Continuity of care refers to the degree to which a series of health care services are experienced by patients as coherent, connected, and consistent with their medical needs and personal circumstances [1-3]. It encompasses informational continuity (shared data across health care providers), relational continuity (stable patient-provider relationships), and management continuity (coordination across care settings) [4-6]. Effective continuity of care has been linked to improved patient satisfaction, better health outcomes, and reduced health care use, particularly for individuals requiring long-term management or multidisciplinary support [7-11].

Despite its importance, ensuring continuity of care remains a persistent challenge. Fragmentation in health systems, barriers to data sharing, and transitions between health care providers often disrupt seamless patient management [12-16]. Discontinuities are particularly frequent at the interface between primary, secondary, and community care. Limited interoperability of electronic health records, variations in clinical practices, and resource constraints further impede coordinated care delivery [15-18]. Consequently, patients may experience inconsistencies in information exchange [19], duplication of tests [20,21], or complete disengagement from necessary follow-up care [14,22-24].

Continuity of care poses additional complexities in psychiatry, where patients often belong to vulnerable populations requiring sustained, holistic management [25-27]. Mental health conditions are characterized by chronicity; fluctuating symptoms; and the interplay of biological, psychological, and social factors [28-32]. Disruptions in care continuity—for instance, due to health care provider turnover, care transitions, or inadequate follow-up—can lead to relapse, hospitalization, or disengagement from treatment [33-39]. The stigma surrounding mental illness, along with insufficient integration between psychiatric and general health services, further compounds these challenges [40,41]. As psychiatric care heavily depends on therapeutic relationships and consistent communication, fragmented care delivery has particularly adverse implications for patient well-being [42-47].

Over the past decade, AI has evolved from a niche technological pursuit into a transformative force with the unprecedented capacity to handle complex data, recognize patterns, and automate decision-making processes. In health care, this automation extends beyond administrative functions to diagnostic support, personalized treatment recommendations, and longitudinal patient monitoring [48-51]. AI applications have increasingly focused on patient management through tools that enable data-driven clinical decision-making, early detection of health risks, and adaptive care coordination [52]. Predictive algorithms analyze large-scale patient data to identify potential deterioration in health status, while conversational AI tools facilitate continuous patient engagement beyond clinical settings. AI-powered systems also streamline information exchange among multidisciplinary teams, contributing to more coherent and responsive care pathways. These technologies pave the way for a model of proactive and personalized health care, in which patient interactions are not episodic but are sustained through data connectivity and intelligent automation. The integration of AI-driven automation has thus become central to improving efficiency, reducing clinician workload, and enhancing the precision of care delivery [52-59].

Within psychiatry, AI offers promising avenues to strengthen continuity of care by enhancing information flow, patient monitoring, and clinical decision-making. Machine learning (ML) algorithms can track longitudinal patterns in patient data to flag early signs of relapse or risk behaviors, allowing timely interventions. AI-driven chatbots and mobile mental health apps support patient engagement, bridging gaps between appointments and promoting adherence to treatment plans [45,47,60,61]. Furthermore, AI can facilitate integration across care settings by synthesizing multimodal data from electronic health records, therapy notes, and wearable devices (digital phenotyping) and recognizing subtle patterns and longitudinal trends that may be missed even by experienced health care professionals, thereby generating cohesive, patient-centered insights. By automating routine tasks and augmenting human oversight, AI removes the necessity for clinicians to provide constant manual oversight for every individual in their caseload, thereby supporting systems facing acute staff shortages. This allows the system to “triage” human attention to patients flagged by AI as being at high risk or in crisis, thereby fostering more continuous, personalized, and coordinated psychiatric care [62-64]. Nevertheless, recent work has similarly emphasized that AI systems used for conversational or risk-oriented support in health care settings should be evaluated not only for task performance but also for safety-sensitive reasoning and oversight requirements [65].

Despite a growing number of studies in this area, a rigorous synthesis of the evidence is needed. This systematic review aims to comprehensively evaluate the effectiveness and implementation factors of AI and ML interventions specifically designed to improve the continuity of patient care in psychiatric care settings, thereby contextualizing the current state of knowledge for researchers, clinicians, and policymakers [45,66,67].

Following the search, all identified records were exported to Covidence (Veritas Health Innovation) systematic review software for centralized management. To ensure a clean dataset, duplicate records were removed using the software’s automated deduplication algorithm, followed by a manual verification process to identify any remaining duplicate records.

The selection of studies followed a rigorous 2-stage process conducted independently by at least 2 reviewers (EJT, WJDY, XEO, JF, IHAP, and MA). In the first stage, reviewers screened the titles and abstracts of all retrieved records in parallel. To ensure interrater reliability, a pilot screening of 20 articles per reviewer pair was conducted with a target agreement threshold of 80%; disagreements during this pilot were analyzed to refine the application of the eligibility criteria.

In the second stage, the full-text documents of all potentially eligible records were retrieved and independently assessed against the inclusion criteria. Throughout both stages, any discrepancies between reviewers were resolved through consensus or, where necessary, through arbitration by a third reviewer. The final results of the selection process, including specific reasons for the exclusion of full-text articles, will be documented and presented in a PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram.

Data will be extracted from the included studies using a standardized data extraction form, which will be developed and prepiloted in a spreadsheet program. To ensure the reliability of the findings and minimize the risk of transcription errors, the extraction will be performed by 2 reviewers independently and in duplicate. Following the extraction phase, the reviewers will meet to compare their findings, and any discrepancies will be resolved through consensus or, if necessary, through arbitration by a third member of the research team. If crucial data—such as SDs or specific outcome values—are missing or insufficiently reported, the authors of the original studies will be contacted via email. A total of 3 contact attempts will be made at weekly intervals over a 3-week period before the study is assessed based solely on the published information or excluded from the meta-analysis.

The variables to be extracted will be categorized into study characteristics, participant details, and technical and clinical outcomes. Regarding study characteristics, we will record the author, year of publication, country, study design, clinical setting, and funding source. The clinical setting will document the health care tier (eg, primary, secondary, or tertiary) and the precise care settings and direction of transition (eg, inpatient psychiatric ward to primary care and general practitioners, or emergency triage to community teams). To address the PICOS framework, we will extract data on the patient diagnosis, population size, the specific type of AI or ML architecture used (including the algorithm and its primary function), safety fallback pathways with protocols for emergency human-in-the-loop override, a description of the comparator, and the specific outcomes measured. For studies reporting quantitative results, we will extract the means and SDs, incidence rates, relative risks (RRs), hazard ratios (HRs), model sensitivity, specific false-negative rates, undertriage error frequencies, and 95% CIs for all primary and secondary outcomes. Furthermore, for studies involving qualitative components, we will extract relevant themes, categories, and illustrative quotes concerning patient experience and the implementation of the AI technology.

To synthesize quantitative efficacy data alongside qualitative experiential findings, this review will use the Joanna Briggs Institute (JBI) convergent segregated approach to synthesis. Quantitative and qualitative data streams will be analyzed independently and then integrated during the narrative discussion to map intervention efficacy directly against human implementation factors. For quantitative data streams, a formal random-effects meta-analysis will be conducted only if a minimum of 3 independent primary studies exhibit clinical homogeneity—specifically evaluating identical stratified AI architectures (eg, predictive risk algorithms) and using uniform clinical indicators. If clinical or methodological heterogeneity is excessive, studies will be synthesized purely using qualitative narrative synthesis following the synthesis without meta-analysis (SwiM) reporting guidelines.

To ensure the integrity of the cumulative findings, the risk of bias due to missing results will be addressed at both the study and outcome levels. Publication bias will be formally assessed if a minimum of 10 studies are included in a meta-analysis for a specific outcome. Primary studies with irrecoverable variance data will be omitted from quantitative meta-analyses and synthesized exclusively within the narrative framework. For the assessment of publication bias via funnel plot asymmetry, the Egger test and Peters test will be used for continuous and binary outcomes, respectively. If significant asymmetry is detected, the “trim and fill” method will be used as a sensitivity analysis to estimate the potential impact of missing literature on the summary effect size.

Outcome reporting bias will be evaluated at the within-study level by comparing published results against original trial registrations (eg, ClinicalTrials.gov or PROSPERO). Any discrepancies, such as the omission of nonsignificant outcomes or the addition of nonregistered outcomes, will be documented and incorporated into the study’s overall risk-of-bias assessment. Furthermore, if numerical data essential for meta-analysis remain missing after author contact, conservative statistical imputation methods—such as substituting missing SDs with the maximum value observed in other studies—will be used in sensitivity analyses to test the robustness of the findings.

The overall certainty of the body of evidence for the primary outcomes will be evaluated using the GRADE (Grading of Recommendations Assessment, Development and Evaluation) methodology. Evidence will be categorized as “high,” “moderate,” “low,” or “very low” based on assessments of risk of bias, inconsistency, indirectness, imprecision, and publication bias. This systematic grading will provide a transparent framework for interpreting the strength of the review’s conclusions.