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 services40 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.