A systematic literature search was conducted on clinical trials on mental health disorders involving AI techniques using 4 electronic databases: PubMed, Scopus, IEEE Xplore, and Cochrane (Table 1). Search string queries are detailed in Appendix S1 in Multimedia Appendix 1.

We specified 3 inclusion criteria for screening. Records were included if they reported a clinical trial (either interventional or observational), were related to mental health disorders, and featured an application of AI. For the final eligibility assessment, exclusion criteria were defined to constrain the review: the reported AI case is not applied for a mental health outcome (ie, applying tools to improve image quality), the record was not published in English, or the report was not published in the last 5 years to review the specific application of these techniques in clinical mental health research. These criteria were designed to evaluate the researching lines in mental health disorders in the last few years, which include the democratization of frameworks and tools for AI application.

Figure 1 shows the flow diagram of the selection process. Records from the scientific literature were identified in the 4 databases defined in Table 1. The resulting data sets were combined in a Microsoft Excel spreadsheet (.xlsx), rearranged by DOI (digital object identifier), and checked for possible erroneous entries. Duplicated records were assessed by comparing DOI names and titles of the publication. A simple code in R 4.2.0 win32 (R Foundation for Statistical Computing) was used to find and tag records whose DOI name and title were already found in the database. The results were manually reviewed to correct minor errors due to misspellings of DOI or the title in the record database. The eligibility criteria for inclusion were then manually evaluated by the title and abstract of each record, and selected records were sought for retrieval. Retrieved records were fully screened and were dismissed if they did not meet the inclusion criteria or met the exclusion criteria. Finally, data and details were extracted for included AI studies.

For the included AI applications, 11 categories and 35 data indicators are reported. These indicators were adapted from the CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies) guideline [10]. In addition, publication-related data such as author(s), title, journal, year of publication, and DOI were extracted for the analysis.

Categories were designed to evaluate the goal of the clinical trial and AI, the accessibility and quality of the development process of the data, how the AI model was designed and developed, results, and the reported discussion. Not only are the categories for data extraction designed for detailing the AI models, but also they evaluate the differences and linkages between trial design, data collection, and AI implementation. For further details, see Appendix S2 in Multimedia Appendix 1.

Risk of bias was evaluated by combining the Cochrane tool for randomized controlled trials [11] and PROBAST (Prediction Model Risk of Bias Assessment Tool) guidelines [12]. The Cochrane Handbook for Systematic Reviews of Interventions [13] accounted for the trial design and whether masking and blindness were applied or should have been. The PROBAST guidelines accounted for the suitability of the methodology for collecting the data, candidate predictors, and outcome definition for the AI model as well as how the AI model was applied and analyzed. Both guidelines were considered together to evaluate possible biased relations between trial design and AI applicability. Details of the methodology are provided in Appendix S3 in Multimedia Appendix 1.

The search identified 540 records, all published in English. Excluding 111 duplicates, 429 articles were screened according to the eligibility criteria based on the title and abstract. The screening process concluded with 241 records excluded for not meeting the inclusion criteria. Of those, 188 records were sought for retrieval, with 12 found impossible to retrieve. Thus, 176 records were assessed for eligibility and 65 were excluded due to not being a clinical trial (n=37), not related to AI (n=23), partially related to mental health (n=3), or not related to mental health (n=2). Furthermore, limitations of search queries were minimized as much as possible by manually adding a selection of AI studies that were not found in the search (n=25). Records from this selection were also screened and sought for retrieval and eventually 18 studies were included. Ultimately, 129 records were included in the analysis. A record could involve 2 or more different cases of AI use for a different outcome, from now on referred to as an AI study. A total of 153 AI studies or AI applications were analyzed. Table 2 summarizes the most important information extracted for this systematic review. Details on the final analysis for each study can be found in Multimedia Appendix 2 (see also [14-142]).

Most used private data (n=142), and a small fraction used public data (n=10) or a mix (n=1). Most studies aimed to develop a new model (n=152); only 1 study aimed to validate a current model. No AI study was intended to update a previously developed AI model. Concerning mental health categories based on the International Classification of Diseases 11th Revision (ICD-11), nearly one-half of the studies (77/153, 50.3%) related to mood disorders, which combines the Depressive disorders (n=70) and Bipolar or related disorders (n=4) categories; 3 other studies used data from patients within both categories, labeled as “mood episodes.” The second most common category was Schizophrenia or other primary psychotic disorders (n=26), and the third was Disorders specifically associated with stress (n=12). Some studies included participants with different mental disorders (n=10).

Other categories were Anxiety or fear-related disorders (n=7); Secondary mental or behavioral syndromes associated with disorders or diseases classified elsewhere (n=5); Disorders due to substance use (n=5); Neurocognitive and dementia disorders (n=4); Neurodevelopmental disorders (n=1); Obsessive-compulsive or related disorders (n=1); Feeding or eating disorders (n=1); Bodily distress disorder (n=1); Personality disorders (n=1); and Mental or behavioral disorders associated with pregnancy, childbirth, or the puerperium, without psychotic symptoms (n=1).

Only 28.1% (43/153) of studies used original data collected within the study, while 71.9% (110/153) of studies retrieved data from databases or were a secondary analysis of clinical trials not designed for that purpose. The most common type of trial design was randomized clinical trial (n=62), followed by prospective cohort study designs (n=30) and nonrandomized clinical trial designs (n=15). Further, we found longitudinal naturalistic studies (n=15), cross-sectional designs (n=14), case-control designs (n=9), and case reports (n=2). Two reports of AI cases used a mix of trial designs and 4 did not report this or the references were unclear. Figure 2 shows the distribution of study design based on the prospective or retrospective collection of data.

Not all studies reported enough details to evaluate recruitment of participants (n=17). Almost one-half of the studies collected data from different locations (n=75), whereas the rest only reported 1 location (n=61). Of the multisite studies, only one-third used international collection (17/153, 11.1%). Only 13 of the 43 (30%) prospective collection studies followed a multisite collection method (n=13), and only 1 study was international.

Studies were categorized according to the intended use of the AI models in the research. The most common category was studies for evaluating treatments, Treatment quality (n=44), followed by Subgroups/patterns identification (n=31), Predictor identification (n=28), Prognosis (n=23), Diagnose (n=20), and Forecasting symptoms (n=7).

Most Treatment quality applications used retrospective data from previous randomized clinical trials (n=28) and in these studies the clinical arms were treated as different cohorts to compare AI outcomes and performance (n=28). The same results can be found in Predictor identification, where close to one-half of the studies collected data from previous randomized clinical trials to compare different clinical arms (n=13). In the Subgroups/patterns identification category, studies collected data from a balanced mix of study designs, while in Prognosis the most common method was prospective cohort studies. In the Diagnose and Forecasting symptoms studies, none of the categories stood out in particular. More detailed results are presented in Appendix S4 in Multimedia Appendix 1.

Figure 3 presents a dashboard that summarizes the AI model results regarding candidate predictors, preprocessing pipeline, AI techniques, and validation. For candidate predictors, many studies used a combination of data (n=73). The most individually common category was Medical image (n=33), which relates to medical imaging analysis (ie, region of interest or voxel-based morphometry), and the second was Questionnaires and scales (n=20), defined as any self-reported or interview-reported scale for symptom severity, conditions, or actual mood. The third was Biosignal collection (n=11), such as electroencephalography or electrocardiography and related analyses. Other data categories were Biomarkers (n=5), Genomic data (n=3), Electronic health records or I (n=2), Text (n=2), Video image analysis (n=2), and Audio recording (n=2). I refers to historical, demographic, and clinical information collected in hospitals and specialty care sites. Text refers to any data that are used for natural language processing analysis, such as written text or speech. Audio recording was introduced as the analysis of audio and voice features unrelated to language processing. The Mixed category (n=73 studies) combined data from I and different questionnaires and symptom scales (49/73, 67%); the remaining studies included other categories such as Biomarkers (n=7), Medical image (n=4), Genomic data (n=3), Biosignal (n=3), and Text (n=2). Medical image was also combined with Genomic data (n=1) and Biomarkers (n=1). Besides, Biomarkers were combined with Questionnaires and scales (n=2) and with Biosignals (n=1).

When evaluating data quality, methods to assess data suitability, and preprocessing pipelines, only 12/153 studies (7.8%) considered the statistical power of the sample size; 37.3% (57/153) of studies used a sample size of 150 or less to train the AI models. Only 13.7% (21/153) reported external validation (n=5) or reported both internal and external validation (n=16). The rest of the AI studies used only internal validation (n=108) or did not report the validation method (n=24). Only 38.6% (59/153) of studies reported a method to assess significance of their performance results, while the majority did not detail any (n=94).

AI studies used supervised learning, semisupervised learning, and unsupervised learning methods. No reinforcement learning algorithms were found. Regarding AI algorithms, the most popular family of techniques was regression (n=34), followed by trees (n=26) and support vector machine (n=23), which constitutes most AI studies. Other algorithms were Bayesian (n=6), statistical clustering (n=5), hierarchical clustering (n=5), mixture model (n=3), deep learning (n=1), manifold (n=1), Markov model (n=1), and relevance vector machine (n=1). In some cases, an Ensembled model was designed with the inclusion of different types of AI algorithms (n=12). Another category, Competing models (n=35), refers to the AI studies that did not predefine a specific AI technique or algorithm based on features of the data and instead applied different techniques with the intention to retain the model algorithm with the best performance to their outcome definition. These 35 studies used 144 AI techniques in total.

Regarding preprocessing methods, only 63.4% (97/153) of studies reported whether they applied any preprocessing technique to data or that preprocessing was not needed, while the rest did not report any (n=56). Regarding data gaps, only 52.3% (80/153) of studies reported or mentioned if there were some missing data in samples or not, while 47.7% (73/153) did not. Of the studies reporting missing data, 2.6% (4/153) did not report any method to handle missing data bias, whereas 24.2% (37/153) opted for excluding the samples and 25.5% (39/153) chose to impute the missing values from the data distribution by different imputation methods. Of these, only 2 studies detailed the type of missingness. The proportion of reporting missing data was similar for both retrospective and prospective data collection.

The risk of introducing bias is defined in Appendix S3 in Multimedia Appendix 1. Only 23 studies were found unlikely to introduce bias due to the trial design and evaluation of participants, whereas in the majority of studies the risk was high or unclear. In most cases, the risk of bias due to participants and the trial features was introduced by bias in the distribution of participants—that is, inclusion and exclusion criteria, loss of follow-up, and participants withdrawal—or the sample was not enough to be considered a good example of the target population. The definition and collection of the candidate predictors were mostly a low risk of introducing bias (n=16), with a few studies possibly introducing bias (n=21) or having an unclear risk of bias (n=16). Results for the outcome definition in the AI model are similar, with most studies evaluated as having a low risk of bias (n=101). Some studies were categorized as unclear (n=15) or high risk (n=37) due to unclear definitions of outcomes or combining data set of different populations whose outcomes were evaluated with different methods. The most important risk of bias was found when applying AI algorithms and their evaluation. Only a few studies were evaluated as unlikely to introduce some bias (n=5) and the vast majority of AI analysis introduced a high risk of bias (n=139), while 9 could not be assessed properly (n=9). The main issues for bias in the AI analysis were not appropriately preprocessing and arranging the data for the specifications of the applied AI model, a bad handling of missingness, or an insufficient validation of the performance to account for overfitting and optimism (Figure 4). Appendix S5 in Multimedia Appendix 1 shows a stratified analysis of the risk of bias based on disorders, study designs, and outcome.

Overall, only 1 AI study could be assessed as a low risk of being biased. The most contributing categories to the overall risk were Participants and AI analysis. Most studies were likely unbiased about the definition and collection of predictors and the outcome but they failed to apply these data later in the model—bad data engineering or bad validation of the models—or the trial design had some flaws from the beginning. It is worth mentioning that only 9 of the 153 models reported any hyperparameter tuning or coefficients of the models and most of them reported basic trees models coefficients and decision rules. Only 58 studies mentioned or reported predictor importance and less than one-half reported the ranking and evaluated the methodology to test it (Figure 5).

This overview summarizes the development and status of AI applications in mental health care. The review focused on the period from 2016 to 2021 to understand the latest advancements of these applications in mental health research, including aspects related to the methodological quality, the risk of bias, and the transparency. Results may be limited by the keywords applied in the search queries. This analysis is only representative for the records retrieved in this search. However, the samples analyzed may be sufficient to judge the quality and current status of this field. Significant methodological flaws were found involving the statistical processes of AI applications and data preprocessing. Many studies overlook evaluating or reporting on data quality and how it fits the model. Some studies applied several AI techniques, here aggregated as “compelling models” studies, to select the most efficient technique without assessing their suitability for the problem they face, which may lead to overoptimism. Further, missing data management and the recruitment of participants are rarely reported, making it difficult to account for the risk of model overfitting. Preprocessing pipelines are not sufficiently reported, which hampers the reproducibility of the study or the adaptation of the AI techniques to the specific type of study. The use of reporting guidelines, such as the CONSORT-AI (AI version of the Consolidated Standards of Reporting Trials) for clinical trial reports involving AI [143], the SPIRIT-AI (AI version of the Standard Protocol Items: Recommendations for Interventional Trials) for clinical trial protocols involving AI [144], or the MI-CLAIM (Minimum Information About Clinical Artificial Intelligence Modeling) checklist on minimum information for clinical AI modeling [145], would be very useful to ensure that the basic information about the study design and implementation is reported in these types of studies.

Some predictive models are being updated for AI reporting, such as TRIPOD-AI (AI version of the Transparent Reporting of a Multivariable Prediction Model of Individual Prognosis or Diagnosis) and PROBAST-AI (ie, the AI version of the PROBAST) [146]. However, these guidelines are rarely followed in the reviewed records. They lack transparent reporting on AI model features such as coefficients, hyperparameters, or predictor selection. Encouraging transparent reporting should be prioritized, as it would benefit second external validations and provide better accountability for reported models.

The incorporation of AI in mental health research is unbalanced between ICD-11 mental health disorders. Most research focusses on depressive disorders, where it combines severity questionnaires and scales with electronic health record and psychotic disorders using medical image data. External validation is very uncommon. Conducting suitable trial designs for the intended AI outcomes is understandably difficult in terms of money, time, and resources. Thus, it is common to apply data collected retrospectively. However, the original trial designs do not fit the specifications for AI development and most studies do not assess the appropriateness of these data. Notably, many authors may not understand the need to ensure an optimal preprocessing pipeline. In these cases, authors are aware of the poor performance of the models, but the proposed approach for improvement is suggested directly from a trial perspective rather than from assessing possible statistical bias or mistakes in model development, which could save cost and time over designing new studies.

AI studies were analyzed to identify challenges and opportunities involving the use of AI in mental health. Typically, AI studies reported insufficient samples to ensure model generalizability [68,84,103]. Several authors reported bias because of the difficulty in adapting typical trial designs to an AI context. For example, some authors detail the constraining boundaries for selecting participants in randomized clinical trials as a limiting factor, which reduces the sample size and could overlook confounders [68,90]. Most randomized clinical trials noted possible variance between the collected data and the real-world data. However, observational studies can also introduce bias in AI models if the imbalance between cohorts is not adequately addressed [84,128]. In these studies, the variety in features such as prescribed medication could introduce confounders and bias that are difficult to manage [94]. Furthermore, in long-term studies, lack of follow-up or other conditions leading to a decrease of patients is an important limitation, mostly for prognosis studies or predictive evolution of condition severity [30,58]. These issues are worse for retrospective collection of data, where trial designs tend to diverge from the problematics of AI. Besides, some authors are aware of bias due to gaps, but most did not properly evaluate this risk.

A noticeable lack of internationalization was detected. Many studies focused only on local data, which contributes to small sample sizes and poor generalizability [115,127]. Encouraging partnerships and collaborations across countries and centers should be a priority, as it could facilitate external validation [71]. Some authors mention difficulties reconciling clinical practices with AI study requirements, usually due to ethics problems related to clinical practice in patients that can overlook confounders, that is, making it difficult to apply placebo controls in some interventions [82,115].

Another challenge is the explainability of complex AI models, which could make researchers reluctant to adopt techniques that map high-order interactions or “black-box” algorithms [81,122]. Researchers prefer simpler algorithms. The few studies that reported model coefficients and some explanation used decision trees. Another challenge is that contradictory findings could occur among studies [85].

Finally, some authors are aware of the opportunities that everyday devices and platforms such as phones and social networks offer but find it difficult to take advantage of these tools due to lack of standardization, which reduces the target population for defining a study [92].

Some studies introduced devices and platforms to improve the monitoring of patients. The application of everyday digital tools could reduce necessary resources and therefore facilitate data collection [99,127]. Promoting the use of frequently used devices combined with the application of AI models seems like a future trend that could improve the treatment of many conditions where the chance of treatment response decreases over time [126]. Further, it opens possibilities of internet-based treatments that could be conducted in real time with digital technology, easing the load on hospitals [99].

Data sharing and public databases should be encouraged to develop and implement more trustworthy AI models. AI models from clinical stage to clinical practice could be difficult but powerful tools to gain insights into predictor collection, human-based decisions, and AI biases while these techniques are being implemented in clinical world. Many studies report the high potential of AI in mental health for clinical support, computer-aided systems, and possibly preliminary screening [94,127].

Currently, many guidelines and initiatives exist to which researchers could adhere to in order to increase transparency and better use AI models. Currently, the EQUATOR (Enhancing the Quality and Transparency of Health Research) network initiative reports useful guidelines that could foster collaboration and implementation [147].

AI algorithms are increasingly being incorporated into mental health research; however, it is still uneven between ICD-11 categories. Collaboration is merely anecdotal, and data and developed models mostly remain private. Significant methodological flaws exist involving the statistical process of AI applications and data preprocessing pipelines. Only 1 study was found reporting second validation, and 13.7% (21/153) reported external validation. The evaluation of the risk of bias and transparent reporting was discouraging. Model hyperparameters or trained coefficients are rarely reported, nor are insights about the explainability of the AI models. The lack of transparency and methodological flaws are concerning, as they delay the safe, practical implementation of AI. Furthermore, data engineering for AI models seems to be overlooked or misunderstood, and data are often not adequately managed. These significant shortcomings may indicate overly accelerated promotion of new AI models without pausing to assess their real-world viability.