This scoping review adhered to Arksey and O’Malley’s framework for scoping review [15] and the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines [16] (Multimedia Appendix 1). Our primary research question was: “What are the methodologies, clinical applications, and implementation challenges of causal discovery in observational medical research?” The aim was to synthesize evidence to guide future methodological development and clinical translation. Given the rapidly evolving nature of this field, no temporal restrictions were applied.

To comprehensively review studies of causal discovery algorithms in medical research and its subtypes, 2 researchers (XY and ZH) performed a systematic literature search of Scopus, Web of Science, PubMed, MEDLINE, Embase, and CINAHL in May 2025 and was restricted to studies published in English, using the following search strategies: “causal discovery/causal structure discovery/causal structure learning/causal graph discovery/learning causal models.” The search was focused on biomedical databases (PubMed, Embase, etc) as this review aims to map applications in medical research. While computer science–focused databases, such as IEEE Xplore or ACM Digital Library, include methodological developments, they primarily emphasize algorithmic innovation; seminal methodological papers with medical applications are typically also indexed within the selected biomedical sources. The search strategy was designed in consultation with an information specialist. The entire search strategy for the 6 databases is displayed in Multimedia Appendix 2.

Two researchers (HM and JM) independently screened studies based on predefined inclusion and exclusion criteria by examining their titles, abstracts, and full texts. Studies met the following inclusion criteria: (1) application focus: empirical implementation of causal discovery algorithms in medical or health contexts. (2) Data and methodology: the study applied a causal discovery approach within the observational research paradigm. This included analyses of real-world observational medical data as well as method-validation studies that used synthetic or benchmark datasets to address a medical research question or simulate a medical data environment. (3) Output: the study generated an explicit causal graph or causal structure. (4) Language: English-language publications. The exclusion criteria were: (1) The study purely focused on methodological or theoretical papers without medical applications. (2) Experimental (RCT) or basic science studies. (3) Duplicate publications, brief abstracts, or conference summaries lacking full methodological and results sections. (4) Studies where causal discovery was not the primary methodology. Discrepancies in study eligibility were resolved through consensus between 2 independent reviewers (JK and JL).

Screening was performed using Rayyan [17], with dual-reviewer verification at all stages. A standardized electronic form was used to extract the following data items: (1) study characteristics: year of publication, country or region of corresponding author affiliation, study design, and sample size; (2) causal discovery methods: specific algorithms, methodological category, and the primary software or package used for implementation; (3) data properties: data type, dimensionality, and temporal features; (4) application domain and specific clinical or public health objectives; and (5) validation approaches and key reported limitations.

Data synthesis was conducted in 2 complementary ways to align with the scoping review’s descriptive aims. First, a descriptive quantitative analysis summarized structured characteristics (eg, causal discovery algorithms, software tools, and application domains) using frequencies and percentages. Second, for textual data on reported methodological considerations and challenges, an inductive content summary was performed. Two reviewers (ZL and TL) independently identified recurrent issues and patterns, which were then discussed, consolidated, and organized into overarching descriptive themes (eg, methodological landscape and challenges). This process ensured the results reflected both the quantitative distribution and qualitative substance of the included literature. While this scoping review did not undertake a formal critical appraisal of individual studies using standardized quality assessment or risk-of-bias tools, we systematically extracted study-reported methodological limitations and implementation challenges as a prespecified data item during full-text screening. These author-reported limitations were subsequently subjected to inductive qualitative synthesis, alongside other extracted textual data, to identify recurrent patterns and to develop the descriptive themes presented in the “Challenges” section. This approach was intended to map commonly reported methodological issues in the literature rather than to appraise or grade the quality of individual studies.

After the removal of duplicates, 2296 unique records were identified for screening. Of these, 72 studies [8-14,18-82] (3.1%) met the inclusion criteria and were included in the final synthesis (Figure 1). After eliminating 2548 duplicates (52.6%), title or abstract screening excluded 1909 records (39.4%). A full-text review of 387 articles led to the exclusion of 315 (81.4%) due to reasons such as a methodological focus (201/315, 63.8%), duplicate publications or abstracts (50/315, 15.9%), and basic research or experimental designs (64/315, 20.3%).

The characteristics of the included studies (N=72) [8-14,18-82] are summarized in Table 1, covering publication year, country or region, and document type. The number of annual publications progressively increased from 2017 to 2025, with the years 2023-2025 accounting for the majority (46/72, 63.9%). Publications peaked in 2023 and 2024. The majority of studies were peer-reviewed journal articles (64/72, 88.9%), with the remaining studies being conference proceedings. The studies originated from 18 distinct countries or regions, with the United States contributing the largest proportion (25/72, 34.7%), followed by China (8/72, 11.1%), the Netherlands (6/72, 8.3%), and Japan (4/72, 5.6%), accounting for a combined total of 59.7% of publications. Three primary themes emerged from the analysis: (1) methodological landscape, that is, dominant causal discovery approaches and typologies in medical research; (2) application domains, that is, multidimensional implementation scenarios across observational studies; and (3) challenges, that is, methodological limitations and barriers to clinical translation.

This scoping review represents the first systematic mapping of causal discovery algorithms in observational medical research. It synthesizes key methodological, translational, and implementation gaps across clinical and public health domains. Despite the field’s nascent stage, research momentum has accelerated markedly, reflecting growing translational potential. The scarcity of qualifying studies (72 of 2296 screened) underscores both the emergent nature of these methods and the novelty of this review. Our integration of fragmented evidence establishes a foundational framework for interdisciplinary refinement, advancing beyond traditional observational research limitations and supporting the causal discovery’s role in medical settings.

Clinical research concentrates on etiological exploration, leveraging causal graphs to deconstruct complex diseases (eg, mental health and chronic conditions). However, a severe imbalance exists between mechanism discovery and clinical decision support development (1/72, 1.4%), revealing a critical mechanism-application dissociation. This translational gap primarily stems from 4 interconnected barriers: most methods remain at the proof-of-concept stage without large-scale clinical validation, creating a fundamental trust deficit; existing tools lack seamless integration with clinical workflows and EHR systems; real-world data quality issues and computational demands hinder practical implementation; and crucially, statistically robust findings often fail to translate into clinically actionable recommendations that clinicians can readily understand and apply at the point of care. Public health applications focus on behavior interventions (eg, lifestyle-outcome linkages) and system optimization (eg, health care use patterns), but face domain-specific confounders (eg, policy shifts). Cross-domain translation requires context-adapted methodological frameworks.

Causal discovery in observational medical research remains dominated by constraint-based (eg, PC and FCI) and score-based methods (eg, GES and FGES), valued for their reliability in exploratory analysis and high-dimensional data, as well as their inherent interpretability; their graphical outputs and testable assumptions align well with clinical reasoning. However, their limitations in identifying causal directions, handling unmeasured confounding, and computational efficiency have spurred new approaches: continuous optimization (eg, NOTEARS) enhances scalability, although often at the cost of transparency; hybrid methods (eg, GFCI) improve robustness; functional models (eg, LiNGAM) enable direction identification; and time-series approaches (eg, PCMCI) capture dynamic causality. This interpretability-scalability trade-off critically guides model selection in clinical settings, where trust and actionable insight depend on understandable outputs. This evolution marks a transition toward more robust, scalable, and temporally-aware causal inference.

Unmeasured confounding remains fundamental, distorting effect estimates and causal structures. Partial ancestral graph algorithms (eg, FCI and RFCI [91]) and negative control designs [92] offer partial solutions but lack universal robustness. Small-sample limitations undermine statistical power and generalizability, particularly in rare diseases or early-phase studies. Transfer learning [93], multimodal data fusion (eg, Cov-Pneum dataset [94]), and cross-disease validation mitigate data scarcity [95], while embedding expert knowledge (eg, in Bayesian networks) reduces model uncertainty. Beyond these challenges, translating causal discovery into practice requires concerted action on 2 fronts: establishing benchmark datasets and implementing a rigorous, explainable, multitiered validation framework. We strongly recommend the curation of medical benchmark datasets derived from gold-standard sources, such as large-scale RCTs or well-validated longitudinal cohorts, to provide a critical foundation for fairly evaluating and comparing algorithms. For instance, the study by Gururaghavendran and Murray [18] used a zero placebo effect as the benchmark to compare multiple causal discovery algorithms, demonstrating that guidance from expert insight and prior knowledge significantly enhances their performance.

Explainable artificial intelligence techniques hold significant potential for improving the causal discovery process. Hasan and Gani [96] introduced a knowledge-guided causal AI system that integrates domain-specific prior knowledge as structural constraints with observational data to refine causal graph discovery. The results show that this approach also lowers computational demands while improving reliability [96]. One study introduced a new causal discovery approach, called REX, leveraging ML models coupled with interpretability techniques, showing REX’s effectiveness and robustness in accurately recovering true causal structures, along with its applicability to real-world problems [97]. In addition, moving beyond theoretical sensitivity analyses, a robust validation strategy must integrate internal validation (eg, using bootstrap to assess graph stability), external validation on independent datasets, and most importantly, rigorous evaluation of biological and clinical plausibility against established knowledge or new experimental findings. This comprehensive approach is essential for bridging the current trust gap and generating clinically actionable insights.

To propel the field beyond current methodological constraints and accelerate clinical translation, the following strategic advancements warrant prioritized investigation:

Ultimately, the real‑world impact of causal discovery hinges on its ability to deliver comprehensible and actionable insights that resonate with clinical researchers and practitioners. However, whether causal discovery can genuinely drive scientific progress, generate novel insights, and inform scholarly discourse in fields can ultimately be determined only through its repeated application in practice, coupled with rigorous and transparent reporting of the results [101].

This review provides a timely, comprehensive synthesis of causal discovery algorithms across diverse medical domains. Nevertheless, limitations exist: First, restricting the search to English-language publications may have excluded valuable non-English evidence. Second, our review reveals a pronounced geographic concentration of research, with the majority of studies originating from the United States and China. This likely reflects disparities in access to large-scale digital health infrastructure, advanced computational resources, and targeted research funding. Consequently, the current evidence base and its associated methodological trends may not be fully representative of global research activity or directly generalizable to health care settings in other regions. Third, the search was conducted within biomedical databases and relied on a core set of terms centered on “causal discovery.” While this approach efficiently captured studies self-identifying with this paradigm, it is possible that some relevant applications published under different terminology (eg, “Bayesian network structure learning”) in computational science venues were not retrieved. Finally, while we mapped application breadth, a formal assessment of reporting quality or bias risk was not conducted; future work should implement standardized appraisal tools to systematically evaluate methodological rigor.

Causal discovery is emerging as a pivotal approach in medical observational research, offering transformative potential for etiological exploration and clinical decision-making. However, critical barriers impede clinical translation, notably the absence of standardized validation frameworks and persistent confounding. Future progress necessitates rigorous evaluation standards and AI-causal methodology integration to enable robust, generalizable real-world decision support.