A systematic review of the use of artificial intelligence in mental health–based diabetes care: Current applications and future directions

Abstract

Aims

To map and systematise existing research on the use of artificial intelligence (AI) in mental health–based diabetes care contexts, identify trends and potential gaps in the literature, examine methodological limitations and highlight future research directions.

Methods

The review adhered to PRISMA guidelines and was pre‐registered on PROSPERO (CRD420251167053). Comprehensive searches were conducted across nine databases, including MEDLINE, EMBASE, PsycINFO and IEEE Xplore, using a Boolean strategy that combined terms related to diabetes, AI and clinical mental health. Inclusion criteria encompassed peer‐reviewed, empirical, quantitative studies involving humans, diabetes contexts, mental health factors and AI‐based methodologies. Screening and data extraction were performed independently by two reviewers. Forty‐one studies ultimately met the inclusion criteria.

Results

Research on AI in mental health–based diabetes care contexts has grown substantially since 2020. Most studies employed observational (83%) and cross‐sectional (56%) designs, focused on assessment rather than intervention (88%) and targeted depression (56%). Supervised learning algorithms were most frequently used (83%); however, deep learning models achieved the highest performance. Despite technological advances, no temporal improvement in algorithmic performance was observed. Methodological limitations included limited diversity in samples and outcomes, minimal use of prospective experimental and randomised controlled trial‐based designs and overreliance on supervised learning algorithms.

Conclusions

AI shows promise in addressing mental health needs in diabetes care. However, current research is narrow in scope and lacks methodological rigour in some respects. Future studies might profitably prioritise diverse populations, prospective designs, interpretability and clinical utility to enable safe, effective and equitable integration of AI into person‐centred diabetes care.

What's new?

• Diabetes significantly impacts mental health (MH) and vice versa.

• Many healthcare systems struggle to meet the needs of people with diabetes. AI technologies show promise, but their role in MH‐based diabetes care remains underexplored.

• This is the first systematic review on artificial intelligence (AI) in MH‐diabetes contexts. Findings reveal rapid research growth post‐2020, dominance of assessment‐focused studies, reliance on supervised learning models and superior performance of deep learning algorithms.

• AI can bridge gaps in MH‐based diabetes care, but limitations need to be considered. Future research should prioritise a wider range of populations and outcomes, ethical design, prospective experimental designs and clinical actionability.

1. INTRODUCTION

1.1. Background

Diabetes affects an estimated 589 million people globally, ¹ and can have significant impacts on emotional well‐being. ² Understandably, many people living with diabetes experience psychological challenges, such as depression, anxiety and diabetes distress, which can negatively impact self‐management and glycaemic outcomes. ³ For example, people with diabetes are twice as likely to have depression relative to people without. ⁴ Clinically, mental health needs are often deprioritised rather than viewed as interconnected within diabetes care. ⁵

Regardless of diabetes type, individuals face psychological demands related to daily diabetes care, regular monitoring of glycaemic outcomes and complications, and the need for coordinated care across multiple specialties. ⁶ However, uneven resource distribution, limited availability of qualified professionals and insufficient primary care capacity create challenges. ⁶ Over half of individuals with diabetes do not receive adequate care, and many remain undiagnosed for years, ⁶ , ⁷ with both the prevalence and the proportion of untreated cases continuing to rise over time, ⁷ , ⁸ particularly in Low‐ and Middle‐Income Countries (LMICs). ⁹ For example, in Nicaragua, for every person diagnosed with diabetes, over eight remain undiagnosed, and in Mozambique, in a population‐based survey screening for diabetes, of those found to have diabetes when screened, only 13.3% were aware of their condition. ⁹

With diabetes care accounting for 12% of the world's health expenditures, ¹⁰ these systemic challenges are further compounded by growing demand for psychosocial services and a lack of ability to meet them, ¹¹ long waitlists, ¹² fragmented or complex service delivery pathways, ¹³ , ¹⁴ non‐uniformity of access, ¹⁴ , ¹⁵ financial costs, ¹⁶ and a shortage of mental health professionals. ¹⁷ , ¹⁸ Challenges in addressing psychosocial burdens associated with diabetes care tend to be particularly pronounced in LMICs and are further exacerbated by limited psychosocial research in these countries and limited mental health and diabetes screening in low‐resource areas. ⁴ , ⁹ The advent of artificial intelligence (AI)–based technologies can offer promising solutions to these growing challenges within healthcare systems. ¹⁹

Whilst there is no general legal definition for what constitutes ‘artificial intelligence’ outside of a specific application, ²⁰ AI is a term that broadly refers to technology designed to perform activities in a manner that emulates human intelligence. ²¹ , ²² The use of AI‐driven technologies can offer several advantages over traditional forms of healthcare provision, including prompt and efficient data‐driven insights, ²³ prediction and early detection of issues, ⁶ , ²⁴ , ²⁵ ongoing monitoring, ⁶ , ²⁴ reduced stigma, ¹⁹ , ²³ prompt creation of personalised treatment plans, ²³ scalability, ²³ resource efficiency, ²³ reduced practitioner burnout ²³ and reduced access barriers. ¹⁹ AI technologies may also incorporate customisation of mannerisms, dialect and other characteristics to enhance rapport between the virtual clinician and individuals living with diabetes, ²⁶ and consequently further reduce mental health support‐related stigma.

Hence, as AI‐based tools become more sophisticated, evidence‐based, safe and accessible, they hold the potential to bridge critical gaps in mental health care, particularly in resource‐limited settings, LMICs, or amongst people living with diabetes who have limited access to psychological services.

Despite (a) the rising prevalence of diabetes (particularly in LMICs) ⁷ , ⁹ ; (b) the health and economic burden of diabetes ¹⁰ ; (c) the importance of mental health factors in exacerbating (or potentially improving) diabetes management ² , ²⁷ ; (d) the documented limitations of current healthcare systems in meeting the integrated psychological and physical health needs of people living with diabetes (especially in LMICs) ⁴ , ⁹ , ¹¹ , ¹² ; and (e) the potential of AI to minimise and mitigate these challenges, ¹⁹ there have been no attempts at systematically synthesising the studies focusing on the use of AI in mental health–based diabetes care contexts.

More specifically, a preliminary search of MEDLINE, PROSPERO, the Cochrane Database of Systematic Reviews and JBI Evidence Synthesis revealed no current or ongoing systematic reviews or scoping reviews on the topic, as of 15 October 2025. A lack of published or registered reviews on the subject further underscores the novelty and necessity of this work.

A lack of review, systematic analysis, or regulation of AI‐based psychosocial tools is problematic and poses a range of ethical and legal issues. ²⁰ Some of these issues include privacy concerns, ²⁸ , ²⁹ concerns regarding risk and associated safety management, ²⁸ , ²⁹ , ³⁰ need for ongoing human oversight (which may not always be feasible), ³¹ absence of ethical consideration during design or implementation process, ²⁸ , ³⁰ lack of a defined duty of care, ³² lack of transparency, ²⁹ lack of balance between technological advancement and human‐centric care, ¹⁹ , ²⁹ and lack of frameworks for their regulation, accountability and transparency. ¹⁹ , ²⁸ , ²⁹ , ³⁰ , ³² Many of these issues can be attributed to the rapid development of AI technologies, ²⁸ and the creation of AI‐based tools within siloed professional communities, which tend to involve minimal interactions between computer scientists and healthcare professionals. ³³

The following systematic review was thus designed to: (a) map and systematise existing research on the use of AI in mental health–based diabetes care contexts; (b) identify trends and potential gaps in the literature; (c) examine methodological limitations and propose strategies to mitigate them; and (d) highlight future research directions.

A systematic review methodology was selected for the study, as it provides the most comprehensive, rigorous, critical and reliable appraisal of the evidence in relation to the research aims, compared with other types of reviews (except meta‐analyses). ³⁴ , ³⁵ It does so via defined inclusion and exclusion criteria, systematic data extraction and synthesis, quality assessment, enhanced statistical power and greater generalisability of results, critical appraisal of the included studies, and mitigation of publication bias (through the inclusion of studies with positive and negative results). ³⁴ Due to the heterogeneity of the studies (in terms of their methodology and outcomes) included in the review, a meta‐analytic methodology could not be employed.

2. METHODS

The review was conducted in accordance with PRISMA guidelines, ³⁶ and was pre‐registered on PROSPERO (CRD420251167053). ³⁷ The search strategy was devised in consultation with senior academics and conducted on 13 November 2025, utilising the Cochrane Library (including CENTRAL), AMED, MEDLINE, EMBASE, APA, PsycINFO, Ovid Journals, IEEE Xplore and Scopus. The search consisted of the following Boolean string:

(‘diabetes’ and (‘artificial intelligence’ or ‘AI’ or ‘machine learning’ or ‘deep learning’ or ‘neural network’ or ‘natural language processing’ or ‘chatbot’) and (‘mental health’ or ‘depress*’ or ‘anxi*’ or ‘eating disorder*’ or ‘disordered eating’ or ‘quality of life’ or ‘diabetes distress’ or ‘burnout’ or ‘fear of hypoglycemia’ or ‘needle phobia’ or ‘stress’ or ‘psychosocial’)).

Inclusion criteria included studies written in English, peer‐reviewed, empirical, quantitative, focusing on humans, related to a diabetes context, examining clinical mental health factors and involving the use of AI. Conversely, the exclusion criteria includes studies that were not empirical journal articles (e.g., reviews, book chapters, conference proceedings, opinion pieces and corrigenda), used qualitative methodologies, used technologies and processes that are considered para‐AI but not AI (e.g., Structural Equation Modelling), had addictions‐based outcomes, and focused on psychosocial outcomes beyond mental health indices (e.g., treatment adherence, quality of life, social support and stigma). Studies focusing on alcohol or other drug addictions as outcomes were excluded due to the addictions sector being a niche area of clinical practice, having a relatively distinct literature base, and unique aims and challenges. ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴²

The de‐duplication and screening were conducted via Rayyan, an online screening platform. ⁴³ Identified records were manually de‐duplicated via Rayyan (n = 2695), and two coders (VD and AB) independently screened the articles based on their titles and abstracts. Of the 2503 screened studies, there were conflicts regarding 34 studies (i.e., a 1.36% disagreement rate). The conflicts were primarily in relation to (a) ambiguity regarding what constitutes a ‘clinical mental health factor’ (e.g., stress, and cognitive function or impairment); (b) ambiguity regarding the inclusion of studies using well‐being aggregate measures that include clinical mental health factors as one of many variables; and (c) conference abstracts being mistaken as journal articles. The conflict resolution process was unbiased, whereby 15 decisions were made in favour of one co‐screener and 11 in favour of another. The remaining 8 studies, which primarily related to ambiguity regarding the inclusion of studies having diabetes‐related factors as outcomes instead of predictors, were reviewed by the supervisory team (AS and NC), who recommended the exclusion of studies having non‐diabetes or mental health factors‐related outcomes (n = 3), but inclusion of studies having diabetes‐related outcomes (n = 5). Of the articles assessed as meeting the inclusion criteria (n = 42), one was removed by its publisher ‘at the request of the Editor‐in‐Chief due to an issue relating to the ownership of the data’. ⁴⁴ Ultimately, 41 articles were included in the review (see Figure 1).

FIGURE 1.

Flowchart of study identification, screening, eligibility and inclusion process, based on PRISMA 2020 ³³ guidelines.

Data were extracted from the included articles by the primary reviewer (VD) and verified by the second reviewer (AB). Trends across these articles were analysed by coding and examining common variables using Microsoft Excel.

3. RESULTS

A summary of the findings has been provided in a table in the Supporting Information. The data extracted from the reviewed studies included year of publication, sample size, origin and type (e.g., adults or children), demographics, type of diabetes (e.g., Type 1, Type 2 or preclinical), AI algorithm type (e.g., random forest, neural network, support vector machine, etc.), study methodology (quantitative or mixed‐methods), type (assessment or intervention) and design (e.g., cross‐sectional or longitudinal), mental health component (e.g., depression, anxiety and diabetes stress), outcome type (i.e., primarily related to diabetes or mental health) and the highest AUC (Area Under the Receiver Operating Characteristics Curve) values. AUC was selected as the primary performance and comparison metric due to it being the preferred validation measure in machine learning research, particularly over accuracy, error rate, or error cost. ⁴⁵

3.1. Year of publication

No trend was observed in the number of studies evaluating the use of AI in mental health related to diabetes contexts over time, prior to 2020. Post‐2020, however, the number of published studies on the topic increased rapidly over time (see Figure 2), with one study published in 2020 and 14 published in 2025 (as of 13 November 2025). No studies published prior to 2013 met the inclusion criteria.

FIGURE 2.

Number of included peer‐reviewed studies focusing on the use of artificial intelligence (AI) in mental health–based diabetes contexts over time.

3.2. Sample

No trends were observed in the average sample sizes of the reviewed studies across time, with the average sample size being 26 in 2013, 1323 in 2015, 0 in 2016, 2017 and 2018 (due to no studies in those years meeting the inclusion criteria), 485 in 2019, 1,910,674 in 2020, 50,578 in 2021, 13,381 in 2022, 6920 in 2023, 53,453 in 2024 and 7789 in 2025.

The overall average sample size was 65,799, with a range of 10 (in 2024) to 1,910,674 (in 2020). Twenty‐seven percent of the reviewed studies (n = 11) had a sample size of 100 or less, 20% (n = 8) had a sample size between 101 and 500, 12% (n = 5) had a sample size between 501 and 1000, and 41% (n = 17) had a sample size above 1000.

Most studies (n = 28; 68%) had an adult sample, with 7% (n = 3) of the studies having a sample consisting of young people and another 5% (n = 2) consisting of people of all ages.

Most studies included samples originating in the United States (n = 11; 27%), China (n = 7; 17%), the United Kingdom (n = 6; 15%), South Korea (n = 3; 7%), Israel (n = 2; 5%), Denmark (n = 2; 5%) and Taiwan (n = 2; 5%). The remaining studies (n = 1; 2% each) obtained samples from New Zealand, Morocco, Canada, Bangladesh, Panama, Japan, Australia and India. Hence, 15 countries were represented in the samples recruited in the reviewed studies.

3.3. Demographics

Most of the reviewed studies reported descriptive information on age (n = 37; 90%), sex or gender (n = 34; 83%), BMI (n = 14; 34%), residence (n = 12; 29%), smoking status (n = 11; 27%), educational status (n = 10; 24%), marital status (n = 10; 24%) and ethnicity (n = 10; 24%). A smaller proportion of the studies also reported information on income or financial status (n = 8; 20%), occupation (n = 7; 17%), alcohol use status (n = 2; 5%), insurance status (n = 2; 5%), number of children (n = 2; 5%), family status (n = 2; 5%), living situation (n = 2; 5%), preferred language (n = 1; 2%) and physical activity level (n = 1; 2%). Two studies (5%) did not report any demographic information.

3.4. Diabetes type

Most of the reviewed studies (n = 17; 41%) focused on both Type 1 and Type 2 diabetes, followed by those that focused exclusively on Type 2 diabetes (n = 14; 34%). A relatively limited number of studies focused exclusively on Type 1 diabetes (n = 9; 22%) and preclinical diabetes (n = 1; 2%).

3.5. AI algorithms

Most studies (n = 25; 61%) used multiple algorithms instead of a single algorithm (n = 16; 39%), and the majority (n = 34; 83%) used a type of supervised learning algorithm; for example, random forests (n = 16; 39%), regression (n = 13; 32%), support vector machine (n = 11; 27%) and XGBoost (n = 11; 27%). Some studies also used deep learning (n = 13; 32%), dimensionality reduction (n = 2; 5%) and unsupervised learning (n = 3; 7%) algorithms. None of the reviewed studies used semi‐supervised or reinforcement learning algorithms.

Of the studies for which AUC data were available, those using deep learning‐based algorithms demonstrated the best AI performance on average (average highest AUC = 0.905), followed by those using unsupervised learning algorithms (0.903), supervised learning algorithms (0.832) and dimensionality reduction algorithms (0.805), respectively. No AUC data were available to calculate the average AUCs for studies using semi‐supervised and reinforcement learning algorithms. Notably, despite being the most widely used set of AI algorithms in mental health–based diabetes care literature, supervised learning algorithms demonstrated one of the lowest performance levels relative to other types of algorithms.

3.6. Study methodology and type

Most of the reviewed studies used an observational design (n = 34; 83%), followed by a non‐randomised experimental design (n = 5; 12%), and a randomised controlled trial (RCT) design (n = 2; 5%). Additionally, most of the included studies were cross‐sectional (n = 23; 56%) rather than longitudinal (n = 19; 46%), and one study used a mixed and multi‐component design (observational and cross‐sectional, and experimental and longitudinal).

Most studies were exclusively quantitative (n = 39; 95%), whilst two studies (5%) had a mixed‐methods design (including a qualitative component). Importantly, most studies focused on assessment (n = 36; 88%), whilst the remaining studies (n = 5; 12%) focused on intervention.

3.7. Mental health components

Most studies focused on a single mental health condition (n = 24; 59%). Across studies, depression was the most common focus (n = 23; 56%), followed by diabetes distress and psychological stress (n = 13; 32%), anxiety (n = 8; 20%) and cognitive functioning or impairment (n = 4; 10%), respectively.

3.8. Primary outcomes

Most studies focused on mental health factors as primary outcomes (n = 35; 85%), with some focusing on diabetes (or related factors, such as, diabetic peripheral neuropathy) as the primary outcome (n = 5; 12%)—that is, using mental health factors to predict diabetes occurrence or progression. One study (2%) included both diabetes and mental health–related outcomes.

3.9. Performance evaluation

The average highest AUC across the reviewed studies was 0.84, with a range of 0.69 (in 2021) to 0.98 (in 2025). Despite ongoing improvements in AI‐based technologies, no improvement in AI performance was observed over time (Figure 3).

FIGURE 3.

Area under the receiver operating characteristics curve (AUC)–based comparison of artificial intelligence (AI) performance in mental health–based diabetes contexts over time.

3.10. Intervention studies

None of the intervention studies (n = 5; 12%) evaluated AI model performance using AUC. Instead, they used pre‐post changes in scores pertaining to participants' psychosocial well‐being (measured using psychometrics), HbA1c levels, psychological distress levels, psycho‐emotional adjustment, and self‐reported feasibility and acceptability of the intervention (which included a qualitative component) as measures of AI performance. The two mixed‐methods studies (5%) included in the review were both intervention studies. Notably, intervention studies tended to be less transparent in their overview of specific algorithms underlying their AI‐based applications than assessment studies. More specifically, all intervention studies referred to their AI models using commercial names (i.e., ‘COMPASS’, ‘Spotlight‐AQ’, ‘Wysa’, ‘emoTICare’ and ‘DigiBete’) rather than by specific types of algorithms (e.g., random forest, support vector machine or Naïve Bayes), whilst providing a limited overview of the technical functioning of their models.

4. DISCUSSION

To our knowledge, this is the first systematic review of the use of AI in mental health–related diabetes care contexts. A review on this topic is important, given the rising prevalence of diabetes (particularly in LMICs), ⁷ , ⁹ the health and economic burden of diabetes, ¹⁰ the important role of mental health factors in exacerbating (and potentially improving) diabetes management, ² , ²⁷ the limited ability of current healthcare systems to cope with the increasing psychological and physical health needs of people who live with diabetes (especially in LMICs), ⁴ , ⁹ , ¹¹ , ¹² and the potential of AI to minimise and mitigate these challenges. ¹⁹

Data from this review suggest that research on this topic began to be conducted in 2013, with no studies prior to that meeting the inclusion criteria and has dramatically increased since 2020. Most included studies had a relatively large sample size, recruited participants from high‐income countries, employed an observational and cross‐sectional design, focused on assessment (as opposed to intervention), had depression as the primary outcome and primarily focused on an adult sample with Type 1 diabetes. Most studies employed multiple algorithms, primarily supervised learning algorithms, with those utilising deep learning algorithms performing the best. Despite an ongoing improvement in AI‐based technologies, no improvements in AI performance were evident over time.

AI‐based technologies offer transformative potential for diabetes care through personalised precision medicine, early detection of complications, real‐time continuous monitoring integrated with closed‐loop systems, enhanced clinical decision support optimising treatment selection and improved patient empowerment through explainable feedback. ⁶

As reflected in the results of this review, AI‐based technologies offer promising solutions to gaps in traditional healthcare systems, enabling the assessment and treatment of mental health conditions in individuals with diabetes, whilst also being able to use mental health conditions to predict the development of diabetes. These AI‐based capabilities hold notable clinical significance in providing necessary support to individuals with diabetes who develop mental health conditions that may exacerbate their ability to manage diabetes, and individuals with mental health conditions who have, or may be at risk of developing, diabetes that may further exacerbate the impact of their mental health conditions. This contribution to diabetes psychosocial care is important to consider as diabetes care currently accounts for 12% of the world's health expenditures, ¹⁰ whilst long waitlists, ¹² fragmented and complex service delivery pathways, ¹³ , ¹⁴ non‐uniformity of access, ⁶ , ¹⁴ , ¹⁵ financial costs, ¹⁶ and a shortage of health, particularly mental health, professionals, ⁶ , ¹⁷ , ¹⁸ can lead to over half of individuals with diabetes not receiving adequate care or remaining undiagnosed for years ⁶ , ⁷ —particularly in LMICs. ⁴ , ⁹

Despite their notable potential to improve outcomes for people living with diabetes and mental health conditions, AI‐based tools can have important limitations within diabetes contexts. ⁸ As per Ellahham (2020), ⁸ these limitations can be grouped into four categories, that is, those related to (a) human factors, (b) technical factors, (c) data limitations and (d) design limitations. ⁶ , ⁸ Of these categories, human factors, technical and design‐based limitations were the most salient in this review.

4.1. Human factors

Most reviewed studies did not use diverse samples; for example, (a) only 15 countries were represented, with 59% having a sample originating in high or high‐middle‐income countries, such as, the United States, China or the United Kingdom; (b) less than a quarter of the studies reported ethnicity‐related information, with no specific focus on indigenous populations; and (c) no studies focused on gestational diabetes, whilst a relatively limited number specifically focused on Type 1 or preclinical diabetes, despite these conditions having different physical and psychological needs compared with Type 2 diabetes. ⁴⁶ The geographical and methodological concentrations create substantial equity concerns, as the analysed AI models may fail to generalise to populations with non‐Type 2 diabetes, particularly in countries or amongst underrepresented demographic groups (e.g., indigenous populations) where the diabetes burden is high. This human‐factor‐based limitation is further compounded by the methodological limitation of the lack of cross‐cultural analyses in any of the reviewed studies. An associated data limitation, where authors do not share their source code, further exacerbates the issue by not revealing any cultural biases that may arise from the training data and coding, whilst also preventing comprehensive validation and reproduction of their findings. A survey of 400 algorithms presented in papers at leading AI conferences revealed that only 6% of the presenters shared the code for their algorithms, whilst only a third shared the data used to test their algorithms, and just half shared the ‘pseudocode’ (i.e., a limited summary of a source code algorithm). ¹⁰

The reviewed studies also remained limited in their examination of mental health factors, with most focusing exclusively on individuals with either mood and anxiety disorders or distress. This narrow scope further limits ecological validity and generalisability. For instance, depression and diabetes distress are clinically distinct constructs with different implications for care planning. However, systems trained primarily to detect ‘depression’ may inadvertently conflate diabetes‐related distress with a depressive disorder, which can affect clinical decision‐making and referral pathways. Future work should, therefore, (a) report construct definitions and measures clearly; (b) train AI models using a nuanced, differential diagnosis‐focused approach; and (c) evaluate clinical utility in terms of the models' impact on clinical decision‐making and subsequent care pathways.

An associated limitation of the current review is the absence of cross‐methodology and cross‐outcome AUC comparisons (and risk of bias assessment and analyses), due to insufficient statistical power and heterogeneity in the methodologies and outcomes of the included studies. Such analyses could have provided novel and valuable insights into whether AI (and which specific types of AI algorithms) perform the best for which particular mental health conditions and contexts.

4.2. Technical factors

Most studies included in this review employed supervised learning models (83%), despite neural network‐based deep learning models consistently outperforming simpler, linear models in the literature ⁴⁷ , ⁴⁸ —a trend reflected in the AUC‐based comparisons of the AI algorithms deployed by the reviewed studies. This apparent paradox is important to consider, as the predominance of simpler supervised learning approaches, despite evidence of deep learning superiority in research settings, likely reflects critical implementation barriers that prevent research findings from effectively translating into the most clinically applicable and impactful AI‐based technologies in diabetes care contexts.

This research‐practice translation gap manifests in several diabetes‐specific challenges that hinder the deployment of more sophisticated neural network models, despite their demonstrated performance advantages. Foremost amongst these is the reliance of deep learning models on substantially larger datasets and greater computational resources and costs than alternative AI algorithms, ⁴⁹ , ⁵⁰ which may not be deliverable in diabetes care contexts that already struggle to meet resource demands, particularly in LMICs. Beyond data requirements, deep learning models depend heavily on ‘black box’ mechanisms of information processing and consequently offer substantially reduced explainability, which likely diminishes clinician trust and acceptance. ⁴⁸ So, whilst the literature and current review highlight deep learning models as delivering superior performance, a fundamental trade‐off exists between (a) performance and (b) explainability, cost, computational efficiency and clinician trust. ⁴⁸ , ⁴⁹ , ⁵⁰ This trade‐off underscores the critical nuances that warrant careful consideration when implementing AI technologies in healthcare contexts, particularly where practical deployment constraints and clinical acceptability may ultimately outweigh performance gains.

4.3. Design limitations

In terms of design limitations, most studies employed cross‐sectional or retrospective data collection methods. This trend is reflective of the larger AI‐based diabetes literature, whereby studies tend to validate their models retrospectively rather than prospectively. ⁸ It is important that clinical study endpoints evolve to capture digital biomarkers from apps, monitors and trackers. ⁸ Furthermore, it is important to note that only 12% of the reviewed studies used a non‐randomised experimental design and a further two studies (5%) used an RCT design. This predominance of cross‐sectional observational studies presents a critical evidence gap that fundamentally limits causal inference regarding whether AI can effectively assess or intervene in mental health–related diabetes contexts, devoid of any confounding factors.

Furthermore, none of the intervention‐based studies used AUC; therefore, the interventional models could not be compared with their assessment counterparts in terms of performance. This divergence between assessment and intervention studies in outcome metrics could be seen as reflecting their distinct, yet equally important, objectives rather than a fundamental limitation. AI‐driven assessment can offer substantial standalone value by accurately identifying depression and diabetes distress in populations experiencing significant under‐detection, thereby liberating clinician time for personalised, evidence‐based intervention and therapeutic support. Assessment studies can be seen as appropriately employing AUC and discriminative metrics to determine whether at‐risk individuals can be identified (and their trajectory predicted), to what extent and how accurately. Whereas intervention studies primarily seek to determine whether the employed approaches meaningfully improve targeted clinical outcomes. Both sets of studies focus on questions that are essential for comprehensive clinical validation but are fundamentally different.

4.4. Future directions

In accordance with the gaps and limitations highlighted above, future research in this area should focus on five important areas:

1. Testing AI models across different clinical settings, and amongst people of different ages, genders, ethnicities (especially indigenous populations), rurality and accessibility to technology, socioeconomic deprivation, mental health conditions and various types and manifestations of diabetes, to evaluate the adaptability and performance variability of the models.

2. Adopting more prospective experimental designs, particularly with assessment‐based studies, whereby, via temporal validation, the models are tested to predict future outcomes.

3. Using more neural network‐based deep learning models instead of simpler and linear supervised learning models, whilst explicitly considering data representativeness, accessibility and potential inequities arising from differential deployment of these technologies or from existing access barriers to devices and connectivity required to access these technologies.

4. Reporting any adverse events encountered during the course of the study, as well as user acceptability and engagement.

5. Emphasising interpretability and clinical actionability (i.e., providing clear clinician‐oriented explanations and connecting predictive signals to potential interventions), and emphasising the evaluation of clinical utility (impact on outcomes) rather than statistical significance and discrimination.

More broadly, the use of AI in the mental health‐diabetes sector could profitably focus on the following five areas:

1. Risk‐stratified early assessment and intervention to identify individuals with preclinical diabetes or mild subclinical mental health disorders to enable primary prevention at critical junctures where intervention can alter disease trajectories and prevent progression.

2. Creation and implementation of precision stepped care models to match intervention intensity to individual needs, directing intensive resources towards the highest‐risk individuals whilst providing lighter support to lower‐risk patients, improving both clinical outcomes and cost‐effectiveness.

3. Propagation and application of precision mental healthcare as essential to precision diabetes medicine, with AI systems simultaneously modelling both mental health and metabolic domains to identify subgroups requiring integrated interventions that recognise the associations between mental health, diabetes management, quality of life and glycaemic control.

4. Adaptive assessment and intervention tools leveraging real‐time data to deliver precisely calibrated support at high‐risk moments, transforming mental health care from episodic clinical encounters into continuously adapted, real‐time psychological support for those with diabetes.

5. Greater emphasis on the creation, implementation and evaluation of mental health–related diabetes care technologies in LMICs, with a targeted focus on reducing health inequities.

5. CONCLUSION

This systematic review is the first to highlight the emerging role of AI in addressing mental health needs within the context of diabetes care. Living with diabetes imposes a significant psychological burden, increasing susceptibility to mental health challenges, such as, depression, anxiety and diabetes distress. These conditions not only affect emotional well‐being and quality of life but also impair diabetes management and worsen long‐term health outcomes. Whilst AI demonstrates promise as a psychological assessment and intervention tool within diabetes care contexts, current research remains limited in scope and methodological rigour in some respects. Most studies focus on mood and anxiety disorders or distress, employ cross‐sectional and observational designs, and over‐rely on supervised learning models, despite evidence that deep learning approaches have better performance. Key challenges include human factors (e.g., lack of diverse samples and outcomes), and technical and design limitations, such as, retrospective validation, absence of cross‐cultural or outcomes‐based analyses, and minimal use of randomised controlled trials. To realise AI's potential as a safe, accessible and effective solution to the current challenges faced by healthcare systems, future work must seek to prioritise diverse populations (especially in LMICs) and outcomes, prospective experimental designs, interpretability, clinical utility, advanced modelling techniques and equity‐based considerations in the deployment of these technologies. Addressing these gaps will be critical for integrating AI into person‐centred diabetes care and improving biopsychosocial outcomes for those with mental health disorders, diabetes, or both.

CONFLICT OF INTEREST STATEMENT

There are no conflicts of interest to declare. No external research funding was received for this project.

Supporting information

Data S1.

DATA AVAILABILITY STATEMENT

All data extracted and results derived from the studies included in this review will be uploaded to the journal's online platform and made available in accordance with the journal's data access policies and regulations.