Leveraging artificial intelligence in the prediction, diagnosis and treatment of depression and anxiety among perinatal women in low- and middle-income countries: a systematic review

Abstract

Aim

The adoption of artificial intelligence (AI) tools is gaining traction in maternal mental health (MMH) research. Despite its growing usage, little is known about its prospects and challenges in low- and middle-income countries (LMICs). This study aims to systematically review articles on the role of AI in addressing MMH in LMICs.

Methods

This systematic review adopts a patient and public involvement approach to investigate the role of AI in predicting, diagnosing or treating perinatal depression and anxiety (PDA) among perinatal women in LMICs. Seven databases were searched for studies that reported on AI tools/methods for PDA published between January 2010 and July 2024. Eligible studies were identified and extracted based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines using Covidence, and the data were synthesised using thematic analysis.

Results

Out of 2203 studies, 19 studies across eight countries were deemed eligible for extraction and synthesis. The review revealed that the supervised machine learning method was the most common AI approach and was used to improve the early detection of depression and anxiety among perinatal women. Additionally, postpartum depression was the most frequently investigated MMH condition in this study. Further, the review revealed only three conversational agents (CAs)/chatbots used to deliver psychological treatment.

Conclusions

The findings underscore the potential of AI-based methods in identifying risk factors and delivering psychological treatment for PDA. Future research should investigate the underlying mechanisms of the effectiveness of AI-based chatbots/CAs and assess the long-term effects for diagnosed mothers, to aid the improvement of MMH in LMICs.

PROSPERO registration number

CRD42024549455.

WHAT IS ALREADY KNOWN ON THIS TOPIC.

WHAT THIS STUDY ADDS

• We identified 19 studies highlighting the AI-based approaches used in predicting, diagnosing and treatment of PDA in LMICs. Overall, postpartum depression (PPD) was the most common mental health (MH) condition studied, and the supervised machine learning (SML) method was the most common AI-based approach adopted.

• The SML method was employed largely to identify risk factors to enhance early detection and prevention (eg, identifying at-risk women and implementing preventive measures) for women at risk of developing PPD.

• While AI-based tools have been effective in this area, few studies have demonstrated their application in delivering psychological treatment and support for new mothers (ie, therapeutic interventions for women already experiencing PPD).

• Many of the studies lacked details of the ethical considerations for the AI-based tools/methods in predicting, diagnosing and/or treating PDA in the LMIC context.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY.

• This study provides a foundation for developing culturally appropriate AI-based solutions in resource-limited settings. Healthcare providers may be encouraged to implement AI-based interventions carefully, potentially improving PDA management where MH resources are scarce.

• The effectiveness of the AI-based conversational agents/chatbots for delivering psychological treatment and support requires further investigation, with comparison across contexts.

• Ongoing literature synthesis is recommended to inform the ethical considerations made by the researchers and practitioners to ensure LMIC-specific guidelines are adopted. This could help policymakers balance AI benefits with regional challenges and ultimately promote more equitable maternal mental healthcare.

Background

Maternal and child health (MCH), a critical indicator of global healthcare systems, has seen improvements in recent decades^{1 2} and reflects a growing commitment to MCH.² Despite these advancements, many low- and middle-income countries (LMICs) struggle to improve MCH outcomes. The slow progress can be attributed, in part, to the insufficient attention given to certain crucial MCH indicators, including maternal mental health (MMH).³ MMH, a crucial contributor to the overall health and well-being of mothers and young children, is often ignored, with postpartum depression as the most common type of MMH.⁴

Approximately 16% of pregnant women and 20% of new mothers experience depression and anxiety in LMICs.⁵ However, these statistics can be much larger for different outcomes. For instance, Dennis et al⁶ reported that nearly one in three women in LMICs experience anxiety during the perinatal period. In India and China, reports show a 34% and a 30.7% rate of depression, respectively, among postpartum women.⁷ The high rates, coupled with the potential lack of attention to MMH conditions, have profound implications on the health and well-being of expectant and new mothers and their newborns. For instance, Hagatulah and colleagues reported an increased risk of mortality among women with perinatal depression (PD).⁸ Also, perinatal depression and anxiety (PDA) has been found to impact child development from birth to early childhood.^{5 9} This shows that it affects infant growth, health, nutrition and mother–child bonding and delays cognitive and language skills in infants. These underscore the urgent need for comprehensive maternal care that addresses both physical health and mental health (MH) needs.¹⁰

Despite concerted global efforts to eradicate or minimise the prevalence of maternal mental disorders, progress remains slow, especially in LMICs. Maternal mental disorders including PDA often go unrecognised and untreated in many LMICs due to intricate sociocultural factors, including stigma, poor community understanding of MH and limited access to appropriate care.¹¹ Many societies hold idealised views of motherhood that conflict with maternal experiences of PDA, and women may fear being judged or feel pressured to meet unrealistic expectations.¹² Moreover, in some cultures, discussing MH remains taboo.¹³ In other contexts, the lack of inclusion of local language complicates access to MH services.¹⁴ These interconnected challenges result in many women not receiving the MH support they need during and after pregnancy.

Technological advancements have been shown to help fill the existing gap and transform the healthcare sector and systems. Digital health interventions, including telemedicine and mHealth apps, have expanded access to MH services¹⁵ and improved treatment options.¹⁶ Recently, such interventions have included artificial intelligence (AI)—science-based, engineered intelligent tools and methods that replicate human cognitive processes like learning and problem-solving by following algorithms or a set of rules.¹⁷ These tools and methods include machine learning (ML) models to improve prediction and diagnosis or the use of conversational agents to facilitate psychological treatment and support. For PDA, applications of AI in high-income countries include early detection using ML methods,¹⁸ personalised risk assessment^{19 20} and AI-powered chatbots for support.²¹ Natural language processing analyses social media for PDA signs,²² while ML optimises treatment plans.²³ In LMICs, the use of AI tools or methods in addressing PDA is still a growing area, and not much is known about the role it plays in the prediction, diagnosis and/or treatment of PDA. The application of AI in addressing PDA in LMICs is of particular interest, given the high prevalence and often underdiagnosed nature of these conditions in such settings.¹¹ Also, many LMICs grapple with limited healthcare infrastructure and a shortage of MH professionals.²⁴ AI presents a promising avenue for expanding access to care and addressing PDA in LMICs. The potential benefits of adopting AI in MMH care vis-à-vis the existing challenges warrant carrying out relevant research to assess how effective AI can be adopted considering the barriers. Our review integrating the patient and public involvement (PPI) approach, which is uncommon in systematic reviews, highlights common existing AI methods and tools for predicting, diagnosing and treating PDA in LMICs. The study again highlights challenges, including ethical issues, in using AI tools for predicting, diagnosing and treating PDA in LMICs.

Objective

This study aims to conduct a systematic review investigating the role of AI in addressing PDA in LMICs. The review focuses on identifying and summarising the AI tools and methods used in the prediction, diagnosis and treatment of PDA among pregnant women and new mothers in LMICs.

Methods

Registration and deviations from protocol

This review adopted the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for searching, screening, selecting and extracting data from the studies.²⁵ The protocol was registered in PROSPERO (ID: CRD42024549455) and approved on 10 June 2024.²⁶ We made the following changes to the protocol: adjusted the search period to cover articles published from January 2010 to July 2024 and included Google Scholar in the databases (this is limited to the first 100 searches).

Patient and public involvement

The systematic review was conducted with the involvement of persons with lived experience of PDA, non-governmental organizations (NGOs), AI experts, health professionals and researchers interested in MMH, in the codesign process to ensure that the research met the needs of the target population. This was a unique approach as emphasised by Cook et al,²⁷ as its application is unusual for systematic reviews and health research carried out in LMICs.²⁸ Before the project launch, we engaged an NGO—APEC Ghana (APECGH), invited them to partner with the team and shared the research proposal. They shared the proposal with their members with lived experience of PDA, and they shared their views on the research process. We codeveloped plans for 10 focus group discussions and an in-person workshop with maternal health stakeholders in Accra, Ghana. Given their expertise, they led community entry and participant recruitment for both activities. We also engaged new mothers (within the first year post partum), health professionals, AI experts and researchers before the initial search for articles. We shared the initial keywords developed by the coinvestigators (USA, AOO and SK) and the research assistants (CSF and SA-A), and they were able to add keywords to the list. During the in-person workshop, the findings from the systematic review were disseminated to participants, including women with lived experience and other stakeholders, and received feedback. The involvement of patients and the public in the codesign process improved the relevance of this research and fostered a sense of ownership among the stakeholders. For example, since collaborating with APECGH, they have led efforts to raise awareness about the research and advocate for PDA through their platform.

Eligibility criteria, search strategy and selection process

The eligibility criteria were developed using the PICOS (Population-Intervention-Comparator-Outcomes-Study type) framework.²⁹ A combination of subject headings and keywords was used for five key search concepts related to (1) AI, (2) pregnant women and new mothers, (3) prediction, diagnosis or treatment, (4) depression and anxiety and (5) geographical setting. The included studies met the following criteria: (1) the population was pregnant women and/or new mothers in the perinatal period; (2) the setting was an LMIC as defined by the World Bank Group³⁰; (3) the study featured an application of AI tool and/or method in the prediction, diagnosis and treatment of PDA; (4) written in English; and (5) published on or after January 2010 to July 2024, with no restrictions on the study design. We excluded studies not focusing on using AI-based tools or methods for the prediction, diagnosis or treating PDA. Our first search was conducted on 13 May 2024 and the last search date was 26 July 2024. Additional details on the inclusion and exclusion criteria are outlined in table 1.

Table 1. Eligibility criteria for studies to be included in the review.

PICOS	Inclusion criteria	Exclusion criteria
Population	Study participants must have been pregnant women or new mothers from any low- and middle-income country following the World Bank (2024) income classification. These women can be at risk of depression and anxiety, validated using screening questionnaires or interviews with health professionals to identify those at risk. They could also have received a formal diagnosis made using formal clinical criteria (eg, Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition) or the International Classification of Diseases, 11th Revision).	We exclude studies for which the participants (the women) are not currently pregnant or not in the perinatal period. We excluded studies on pregnant women and new mothers whose focus is not on PDA. This implies that other forms of perinatal MH such as post-traumatic stress disorder (PTSD), bipolar disorder, psychosis or adjustment disorders were excluded from the study. Finally, we exclude studies which include pregnant women or new mothers but use data that are not well defined within a low- and middle-income country. For example, data from social media or online chat rooms or discussion boards without clear jurisdiction or cuts across a global audience.
Intervention	Any artificial intelligence, machine learning, deep learning, natural language processing or other predictive analytical techniques. The tools and methods encompass AI-driven software, technology and electronics, mobile applications, machine learning, deep learning and other natural language processing methods.	The following interventions were excluded: focused on AI tool/method, AI-based tool/method was not applied for improving the prediction, diagnosis or treatment of PDA. We excluded other health technologies, for example, mHealth, if they were not driven by AI or their analytical methods were not driven by AI.
Comparator	Usual care provided by health professionals; no treatment or alternative AI algorithms.	No exclusion criteria.
Outcomes	Depression and/or anxiety were the primary outcomes of the study—prevalence and risk factors. Populations in which the AI tool and/or method has been applied; users of the technology; feasibility, acceptance and effectiveness for prevention, diagnosis, or early intervention or treatment of depression and anxiety (eg, accuracy, health outcomes, and harms or adverse events).	No exclusion criteria.
Study type	Randomised controlled trials, non-randomised quantitative studies, qualitative studies; included studies could be experimental (eg, randomised controlled trials and cluster randomised trials) and quasiexperimental studies. They could also be non-intervention studies; quantitative, qualitative or mixed-methods research papers using primary and/or secondary data analysis.The articles should have been published in the English language.	Review studies, protocols, expert opinions, guideline reports and non-peer-reviewed articles.

The included studies had to provide information on artificial intelligence tools or methods used in the prediction, diagnosis and treatment of depression and anxiety affecting pregnant women and/or new mothers. Studies focusing on other MH conditions (eg, stress) were excluded.

AI, artificial intelligence; MH, mental health; PDA, perinatal depression and anxiety.

A comprehensive search was conducted (USA) using the following databases: MEDLINE (via OVID), PsycINFO (via OVID), CINAHL, ACM Digital Library, Web of Science (complete core collection), Scopus and Google Scholar. Additionally, backward and forward citation searches were performed (USA) using CitationChaser.³¹ Covidence Veritas Health Innovation³² was used to deduplicate all imported search records and facilitate a blind screening of the articles. All titles and abstracts were screened by two researchers working independently. USA was the consistent reviewer for all titles and abstracts, with SK, SA-A and CSF acting as the second reviewers working independently. The mean percentage agreement was 74.6%. All screened studies that met the criteria for inclusion were considered for full-text review. Retrieved articles were fully screened by three reviewers (USA, AOO and SK) also working independently. The mean percentage agreement was 87.3%. Articles were excluded if they did not meet the inclusion criteria. All discrepancies for inclusion or exclusion were resolved by a third reviewer. (See table 2 for a sample of our search strategy.)

Table 2. Keywords and search strategies.

Electronic databases	Primary keywords	Sample search string
MEDLINE (via OVID) PsycINFO (via OVID) CINAHL ACM Digital Library Web of Science Scopus Google Scholar	Artificial Intelligence; Machine Learning; Deep Learning; Natural Language ProcessingPerinatal; Pregnancy;PostpartumDepressionAnxiety	Scopus - TITLE-ABS-KEY (‘artificial intelligence’ OR ai OR a.i. OR ‘machine learning’ OR ‘deep learning’ OR ‘deep reinforcement learning’ OR ‘reinforcement learning’ OR ‘intelligent simulation’ OR ‘artificial neural network*’ OR ‘artificial intelligen* decision support system*’ OR ‘artificial intelligen* decision support systems, clinical’ OR ‘artificial intelligen* decision support techniques’ OR ‘Artificial intelligen* internet’ OR ‘Artificial intelligen* internet of things’ OR ‘Artificial intelligen* chatbot*’ OR ‘Artificial intelligen* ChatGPT*’ OR ‘BING AI’ OR ‘Google BARD AI’ OR ‘artificial intelligen* telehealth’ OR ‘artificial intelligen* telemedicine’ OR ‘artificial intelligen* telemonitoring’ OR ‘artificial intelligen* telepractice’ OR ‘artificial intelligen* telenursing’ OR ‘artificial intelligen* telecare’) AND TITLE-ABS-KEY (‘perinatal period’ OR parturition OR ‘peripartum period’ OR ‘postpartum period’ OR pregnan* OR ‘pregnant wom*n’ OR ‘breastfeed* mother*’ OR antenatal OR antepartum OR ‘during pregnancy’ OR gestation* OR ‘expectant mother*’ OR intrapartum OR ‘maternal-fetal period’ OR peri-partum OR peripartum OR ‘peri partum’ OR postdelivery OR ‘post-delivery’ OR post-partum OR postpartum OR ‘post partum’ OR prenatal OR ‘pre-natal’ OR postnatal) AND TITLE-ABS-KEY (anhedonia OR ‘affective disorder’ OR ‘depressive disorder’ OR depression OR ‘Anxiety Disorders’ OR anxiety OR ‘Depressive Disorder, Major’ OR ‘Dysthymic Disorder’ OR ‘Phobic Disorders’ OR ‘blue*’ OR ‘clinical depression’ OR ‘depress*’ OR ‘dysthymia’ OR ‘low mood’) AND TITLE-ABS-KEY (‘low-and-middle income’ OR lmics OR ‘developing countr*’ OR ‘global south’ OR africa* OR ‘Korea, Dem. People&apos;s Rep’ OR cambodia OR kiribati OR ‘Lao PDR’ OR ‘Micronesia, Fed. Sts.’ OR mongolia OR ‘Myanmar (Burma)’ OR ‘Papua New Guinea’ OR phillippines OR samoa OR ‘Solomon Islands’ OR ‘Timor-Leste’ OR vanuatu OR vietnam OR china OR fiji OR indonesia OR malaysia OR ‘Marshall Islands’ OR palau OR thailand OR tonga OR tuvalu OR ‘Kyrgyz Republic’ OR tajikistan OR ukraine OR uzbekistan OR albania OR armenia OR azerbaijan OR belarus OR ‘Bosnia and Herzegovina’ OR bulgaria OR georgia OR kazakhstan OR kosovo OR moldova OR montenegro OR ‘North Macedonia’ OR ‘Russian Federation’ OR serbia OR turkiye OR turkmenistan OR bolivia OR haiti OR honduras OR nicaragua OR argentina OR belize OR brazil OR colombia OR ‘Costa Rica’ OR cuba OR dominica OR ‘Dominican Republic’ OR ‘El Salvador’ OR ecuador OR grenada OR guatemala OR jamaica OR mexico OR paraguay OR peru OR ‘St. Lucia’ OR ‘St. Vincent and the Grenadines’ OR suriname OR ‘Syrian Arab Republic’ OR ‘Yemen, Rep’ OR algeria OR djibouti OR ‘Eqypt, Arab Rep.’ OR jordan OR ‘Iran, Islamic Rep’ OR lebanon OR morocco OR tunisia OR iraq OR libya OR ‘West Bank and Gaza’ OR afghanistan OR bangladesh OR bhutan OR india OR nepal OR pakistan OR ‘Sri Lanka’ OR maldives OR ‘Burkina Faso’ OR burundi OR ‘Central African Republic’ OR chad OR ‘Congo, Dem. Rep’ OR eritrea OR ethiopia OR ‘Gambia, The’ OR ‘Guinea-Bissau’ OR liberia OR madagascar OR malawi OR mali OR mozambique OR niger OR rwanda OR ‘Sierra Leone’ OR somalia OR ‘South Sudan’ OR sudan OR togo OR uganda OR angola OR benin OR ‘Cabo Verde’ OR cameroon OR comoros OR ‘Congo, Rep.’ OR ‘Cote d&apos;Ivoire’ OR eswatini OR ghana OR guinea OR kenya OR lesotho OR mauritania OR nigeria OR ‘Sao Tome and Principe’ OR senegal OR tanzania OR zambia OR zimbabwe OR botswana OR ‘Equatorial Guinea’ OR gabon OR mauritius OR namibia OR ‘South Africa’) AND PUBYEAR>2009 AND PUBYEAR<2025 AND (LIMIT-TO (DOCTYPE, ‘ar’)) AND (LIMIT-TO (LANGUAGE, ‘English’))

Furthermore, two of the authors (AOO and USA) performed a risk assessment on all the studies included in this review to minimise bias. They used the Joanna Briggs Institute critical appraisal tool, which offers a variety of tools tailored for different types of research. The review included five types of studies: analytical cross-sectional studies (11), qualitative studies (2), cohort studies (1), quasiexperimental studies (4) and cluster randomised controlled trial (1). In total, 14 out of the 19 studies received an appraisal score of 70% or higher, indicating high quality. Four studies scored between 50% and 70%, reflecting medium quality, while one study scored below 50%, categorising it as low quality (refer to online supplemental tables 8–12 for details about the risk of bias assessment).

Data extraction and analysis

Data were extracted by reviewers (USA, AOO and SK), with the extraction by one reviewer (eg, USA) verified by a second reviewer (AOO or SK). The data extracted covered study details (eg, author(s), year of study, geographical setting and study design), population (eg, pregnancy stage, age, sample size), interventions (eg, ML, deep learning, conversational agent or chatbot), outcomes (depression or anxiety, measures used to screen) and additional findings (eg, risk factors). Information on ethical considerations such as data privacy, informed consent, algorithmic transparency, potential biases and effects on doctor–patient relationships was also examined and extracted. Extracted data were examined and analysed using a narrative review approach to identify recurring themes and relationships. This method was used due to the substantial diversity among the included studies regarding the methodology, interventions, topics, heterogeneity and method of reporting outcomes. Due to heterogeneity in the quantitative study designs and outcomes, meta-analysis could not be performed as part of this review.

Findings

Study selection

Searches of the seven databases identified 2201 references. Backward and forward citation searching found 1205 references. A total of 122 full-text articles were screened, yielding 19 articles included in the study. See the selection process in the PRISMA flow diagram in figure 1.

Figure 1. Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 flow diagram of article searches. AI, artificial intelligence; LMICs, low- and middle-income countries. Source: Page et al.²⁵.

Study, participant and sample characteristics

Online supplemental table 1 presents details on the study characteristics. The studies were conducted in eight middle-income countries: four each in China and Nepal; two each in Brazil, India, Pakistan, Sri Lanka and Türkiye; and one in Kenya. Studies covered a 6-year period (ie, 2019–2024) with more than half published since 2022 (n=11; ~58%). All articles used quantitative methods purely or as part of a mixed-methods design. Out of these, five were intervention based, with one cluster randomised controlled trial. For the seven articles that employed qualitative techniques, all but one adopted interview as its data collection strategy. There was an equal split between participants recruited during pregnancy (n=10) and the postpartum period (n=10). Some of the studies did not report clear upper or lower limits of the participants’ ages. Of those that did, the minimum age was 15 years, and the maximum age was 54 years. The sample size was varied, ranging from 13 to 5666 women with five articles having <100 women in their study. The reasons for the small sample include issues with consent, pilot nature of the tool being tested and focus on specific women, for example, categorised as depressed or non-depressed. For example, as noted by Montenegro et al,³³ “we opted for a reduced sample because our objective with this first study was to capture perceptions related to the two groups without failing to observe contributions and reported failures about the chatbot.” Online supplemental tables 2 and 3 have additional details for each included study. Online supplemental table 3 describes the study aims and clinical goals of the included studies.

Screening procedures

Online supplemental table 4 presents details of the outcomes and the screening procedures. Out of the 19 articles, 14 (68.4%) studied postpartum depression alone or as part of PD. The Edinburgh Postpartum Depression Scale-10 was the most frequently used screening tool, employed alone or along with other tools in eight articles (~42%). Of those eight articles, half of the participants were recruited during pregnancy. The second most common validated screening tool is the Patient Health Questionnaire-9, which was adopted in six of the 19 articles (31.6%). Other scales used for screening depression are found in online supplemental table 4. Anxiety was scarce as a primary outcome in the reviewed article but was considered mostly as a potential driver for depression. For instance, Hong et al³⁴ included anxiety as a psychophysiological variable with a potential risk for maternal depression, and Montenegro et al³³ examined whether anxiety was a significant driver for engagement with a chatbot developed to improve women’s health literacy during pregnancy. The following tools were used to screen for anxiety: the 20-item Self-Rating Anxiety Scale, Perinatal Anxiety Scale (n=2) and the Generalised Anxiety Disorder-7. Given that the scales are largely questionnaire based, they were mostly self-administered (n=13) and screened once at the time of recruitment (n=12). For example, Javed and colleagues³⁵ reported that participants were screened during routine prenatal care. Other studies reported multiple screenings. For example, Zhang et al³⁶ reported screening the participants in the first trimester through 6 weeks post partum (see online supplemental table 4 for more details).

AI tools and methods

Online supplemental table 5 details the AI-based tools and methods adopted in the articles. The most common approach used is learning models with focus on enhancing prediction and diagnosis rather than treatment of depression. In terms of the learning models, supervised machine learning (SML), which uses algorithms to identify patterns of information in labelled training data and then test on unlabelled test data, was the most common AI technique (15/19 articles). The general purpose of the algorithms as used in the studies was to aid prediction and diagnosis, that is, to identify risk factors that can help classify women at risk of PD, in a bid to enhance early detection and intervention for affected women. The most commonly reported risk factor was socioeconomic status/income/financial-related issues, which were linked to the availability of income and ability to manage available income.37,39 Additional details on identified risk factors are presented in online supplemental table 6. Most studies adopted common performance metrics such as the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity and precision (see ref ⁴⁰ for an explanation of these metrics). The use of AI for treatment of PDA was only reported in a few studies. For treatment, we identified three AI-based interventions where women engaged with chatbots or conversational agents. For instance, the StandStrong platform was developed for non-specialists (lay counsellors or midwives) to deliver the psychological treatment—Healthy Activity Program—to adolescent and young women.40,42 The other examples were the conversational agent developed to interact with pregnant women in Brazil via Facebook³² and the Zuri text-based messaging AI system used to deliver the Healthy Moms PD intervention to pregnant women and new mothers in Kenya.⁴³

Ethical considerations

We also report on the extent to which the articles describe the ethical challenges or implications of their approach (see online supplemental table 7). While AI-driven tools or methods offer a lot of potential advantages, they also pose ethical issues that need to be addressed.⁴⁴ The majority of the articles (n=15) reported that they had received approval to conduct the studies. Of those, 12 articles did not describe the considerations in detail but noted that consent from the participants and/or their family members was required before any data collection. For those who reported, key concerns included privacy, confidentiality and data security. For instance, Maharjan et al⁴¹ noted that the specifically provided information on how participants could pause data collection at any time and the safe use of the devices to ensure no harm comes to the participants.

Conclusions and clinical implications

This review analysed 19 studies focusing on using AI-based tools and/or methods in predicting, diagnosing and treating depression and anxiety among pregnant women and new mothers in LMICs. As far as our knowledge extends, this study represents one of the pioneering efforts to identify and synthesise evidence research focused on LMICs in relation to the topic under investigation. Our synthesis revealed that the most prevalent approach is the use of SML techniques for the prediction of risk factors of PDA in a bid to improve early diagnosis of PDA. The few studies that focused on treatment primarily aimed to reduce symptoms,^{33 42} enhance MMH services^{40 45} and provide support for women diagnosed with PDA,⁴³ thus warranting the use of different AI tools using conversational-based agents or chatbots. Considering the subclassification of the studies (symptoms reduction, service enhancement and provision of support) and the dearth of studies aimed at treating PDA, there exists a notable opportunity for additional research and advancement in this area. Further, the lack of direct comparisons among different AI tools for treating PDA highlights a gap in understanding the relative effectiveness of these AI tools.

Regarding the methods and modalities used by the studies in review, we observed a predominance of primary research methods in these studies. This may be due to the lack of nationally representative databases for MMH outcomes, particularly PDA, in many LMICs. Most available databases in LMICs are focused on physical health outcomes such as body mass index (BMI; measured using height and weight), morbidity and mortality as compared with MH outcomes. This lack leads to a high cost of collecting primary data for researchers in this field.³⁵ Additionally, this may pose challenges to obtaining high-quality data essential for training AI models, result in biases in the data due to small sample sizes, and data privacy issues may compromise the result validity and generalisability.⁴⁶

Furthermore, recent advancements in ML have highlighted the importance of selecting relevant variables to build accurate, reliable and contextually relevant models for the prevention, screening, diagnosis and treatment of PDA.⁴⁷ We observed that the variables selected in AI and PDA research vary depending on the objectives of the studies. For instance, studies that aimed to predict risk factors often used variables directly or closely associated with the individual (the participant and her social relationships), such as sociodemographic and socioeconomic variables, social environmental factors, physical conditions, intimate partner violence, BMI, etc. Conversely, studies that aimed at treatment focused on passive sensing data such as location and proximity-related data, behavioural data (mother’s activity) and MH care utilisation data. This shift from traditionally primary data to passive sensing data suggests a move towards leveraging real-time, behavioural data for nuanced treatment strategies. Given the real-time data collection nature, the passive sensing method may help minimise the risk of bias as compared with the traditional data collection methods like the self-reported questionnaires adopted by most prediction-based studies. Passive sensing techniques in MH research offer advantages such as objective data collection and early symptom detection.⁴⁸

However, they also present limitations, notably safety and privacy concerns.⁴⁶ Although many studies in this review obtained approval from Institutional Review Boards, details on critical ethical issues like participant data confidentiality, data ownership and sharing policies, and participant autonomy were inconsistently reported. These observations align with several ethical challenges highlighted by Fiske et al and Saeidnia et al, including data ethics, harm prevention and gaps in ethical and regulatory frameworks.^{49 50} Moreover, the inconsistency in reporting underscores the necessity for increased transparency and focus on ethical considerations within the realm of AI and MH research, particularly in LMICs, to safeguard the research participants. Another limitation of the included studies is that the majority relied on self-reported measurement tools. This is likely due to the nature of the studies, which were primarily field studies rather than clinical studies. While self-reporting can provide valuable insights, it also introduces the risk of bias, as participants may unintentionally overstate or understate their experiences. This reliance on self-reported tools can lead to skewed results, potentially affecting the validity of the findings and the conclusions drawn from the research.

Our review has the following strengths. First, the adoption of the PPI strategy allowed for a comprehensive and diverse perspective on the findings. Second, the use of the Covidence software and multiple authors to manage and streamline the systematic review helped enhance the quality of the review process and reduce the risk of bias. Another strength of this review lies in the use of a wide range of interdisciplinary databases such as MEDLINE, CINAHL, PsycINFO, Web of Science, ACM Digital Library, Scopus, and Google Scholar. By using these databases, the review would access a diverse range of literature and perspectives related to the topic at hand. Despite these strengths, the review has some limitations. First, the review is limited by its predetermined focus on English language articles published between 2010 and July 2024. This restriction may exclude relevant research published in other languages and before 2010, potentially leading to the omission of important articles or findings. The language limitation arises from the researchers’ language proficiency in only English and the lack of funding to cover the cost of hiring international translators. Additionally, the review includes only articles related to the application of AI tools and methods, excluding other technological approaches and methods used in addressing PDA. This narrow focus may overlook potentially valuable insights from related fields or alternative technological solutions that could be relevant to the literature on PDA. The study included women at all stages of the perinatal period, that is, during pregnancy and the postpartum period and for all ages. This might have hidden important differences in how AI tools work for women at different perinatal periods or ages. Additionally, the current study summarised risk factors associated with depression and anxiety among perinatal women as documented in the studies. However, it did not explicitly separate the specific risk factors associated with clinical diagnoses versus those associated with symptom severity across the included studies. Future studies should look at these groups separately to better understand how AI can help predict, diagnose and treat PDA for specific groups of women. Additionally, it would be crucial to distinguish between risk factors associated with a clinical diagnosis of major/minor depression and/or anxiety disorder from those associated with the symptom severity.

As AI continues to make inroads into the health and healthcare domain, its integration into addressing PDA represents an underexplored area in LMICs. The use of AI-based chatbots or conversational agents for PDA is still growing, with limited data to warrant a meta-analysis to assess the effectiveness. Future research could focus on identifying the most efficient and impactful AI tools for delivering psychological treatment for PDA. This exploration is crucial for advancing the application of AI in addressing PDA effectively and improving MH care in LMICs. Further, future research should aim to provide detailed and transparent accounts of the ethical frameworks and practices employed throughout the study. By doing so, researchers in this field can uphold ethical standards, promote transparency and ensure the participants are well protected. Filling these research gaps is crucial for advancing AI in improving MCH outcomes in LMICs.

The views expressed are those of the authors and not those of the management and staff of MQ Mental Health Research.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.