Search data were collected for 108 search terms (Multimedia Appendix 1) from Google application programming interfaces. The search terms included terms related to chronic diseases, risk factors, diet, and physical activity. To generate a comprehensive list of terms, we used the Google Trends website [44] to identify terms that had similar search trends for chronic diseases and their associated risk factors. We collected the yearly search volume for each country from 2010 to 2016 for 52 countries in English [45]. Google normalizes the search volume for each term relative to the search activity in the country and the specific time period. Two countries (South Sudan and Sudan) were excluded because obesity prevalence estimates were unavailable for these countries.

We also downloaded age-standardized obesity and overweight prevalence estimates for adults aged 18 years and older from 2010 to 2016 from the World Health Organization (WHO) website [46,47]. These estimates were obtained using data from population-based studies on cardiometabolic risk factors, multicountry and national measurement surveys, as well as the WHO STEPwise approach to surveillance (STEPS) surveys for estimating BMI [48]. Overweight was defined as a BMI >25 kg/m² and obese was defined as a BMI ≥30 kg/m² [49]. The reported credible intervals (CIs) for the estimates represented the 2.5th and 97.5th percentiles of the posterior distributions.

We used machine learning methods to identify search patterns that were associated with changes in obesity and overweight prevalence across African countries. Specifically, we employed support vector machine (SVM), random forest (RF), gradient boosting, and Bayes generalized linear model (GLM). The machine learning methods were selected to assess a broad range of approaches from decision tree methods, kernel-based approaches, and least squares regression methods. We implemented these methods using the SuperLearner package in R [50,51], which generates estimates for each individual method and an ensemble of the methods.

RF regression is an extension of bootstrap aggregating (“bagging”). It involves the construction of de-correlated decision trees, which are averaged to reduce the variance of the prediction function. Trees are preferred candidates for bagging because they capture the complex interaction structures in the data and have relatively low bias if grown deep. Since each generated tree in bagging is identically distributed, the average of B such trees is the same as the likelihood of any one of the trees. The gradient boosting algorithm also involves the generation of ensembles of predictive trees. However, trees are built using the gradient boosting approach, which involves a sequential iterative fitting procedure to reduce bias by assigning higher weights to poorly fit samples and optimization via a loss function. An advantage of the gradient boosting algorithm is that nonlinearities and interactions do not need to be explicitly specified.

In contrast, SVM regression is similar to multiple linear regression when the relationship between X and y is linear:  y = ƒ(x)  = W · X + b. However, SVM regression involves the application of kernel functions (eg, gaussian, polynomial, radial basis, and sigmoid kernel) to model nonlinearity between X and y. The SVM regression model parameters are selected to minimize an epsilon-insensitive cost function. The model parameters were selected by applying cross-validation to the training data.

Lastly, Bayes GLMs are a class of GLMs that are a generalization of linear regression models such that the distribution of the dependent variable is of the exponential family (eg, gaussian, poisson, binomial, categorical, multinomial, or beta). In the Bayesian approach, inferences are based on the posterior distribution, prior knowledge is captured quantitatively through the prior distribution, and the data are represented through the likelihood function [52,53]. Two advantages of Bayesian models include the incorporation of domain knowledge via the prior and uncertainty quantification via the posterior distribution.

First, we estimated the Pearson correlation coefficient (r) between the search data and obesity and overweight prevalence across Africa from 2010 to 2016. Next, we excluded all search terms that had zero variance (ie, 20 search terms) and search terms not significantly correlated with obesity/overweight prevalence at a significance level of P<.05. Additionally, because there were zero reported searches for some terms in some countries, we excluded all terms with less than 50% of observations greater than zero, implying that only the most significant and comprehensive variables were used in the modeling. We then fitted separate models to estimate obesity and overweight prevalence using the search data. The coefficient of determination (R²) and root-mean-square error (RMSE) were used to assess the model fit. The out-of-sample estimation involved splitting the data into 2 sets: data from 2010 to 2014 were used to train the model, while data from 2015 to 2016 were used to evaluate the model. In machine learning, the data used to train the model are usually different from the data used to validate it. The training data are used to fit the model (ie, train the algorithm to identify patterns) and the evaluation data are used to assess the predictive performance of the fitted model by comparing the model estimates to true values. The aim is to allow the model to be generalizable to future sets of data. However, in the absence of future data, the evaluation data are used. We also report the correlation between the out-of-sample predictions and WHO-estimated obesity and overweight prevalence. The following R packages were used: SuperLearner, randomForest, kernlab, and arm [51,54].

Some countries had sparse or no data for some of the search terms. Search patterns were similar for several of the terms: lose weight and weight (r=0.93, 95% CI 0.91-0.94), diet and weight (r=0.92, 95% CI 0.90-0.93), diet and weight loss (r=0.89, 95% CI 0.87-0.91), food and weight (r=0.88, 95% CI 0.85-0.90), food and weight loss (r=0.86, 95% CI 0.83-0.88), breakfast and diet (r=0.85, 95% CI 0.82-0.87), weight and ginger (r=0.84, 95% CI 0.81-0.87), weight and breakfast (r=0.83, 95% CI 0.80-0.86), weight loss and weight gain (r=0.83, 95% CI 0.79-0.86), exercise and food (r=0.81, 95% CI 0.77-0.84), ginger and weight loss (r=0.81, 95% CI 0.77-0.84), weight loss and fasting (r=0.81, 95% CI 0.77-0.84), gym and diet (r=0.81, 95% CI 0.77-0.84), lose weight and food (r=0.81, 95% CI 0.77-0.84), lose weight and gym (r=0.81, 95% CI 0.77-0.84), and food and ginger (r=0.80, 95% CI 0.75-0.83). Most of these associations were between terms that capture the same underlying intention. For instance, someone searching for information on how to lose weight might also search for gym, diet, or weight loss plans.

Estimated obesity prevalence was lowest for Ethiopia and highest for Libya during the study period (Figure 1). Obesity prevalence was most statistically significantly correlated with similar and different search terms across the countries with highest obesity and overweight prevalence (Figure 2). For example, for Libya, statistically significant correlations were observed between obesity prevalence and searches for yoga (r=0.95, 95% CI 0.71-0.99), exercise (r=0.89, 95% CI 0.43-0.98), and gym (r=0.91, 95% CI 0.49-0.99). Similarly, for Egypt, significant correlations were observed between obesity prevalence and searches for gym (r=0.98, 95% CI 0.83-0.99), breakfast (r=0.96, 95% CI 0.73-0.99), and yoga (r=0.95, 95% CI 0.67-0.99). In contrast, significant correlations for South Africa were between obesity prevalence and searches for how to exercise (r=0.99, 95% CI 0.91-0.99), green tea (r=0.98, 95% CI 0.89-0.99), and weight gain (r=0.97, 95% CI 0.83-0.99). For Algeria, we observed significant correlations between obesity prevalence and searches for gym (r=0.93, 95% CI 0.58-0.99), yoga (r=0.92, 95% CI 0.54-0.99), and weight (r=0.89, 95% CI 0.44-0.98). Searches for Fitbit were significantly associated with obesity prevalence in some countries (eg, Egypt and Algeria); however, the search volume was much lower than the search volume of other terms listed, suggesting less interest. Findings were similar between overweight prevalence and the search terms.

Twelve of the terms that were significantly correlated with obesity prevalence (ie, hypertension, breakfast, diet, nutrition, obese, green tea, weight gain, lose weight, weight loss, weight, gym, and malnutrition) were used in modeling to estimate obesity prevalence. The estimated variances explained by the various models were 0.97, 0.92, 0.77, and 0.30 for RF (Figure 3), gradient boosting, SVM, and Bayes GLM, respectively; the corresponding RMSEs were 1.15, 1.87, 3.53, and 5.60, respectively. Likewise, the correlations between the out-of-sample estimates (ie, data not used to train the model) and obesity prevalence were 0.96, 0.94, 0.87, and 0.56 for RF, gradient boosting, SVM, and Bayes GLM, respectively.

Similarly, 8 search terms (hypertension, breakfast, diet, nutrition, obese, lose weight, gym, and malnutrition) were used in modeling to estimate overweight prevalence. The RF model was also the best performing model for estimating overweight prevalence (Figure 4). The estimated variances explained by the various models were 0.96 (RMSE 2.26), 0.91 (RMSE 3.56), 0.62 (RMSE 7.72), and 0.23 (RMSE 9.99) for RF, gradient boosting, SVM, and Bayes GLM, respectively; the corresponding correlations between the out-of-sample model estimates and overweight prevalence were 0.95, 0.94, 0.78, and 0.49, respectively.