Among the 74 included studies, there were 23 (31%) studies in which the detection of gender or age was secondary to their research, and previously developed methods were used to detect the demographic information of their cohort. Of the 23 studies, 13 (57%) used open-source models, and 10 (43%) used off-the-shelf software. More details about each study are given in the subsequent sections.

Of the 13 studies that used open-source models, 3 (4%) [74,99,100] drew upon an extant model [141] that used a predictive lexicon for the multiclass classification of age or gender for their applications. None of these studies created a validation corpus to assess the performance of the system, which was originally reported as 89.9% accuracy for gender and 0.84 Pearson correlation coefficient for age. One (1%) study [118] used the same text-based model [141] and an image model [142] to determine the age and gender of their cohort. When tested against their gold-standard corpus of self-reports from profile descriptions, they found that the imaging model performed best for gender (accuracy=90%-92%), whereas textual features gave the best results for age (accuracy=60%). A total of 3 (4%) studies [78,91,114] used demographer [115,139,143] for gender predictions, with 1 (33%) study [91] evaluating the performance against a set of users who had self-reported their gender in a survey, finding an F₁-score of 0.869 for women and 0.770 for men. A total of 2 (3%) studies [61,62] used an ensemble classifier of previously developed models, with a reported accuracy of 0.83 and an F₁-score of 0.83 [122]. Two (3%) other studies [67,120] used M3 [112] to detect gender and age, with 1 (50%) study validating the performance using a manually labeled data set, finding an accuracy of 95.9% and an F₁-score of 0.957 for gender and an accuracy of 77.6% and an F₁-score of 0.731 for age. One (1%) study [56] used Deep EXpectation of apparent age [144] for age and gender detection, which reported a validation error of 3.96 years for age and an 88% accuracy for gender. One (1%) study [98] used the rOPenSci gender package, and no assessment of performance was reported.

In the 10 studies that used off-the-shelf software, Face ++ was the most common software, being used in 6 (60%) studies [63,77,88,97,109,119]. The remaining studies used DemographicsPro [59,60], Microsoft Face API [84], and RapidMiner [103].

In 4 (40%) [88,97,103,109] of the 10 studies, no validation of performance was carried out, and a further 2 (20%) studies simply reported that DemographicsPro requires 95% confidence to make an estimation [59,60]. Other studies (n=4, 40%) compared with manual annotation and identified an accuracy of 82.8% for age using Face ++ [77], 68% for strict age groups, or 83% if the age groupings were relaxed [63]. The performance for age using Microsoft Face API was measured at 0.895 Gwet agreement coefficient (AC) [84], when compared with manually labeled data sets.

For gender, the studies (2/10, 20%) that measured performance against their own gold-standard labeled set of users recorded accuracies of 94.4% [77] or 88% [63] using Face ++. Other studies (3/10, 30%) [88,97,109] reported a confidence level of 95% +0.015 or –0.015 using Face ++ for gender prediction.

Only 1 (10%) [119] of the 10 studies went beyond manual annotation to create a gold standard and used multiple search techniques to manually verify age and gender, including LinkedIn profiles, electoral roll listings, personal websites, Twitter descriptions, and Twitter profile images. In this study, Face++ accuracy for age was reported as 40.4%, and Face++ accuracy for gender was reported as 44.8% (with a valid image accuracy of 32.5% for age and 87.7% for gender), and crowdsourcing annotation accuracy for age was 60.8% and for gender was 86.4% (with valid image accuracy of 56.1% for age and 93.9% for gender).

In this review, we aimed to provide an overview of recent ML methods used to predict the gender and age of Twitter users, as these are key demographics for epidemiology. Our review indicates that both tasks have been popular, but the identification of gender has received more attention than the identification of age. However, no de facto standards for research (ie, data collection and evaluation) have emerged, resulting in a large number of heterogeneous studies that are not directly comparable. Thus, it is not straightforward to conclude where the state-of-the-art stands for these tasks.

Our review found evidence of potential bias that impacts the quality and representativeness of the data used in the studies. One prevalent source of bias lies in the data collection and labeling processes. For instance, some studies may introduce systemic bias through the use of imprecise labeling methods such as name matching for labeling Twitter users’ gender. This approach can lead to mislabeling, especially for individuals with names that are culturally diverse or androgynous and introduce inaccuracies into the training data. Another problem is the introduction of sampling bias through the use of artificially balanced data sets, creating an unrepresentative sample of the Twitter population, which, in reality, has a skewed distribution, with certain age and gender groups being more prevalent than others.

It is important to address and limit these biases because when ML models are trained on biased data, they tend to replicate and amplify these biases in their predictions [145].

The prediction of demographic information is an important task to address to fully realize the potential advantages of using social media data, such as those of using Twitter data in health-related research. In the United States, the National Institute of Health has committed to including women participants in clinical studies and including sex as a biological variable, finding that the disaggregation of data by sex will allow for sex-based comparisons of results to identify any sex-based differences. A recent review [146] found that this disaggregation in the development of ML models led to the discovery of sex-based differences that improved the model performance for sex-specific cohorts. Age is also important, as it can correlate and be a factor in the course and progression of disease [146] or the effects of medication [147]. Given the significance of this information, accurate and reproducible models must be developed. One way to ensure the reproducibility of models is for researchers to make data and codes available, including annotation guidelines. In addition to model performance, studies that create annotated corpora should report annotator agreement measures to assess the quality of the corpora. Few of the included studies made their data or code available (Multimedia Appendices 3, 4, and 6).

A particular difficulty when comparing different systems comes from a lack of a gold standard that can be used to compare the systems. Some studies created their own corpora, collecting data randomly or based on keywords relevant to their studies. Others reused data sets from prior studies or shared tasks. Although outside the scope of this review, there have been shared tasks that aim to advance research through competition, focusing on gender and age prediction. A longstanding shared task focused on author profiling was hosted at the PAN workshop of CLEF [123-129]. More recently, Social Media Mining for Health (SMM4H), 2022, included 2 tasks for age detection, releasing new annotated corpora for the tasks [148]; several researchers reported using the corpora from these shared tasks. Testing and reporting performance metrics against these publicly available data sets, without alteration, would provide a comparable metric of different approaches. However, although reusing annotated corpora provides quick access to labeled data, it does have some limitations, including data loss over time as users delete their tweets, which not only reduces the size of the data but also can result in a data imbalance in the corpus.

A summary of our recommendations to reduce some potential bias in the data and improve the classification, reproducibility, and validation of the ML methods used can be found in Figure 2.

Almost all the included studies approached the gender prediction task as a binary classification task, identifying a user as either male or female. We note that even when focusing on binary gender classification, which is the prevalent approach, the task of gender prediction on Twitter could be better characterized as a multinomial classification task: given a user account, the classifier should return male, female, or nonpersonal. The last label (nonpersonal) can account for Twitter users representing organizations or bots. Although some studies attempted to identify and exclude nonpersonal accounts as a preprocessing step, other studies developed their systems using previously annotated data sets that were exclusively labeled as male or female users or removed nonpersonal accounts during annotation before training and testing. It is unknown how well these systems would perform when extended to unseen data that may contain nonpersonal accounts.

Excluding nonpersonal accounts, the ratio of male users to female users in the training data set is also important, as it should mimic the natural distribution of Twitter users, estimated to be 31.5% female users and 68.5% male users as of January 2021 [149]. However, some authors biased their collections using unconventional methods of collection or using artificially balanced data sets. The most conventional method to collect a set of Twitter accounts is to query for any tweet mentioning functional words without semantic meaning such as of, the, or and from the Twitter API. Whereas collecting Twitter users using functional or neutral keywords, a given language, or geographic areas resulted in a male:female ratio close to the ratio naturally observed on Twitter, other choices resulted in collections with different ratios. Such changes in ratios could have improved (or deteriorated) the training of the authors’ classifiers and biased their evaluations, which did not reflect the performance of their approach on a random sample of Twitter users.

All studies treated gender as a binary determination of male or female. Although some referenced the limitation of this approach, they opted to use these designations given the need to align their data with outside resources, such as the US census or social security administration data. We note that gender, unlike biological sex, is not necessarily binary as it is a social construct and has been shown to influence a person’s use of health care, interactions, therapeutic responses, disease perceptions, and decision-making [150]; this underlies the importance of expanding the efforts of classification beyond binary to improve accuracy and avoid misinterpreting results.

The age prediction task generally had a lower performance than the gender prediction task. This was true for studies that developed their own models as well as those that used open-source or off-the-shelf software. This may be because most studies approached age prediction as a multiclass classification task. The proxies used, such as language, names, networks, or images, may have limited predictive value for age. In addition, the distribution of Twitter users means that any data set will be inherently imbalanced, providing few training examples for age groups at the tail end of the distribution. This data imbalance may lead to too few instances of the minority classes to effectively train the classifier. For classification models based on images, poor performance for age may be unsurprising given that it can be difficult for humans to discern age from a single image. In addition, photos may be subject to photo editing or enhancement or may not be a recent photograph of the user. Because of a lack of error analysis reports in the included studies, it is difficult to determine the source of the classification difficulty for age.

Performance aside, the fact that the number and range of age groups vary across studies suggests that a classification approach is not generalizable to all research applications. Identifying the exact age, rather than age groups, can generalize to applications that do not align with predefined groupings of binary or multiclass models; however, using high-precision rules to extract self-reports of exact age from the user’s profile metadata had been shown not to scale. As we worked on this study, we noted that none of the reviewed systems opted for extracting the exact age. To test the feasibility and utility of a generalizable system that extracts the exact age from a tweet in a user’s timeline using deep learning methods, separate from this study, our group developed a classification and extraction pipeline using the RoBERTa-Large model and a rule-based extraction model [151]. The system was trained and tested on 11,000 annotated tweets. The classification of tweets mentioning an age achieved an F₁-score of 0.93, and the extraction of age from these tweets achieved an F₁-score of 0.86. From a collection of 245,947 users, age was extracted for 54% using REPORTage. A shared task for the classification task ran at the SMM4H 2022 workshop, and we released the annotated data set. We did not include our approach in the scoping review, as there were no comparable systems published before the release of the exact age extraction approach as part of the SMM4H 2022 shared task.

The limitations of using names to distinguish between genders may promote bias, particularly if the names used for training do not represent the ethnic diversity of the population, and some cultures may have more unisex names than others, which cannot be used to distinguish genders. There can be a high degree of uncertainty for many users for whom gender cannot be classified by name; estimates by Sloan et al [152] are that 52% of users will be unclassified using this method. However, studies have suggested that the classifications made may be relatively accurate given that the data from UK Twitter demonstrates a high level of agreement with the UK census data [153]. Furthermore, when used alone, this heuristic may label some organization accounts, such as PAUL_BAKERY, as a person.

Relying on self-declarations may be prone to bias as well. For example, younger people are more likely to profess their age than older adults, as age may be more important to them. With respect to gender pronouns, these may be more likely to be declared by those in some occupations or age groups. Indeed, there may also be other biases to self-declarations of data based on culture, background, social class, or country of origin or residence.

Using users’ profile images for gender and age identification is challenging. Not all Twitter users provide a picture of themselves, with many opting for pictures of their pets, objects, children, scenery, or even celebrities. Identifying the gender and age of even those with pictures of themselves can be problematic if the quality of the pictures is poor, the pictures contain more than 1 face, or the pictures are not recent, particularly for predicting age. A comparison of systems using images to predict demographics [154] measured not only the accuracy in identifying age and gender but also the percentage of images in which a face could be detected, finding that only approximately 30% of Twitter users had a single detectable face.

Methods to filter out organizations in the studies included removing accounts with a large number of followers [71] or explicitly searching for organizations by matching username terms linked to economic activities, such as restaurant and hotel [49]. These methods remove accounts that do not represent a single user. However, they do not remove bots. Although one of the studies created a classifier to detect bots, the filtering of bots was limited to those identified in manual annotation, by simple heuristics, or nonexistent in many studies (Multimedia Appendix 5).

For cases where age or gender are estimated, it is necessary to conduct validation exercises whereby the data are compared with a gold-standard data set to establish accuracy levels. For example, 1 study [119] that used off-the-shelf software also created a manually annotated gold-standard data set for measuring accuracy. This study found that although the accuracy of crowdsourcing was higher than that of software, the accuracy was only approximately 60% for age. This puts into question the use of manual annotations alone as a gold standard.

The most reliable way of generating a gold standard is to obtain the information directly from the user. This may be done in the form of direct correspondence with the user, such as messaging via social media or, the other way around, requesting Twitter handles in surveys that collect demographic data. Other methods for validation, such as manual extraction, may be less rigorous. However, these methods can be improved by multiple independent annotators, using experienced teams.

External validation of the model is also a vital step to assess how the model will perform on unseen data [155,156]. In a validation on a second data set, Yang et al [117] found that performance dropped in all but 2 of their models, stressing the importance of benchmarking existing systems on a targeted corpus. This step is equally important when using existing systems, so a range of expected performances can be reported and used in any analysis of the output.

In addition to the potential biases reported earlier, predicting the age and gender of Twitter users has some potential limitations that should be considered and, when possible, addressed to limit their effects. As evidenced by the performance results of the included studies, determining the precise age or age group of Twitter users solely based on their Twitter profiles and tweet content can be challenging. Although methods to extract a user’s self-reported age can be executed with high precision [151], predicting age, especially for more specific age groups, remains a complex task. Another limitation to consider is the potential for users to misrepresent their reported age or gender, which can introduce inaccuracies and affect the reliability of predictions based on user-supplied data. This phenomenon is not unique to Twitter and has been identified in other data sources such as surveys [157,158]. Many of the included studies used self-reported data to label their training data; therefore, any potential misrepresentations could be approached as a noisy label problem. There are numerous methods that can be used to manage the effect of label noise on classification models, such as distance learning or ensemble methods [159,160]. Furthermore, it is important to effectively address potential noise and uncertainty when using the output data for secondary analysis. Statistical techniques that can handle imprecise or uncertain data, such as Bayesian inference or fuzzy logic, can be valuable in this context. Using these methods, the analysis can better account for uncertain predictions, leading to more robust and reliable results. Finally, users’ age changes over time, and their profiles may not be updated accordingly, or the age tweet may be from an earlier year and not reflect their current age. Researchers should ensure that the users’ labeled age is contemporaneous with the other data included in the prediction model. Predicting the age and gender of Twitter users provides valuable insights, and most identified limitations presented by the data can be mitigated.