if the training data are linearly inseparable. Soft-margin SVMs modify the constraints of the hard-margin SVMs by allowing some data points to violate the margin (ie, misclassified). In practice, data are seldom linearly separable in the original feature space, and kernel methods are applied to map the input space of the data to a higher-dimensional feature space where linear models can be trained [105]. Many kernel functions, such as the Gaussian radial basis, sigmoid, and polynomial kernel, can be chosen. Wang et al [56] used SVM to predict obesity risk based on the features of single-nucleotide polymorphisms. Ramyaa et al [51] applied SVM to predict body weight using physical activity and diet data.

DT Algorithms

DTs are nonparametric supervised learning methods for classification and regression tasks [106]. In DT algorithms, a tree is built by splitting the source set that constitutes the tree’s root node into subsets, which comprise the successor children [107]. The splitting is based on a set of rules applied to input features. Different splitting rules exist, such as variance reduction for regression tasks and Gini impurity or information gain for classification tasks. The splitting process is repeated on each derived subset recursively (ie, recursive partitioning). The recursion is completed when all subsets at a node share the same target value or when splitting no longer adds value to the predictions. DTs have several advantages over other ML algorithms, such as high transparency and interpretability and few requirements for data preprocessing [108]. However, DTs can be prone to overfitting (ie, too confident about the rules learned from the training set, which does not generalize well to the test set) and instability (minor variations in the data resulting in a very different tree). Using features extracted from electronic medical records, Hong et al [52] used DTs to predict obesity and 15 other comorbidities. Taghiyev et al [41] performed DTs to identify risk factors associated with obesity onset.

RF Models

Ensemble methods are approaches that aggregate the predictions of a group of models aiming for improved performance in classification or regression tasks [109]. Various ensemble methods exist, such as bagging, pasting, boosting, and stacking [86]. Bagging and pasting use the same training algorithm for every predictor included in the ensemble and train it on different random subsets of the training set. When sampling is performed with replacement, the method is called bagging; when sampling is performed without replacement, it is called pasting. RF is an ensemble of DTs commonly trained via the bagging or pasting method [110]. Specifically, RF fits many DTs on various subsets of the data and uses averaging to improve the predictive accuracy and prevent overfitting. For classification tasks, the RF output is the class selected by most trees; for regression tasks, the mean prediction of the individual trees is used. Some common hyperparameters of RF for fine-tuning include the number of trees in the forest, the maximum number of features considered for splitting a node, the maximum number of branches in each tree, the minimum number of data points placed in a node before the node is split, the minimum number of data points allowed in a leaf node, and the method for sampling data points (ie, with or without replacement) [86]. RF typically produces more accurate and robust predictions than DTs and is one of the most popular supervised ML algorithms [111]. Using RF models, Hinojosa et al [58] examined the relationship between social and physical school environments and childhood obesity in California, United States. Dunstan et al [46] performed RF to predict national obesity prevalence using food sales data from 79 countries.

Extreme Gradient Boosting

Boosting refers to any ensemble method that combines several weak models into a strong one [112]. The difference between boosting and bagging and pasting is that in boosting, different models are applied to the entire training set sequentially, the new model attempting to address the weaknesses (eg, misclassified targets and residual errors) of the previous model. By contrast, in bagging and pasting, the same models are trained on different random subsets of the training set. A popular boosting algorithm is gradient boosting, in which the new model is trained on the residual errors made by the previous model [113]. Extreme gradient boosting (XGBoost) implements an optimized, parallel-tree gradient boosting algorithm, aiming to be highly efficient, flexible, and portable [114]. XGBoost is considered one of the most powerful ML algorithms, often serving as an essential component of winning entries in ML competitions [86]. A few drawbacks of XGBoost include lacking interpretability and being prone to overfitting. Pang et al [33] used XGBoost to predict early childhood obesity based on electronic health records. Alkutbe et al [27] applied gradient boosting to predict BFP based on cross-sectional health survey data collected in Saudi Arabia.

Multivariate Adaptive Regression Splines

Multivariate adaptive regression splines (MARS) is a nonparametric regression technique that automatically models nonlinearities and interactions among variables by combining ≥2 linear regressions using hinge functions [115,116]. A hinge function is a function equal to its argument where that argument is >0 and 0 everywhere else. MARS builds a model using a 2-phase procedure [117]. The forward phase starts with a model consisting of only the intercept term (ie, mean of the target) and repeatedly adds basis functions (ie, constant or hinge function) in pairs to the model that minimizes the squared error loss of the training set. The backward (or pruning) phase usually starts with an overfitted model and removes its least effective term at each step until the best submodel is found. MARS requires little or no data preparation, is easy to understand and interpret, and can address classification and regression tasks. However, it often underperforms boosting ensemble methods. Shao [65] applied MARS to predict BFP using a small-scale health record data set.

DL Models

In the obesity literature reviewed, DL models were applied to 3 distinct data types: tabular data (eg, spreadsheet data), images, and texts. The model architectures differ systematically across these data types.

DL on Tabular Data

Although shallow ML models perform well on tabular data sets in most cases, some complex relationships between the features and the target could be more effectively learned by a deep neural network model [118]. A fully connected neural network consists of a series of fully connected layers, with each artificial neuron (ie, node) of a layer linking with all neurons in the following layer [76]. A multilayer perceptron (MLP) is a classic fully connected neural network consisting of at least 3 layers of neurons: an input layer, a hidden layer, and an output layer [119]. One advantage of fully connected neural networks is that they are structure agnostic, requiring no specific assumptions about the input. However, neural networks trained on tabular data can sometimes be prone to overfitting [120]. Park and Edington [121] used MLP to identify individuals at elevated diabetic risk. Heydari et al [67] performed MLP to predict obesity status using data from a cross-sectional study of military personnel in Iran.

DL on Images

CV is a field of AI that enables computers to learn from digital images, videos, or other visual inputs and derive meaningful information for decision-making and recommendations [122,123]. Nowadays, most CV applications use DL models, which prove more capable than their shallow-learning (ie, ML models) counterparts in representing and revealing high-dimensional, complex nonlinear patterns inherent in image data. Specifically, CNNs consistently outperform the traditional densely connected neural networks (eg, MLP) and achieve human-like or superhuman accuracy in many challenging CV tasks ranging from image classification to object detection and segmentation [124,125]. The main advantages of CNNs over densely connected neural networks are locality, translation invariance, and computational efficiency [126]. Locality refers to the repeated use of small-sized kernels (or filters) in CNNs to identify local patterns at an increasing level of complexity (eg, from basic shapes such as lines and edges to complex objects such as adipose tissue or brain tumor). Translation invariance refers to CNNs’ capacity to detect an entity independent of its position in the image. The computational efficiency of CNNs is achieved by using kernels, global pooling, and other techniques, which typically make the models much smaller (ie, fewer learnable parameters) than their densely connected counterparts. Over the past decade, numerous CNN-based DL models were built and adopted to tackle domain-specific CV problems [76,127]. Some landmark models include, but are not limited to, LeNet, AlexNet, VGG, Inception, ResNet, Xception, ResNeXt, and U-Net.

Transfer learning plays a crucial role in modern AI, where a model developed for a task is reused as the starting point for a model on a different but related task [128]; for instance, the ResNet model trained on ImageNet data with >14 million images in approximately 1000 categories (eg, tables and horses) has stored many useful visual patterns in its weights, which can help solve other CV tasks (eg, identifying fat tissues in MRI scans) [129]. Transfer learning can substantially reduce the number of images required to train a model for a particular task and boost model performance compared with models trained from scratch [130].

Maharana and Nsoesie [57] adopted the VGG model architecture to examine the relationship between obesity prevalence and the built environment measured by Google Maps images (eg, parks, highways, green streets, crosswalks, and diverse housing types). Similarly, Phan et al [42] used the VGG model to assess the link between the statewide prevalence of obesity, physical activity, and chronic disease mortality and the built environment using images from Google Street View. Bhanu et al [38] applied the U-Net model to identify adipose tissues from MRI data. Snekhalatha and Sangamithirai [30] applied transfer learning on a pretrained CNN model to detect obesity based on thermal imaging data.

DL on Text

Besides CV, NLP is another field where DL dominates [131]. Early NLP models primarily adopted recurrent neural network (RNN) architecture, demonstrating broad applicability to various NLP tasks such as sentiment analysis, text summarization, language translation, and speech recognition [74,132]. RNN differs from feed-forward MLP in that it takes information from prior inputs (stored as memories) to influence the current input and output, which capitalizes on the structure of sequential data where order matters (eg, time series or natural languages) [133]. Some popular RNN models used in NLP tasks include gated recurrent unit and long short-term memory unit [74]. However, in today’s NLP landscape, transformers, invented by a team at Google in 2017, have surpassed RNN models such as gated recurrent unit and long short-term memory unit [134-136]. Transformers are encoder-decoder models that use self-attention to process language sequences [137]. An encoder maps an input sequence into state representation vectors. A decoder decodes the state representation vector to generate the target output sequence. The self-attention mechanism is used repeatedly within the encoder and the decoder to help them contextualize the input data. Specifically, the mechanism compares every word in the sentence to every other word, including itself, and reweighs each word’s embeddings to incorporate contextual relevance. Popular transformer models such as GPT-3, BERT, XLNet, RoBERTa, and T5 have been widely applied to various NLP tasks and achieved state-of-the-art results [137]. Stephens et al [48] tested the efficacy of pediatric obesity treatment support through Tess, a behavioral coaching chatbot built on NLP models. The study concluded that Tess demonstrated therapeutic values to pediatric patients with obesity and prediabetes, especially outside of office hours, and could be scaled up to serve a larger patient population.

Discussion

Overview

This study conducted a scoping review of the applications of AI to obesity research. A keyword search in digital bibliographic databases identified 46 studies that used diverse ML and DL models to study obesity-related outcomes. In general, the studies found AI models helpful in detecting clinically meaningful patterns of obesity or relationships between specific covariates and weight outcomes. The majority (18/22, 82%) of the studies comparing AI models with conventional statistical approaches found that the AI models achieved higher prediction accuracy on test data. Some (5/46, 11%) of the studies comparing the performances of different AI models revealed mixed results, likely indicating the high contingency of model performance on the data set and task it was applied to. An accelerating trend of adopting state-of-the-art DL models over standard ML models was observed to address challenging CV and NLP tasks. We concisely introduced the popular ML and DL models and summarized their specific applications in the studies included in the review.

Despite the variety of ML and DL models used in obesity research, it could well be the beginning of the trend for using AI applications in the big data era. Future adoptions of AI in obesity research could be influenced by a broad spectrum of factors, with 3 prominent ones discussed in the following sections.

Artificial General Intelligence

The ML and DL models reviewed in this study were primarily unimodal and task specific: they were built on a single data type (eg, tabular, text, or image) to solve a specific problem such as obesity classification or BMI prediction. Recent advances in AI showcase the feasibility and possibly superior performance of multimodal, multitask ML and DL models that are trained on diverse data types (eg, tabular plus text, image, video, or audio) and can handle many domains of downstream tasks (eg, text generation, object detection, time series prediction, and speech recognition) simultaneously [138-140]. However, it should be noted that the predictive accuracy of AI models may vary across gender and age groups [27] and sex and age groups [59]. Different from BMI, BMI z scores adjust for sex and age differences [141]. Future research may evaluate the potential disparities in AI model performances in their applications to BMI versus BMI z scores as outcome measures. Artificial general intelligence (AGI) refers to the ability of an intelligent agent to understand or learn any intellectual task performed by a human being [142,143]. It is too early to tell whether these multimodal, multitask ML and DL models may lead to AGI (or whether we could ever achieve AGI through technological innovations) [144]. Nevertheless, we may soon witness increasing applications of these models in obesity-related research.

Synthetic Data Generation

Data access is fundamental to any AI model training. Two primary barriers with regard to data are limited sample size and confidentiality concerns [145-148]. ML and DL models are increasingly used to generate synthetic data as an alternative to data collected from the real world [149,150]. Synthetic data do not contain private information requiring human subject review and, therefore, can be shared with other parties or the public without confidentiality concerns [151]. By contrast, synthetic data preserve the original data’s mathematical and statistical properties, ensuring that the AI model trained on them can be generalized to real-world data [152]. In addition, given the unrestrained availability of synthetic data (only limited by the computational power of data generation), AI models trained on synthetic data can be robust with regard to data variations [153]. Synthetic data of various types, such as tabular, text, and image, have been generated in massive quantities to train ML and DL models cost-effectively. Obesity-related data or, more generally, health-related data can be expensive to collect (eg, MRI scans) and contain confidential information (eg, patients’ names or residential addresses), which could be addressed by synthetic data generation [154].

Human-in-the-Loop

There have been increasing concerns over AI-related data bias and ethical issues [155,156]. Fundamentally, AI models should facilitate but not replace human judgment and decision-making [157,158]. Human-in-the-loop (HITL) is an AI model that requires human interaction [159,160]. HITL ensures that algorithm biases and potentially destructive model outputs can be identified in a timely manner and corrected to prevent adverse consequences. However, such interactions between humans and machines require thoughtful designs in the data-processing pipeline, model architecture, and personnel management [159]. Data- and model-driven decision-making related to obesity, such as behavioral modifications (eg, diet or physical activity interventions) or medical treatment, can be complex [161]. AI-powered wearables and other digital health platforms can detect change in an individual’s physical activity and provide actionable information to improve health outcomes [162-164]. Mobile chemical sensors could offer timely dietary information by monitoring real-time chemical variations upon food consumption, collecting dynamic data based on an individual’s metabolic profile and environmental exposure, thus supporting dietary behavior decision-making to improve precise nutrition [165]. HITL may integrate AI model outputs with expert inputs to make informed decisions that capitalize on the strengths of both and maximize patients’ chances of health restoration and improvement [166].

Limitations of the Scoping Review and Included Studies

To our knowledge, this study is the first to systematically review AI-related methodologies adopted in the obesity literature and project trends for future technological development and applications. However, several limitations should be noted concerning this review and the included studies. As our review focused on ML and DL methods, study-specific findings (eg, the effectiveness of an intervention and estimated associations between covariates and an outcome) were not synthesized in detail. The included studies were heterogeneous in terms of hypothesis and research question, study design, population sampled, data collection method, sample size, and data quality. The analytic approach chosen was endogenous to these study-specific parameters; therefore, across-study comparisons of model performances may not be reliable. Even within the same study, conclusions about relative model performances (eg, the prediction accuracy of logistic regression vs SVM) may lack generalizability because of the interdependency between data and ML and DL algorithms. AI technologies are rapidly advancing, with innovations and breakthroughs almost daily. A review such as this one will have a short shelf life and warrant periodic updates.

Conclusions

This study reviewed the AI-related methodologies adopted in the obesity literature, particularly ML and DL models applied to tabular, image, and text data for obesity measurement, prediction, and treatment. It aimed to provide researchers and practitioners with an overview of the AI applications to obesity research, familiarize them with popular ML and DL models, and facilitate their adoption of AI applications. The review also discussed emerging trends such as multimodal and multitask AI models, synthetic data generation, and HITL, which may witness increasing applications in obesity research.