A detailed search strategy was implemented to locate relevant studies exploring the use of Artificial Intelligence (AI) in preventing and managing bleeding disorders. The search covered prominent databases, including PubMed, Science Direct, Google Scholar, and Wiley. Keywords and Medical Subject Headings (MeSH) terms were combined using Boolean operators (AND/OR) to enhance the search precision. The query included terms such as: (Bleeding Disorders OR Hemophilia OR Von Willebrand Disease OR Coagulopathy OR Clotting Disorders OR Thrombocytopenia OR Hemostasis disorders) AND (Artificial Intelligence OR AI OR ML OR Machine Learning OR Deep Learning OR Predictive Algorithms OR Predictive Modeling OR Computer-Assisted Diagnosis). Searches were finalized on 02/12/2024, with citations managed through Rayyan software to eliminate duplicates and facilitate initial screening (36).

The inclusion criteria for this systematic review were established using the PICO framework to ensure methodological rigor. Eligible studies focused on populations diagnosed with bleeding disorders, including hemophilia, von Willebrand disease, and other coagulopathies. The intervention of interest was the application of Artificial Intelligence (AI) techniques, such as machine learning, deep learning, and predictive algorithms, in the diagnosis, prevention, and management of bleeding disorders. Studies were required to include a comparison with traditional approaches, usual care, or no AI-based intervention. The review prioritized studies reporting outcomes related to improved patient care, including reductions in bleeding episodes, enhanced management of bleeding events, optimized dosing of treatments, early prediction and prevention of bleeding risks, and overall improvements in quality of life and healthcare efficiency. Only human studies employing randomized controlled trials, observational designs, cross-sectional studies, or cohort studies were considered for inclusion.

Exclusion criteria were defined to maintain the focus on high-quality, peer-reviewed evidence. Non-peer-reviewed literature, such as editorials, opinion pieces, conference reports, or abstracts, was excluded, along with case reviews, case series, review articles, and case reports. Studies written in languages other than English and those involving animal models were also excluded. These criteria were applied to ensure that the review synthesized robust and relevant evidence regarding the role of AI in the prevention and management of bleeding disorders.

The initial screening process was conducted systematically, beginning with a review of article titles, followed by an evaluation of abstracts. Each title and abstract were carefully assessed against the predefined inclusion and exclusion criteria. In the subsequent stage, full-text articles were subjected to a detailed review to ensure they addressed the use of Artificial Intelligence (AI) in the diagnosis, prevention, or management of bleeding disorders. Particular attention was given to studies that provided adequate scientific detail on AI techniques, their applications, and their impact on patient outcomes. This rigorous three-step screening process ensured the inclusion of studies that would contribute to a comprehensive and relevant dataset for understanding the role of AI in improving the diagnosis, prevention, and management of bleeding disorders.

Data were systematically extracted from each included study using a structured Microsoft Excel form to ensure a comprehensive and accurate capture of key information. The extracted data included details on study design, country of origin, and total sample size, as well as participant characteristics such as gender, age, and ethnicity. Inclusion and exclusion criteria were documented, encompassing symptoms, medical history, diagnostic methods, and other relevant factors. Specific information related to model development was also recorded, including data sources used, training and testing processes, and model performance metrics such as accuracy, precision, and sensitivity.

Additional variables extracted included key predictors identified by the models, the number of undiagnosed cases, and the main characteristics of these undiagnosed cases. Study limitations, outcome definitions, data processing methods, exploratory data analysis findings, and validation strategies were meticulously noted. Details on treatments administered to patients were also collected. This structured and systematic approach to data extraction ensured the inclusion of all relevant variables necessary for a comprehensive evaluation of the role of Artificial Intelligence (AI) in the diagnosis, prevention, and management of bleeding disorders.

To assess the risk of bias in the included studies, we utilized a variety of validated tools, each specifically designed for different study types, to ensure a thorough and consistent evaluation.

For observational studies, the Risk of Bias in Non-randomized Studies - of Exposures (ROBINS-E) tool was employed. This tool evaluates bias across multiple domains, including confounding, participant selection, exposure classification, deviations from intended exposures, missing data, outcome measurement, and reporting selection. Each observational study was reviewed using these criteria, enabling a detailed assessment of potential biases specific to non-randomized studies. The domains were rated as low risk (L), moderate risk (M), serious risk (S), critical risk (C), or no information (NI) (37).

The risk of bias assessment for randomized controlled trials (RCTs) included in this review was conducted using the Cochrane Risk of Bias 2 (RoB 2) tool, a rigorous and widely accepted framework for evaluating the methodological quality of RCTs. The assessment focused on five key domains: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in the measurement of outcomes, and bias in the selection of the reported result. Each domain was systematically evaluated, and studies were rated as having a low risk of bias, some concerns, or a high risk of bias based on predefined criteria (38).

A quantitative meta-analysis was not performed due to substantial heterogeneity across included studies. Variability existed in AI model architectures (such as Random Forest, XGBoost, Graph Neural Networks), predictor variables (genetic, clinical, laboratory, environmental), outcome definitions (for instance, short-term bleeding risk, inhibitor development, disease severity classification), and performance metrics (accuracy, AUROC, F1-score, PPV). This heterogeneity precluded valid statistical pooling, so a narrative synthesis approach was adopted.

The initial phase of screening the identified studies involved reviewing their Titles and Abstracts to assess relevance based on the defined PICOS criteria for this systematic review. The search yielded 2,927 records, which were imported into Rayyan software to streamline and organize the screening process. Rayyan automatically identified and removed 161 duplicate entries, leaving 2,766 unique records for evaluation.

The first stage of screening focused on assessing Titles and Abstracts based on the pre-established inclusion and exclusion criteria. This resulted in the exclusion of 2,714 studies that did not align with the review’s core focus on the use of Artificial Intelligence (AI) in the prevention and management of bleeding disorders. In the next phase, a more in-depth review was conducted for the remaining 52 abstracts. This stage involved examining the relevance of each study to AI applications in risk prediction, diagnostic advancements, and treatment optimization for bleeding disorders. Studies that did not directly address these topics were excluded, leaving 21 articles for full-text evaluation.

The final phase involved a careful full-text assessment to ensure compliance with the inclusion criteria. Nine studies were excluded due to inadequate focus or lack of relevant data, leaving a total of 12 studies for inclusion in the systematic review. This rigorous selection process ensured a reliable foundation for understanding AI applications in bleeding disorder management.

To provide a clear overview of the multi-stage review process and enhance methodological transparency, a PRISMA flowchart (Figure 1) was created, illustrating the progression from the initial search to the final selection of studies included in the review (25).

The risk of bias assessment for the included studies was conducted in two stages: separately for non-randomized studies and randomized-control studies.

The risk of bias assessment for the studies, presented in Table 2, was conducted using the ROBINS-E tool, which evaluates the quality of non-randomized studies based on seven domains: risk of bias due to confounding, bias arising from the measurement of exposure, bias in the selection of participants, bias due to post-exposure interventions, bias due to missing data, bias arising from the measurement of outcomes, and bias in the selection of reported results (37).

In the domain of risk of bias due to confounding, the majority of the included studies were rated as having moderate risk, likely due to incomplete adjustment for confounding factors or unclear reporting of control strategies (39–48). However, studies such as Ferreira et al., 202, and Sidonio Jr et al., achieved a low-risk rating, suggesting more rigorous confounder control (49, 50).

For bias arising from the measurement of exposure, most studies demonstrated a low risk, indicating reliable assessment methods (39, 43, 44, 46–48). However, moderate concerns were noted in some studies possibly due to measurement inaccuracies or unclear exposure definitions (40–42, 45, 50). The domain of risk of bias in the selection of participants was generally low risk, except for Aleksić et al., which was rated as moderate risk, potentially due to a small sample size from a single center that might introduce bias (46).

In the domain of bias due to missing outcome data, moderate risk was observed in several studies (41–49), often due to incomplete follow-up or inadequate reporting strategies. Notably, Hu et al., was rated as high risk (40), whereas studies like An et al., and Sidonio Jr et al., (39, 50) demonstrated low risk, indicating strong data management practices.

For bias arising from the measurement of outcomes, most studies were categorized as low risk, while studies such as Hu et al., Lopes et al., Ferreira et al., and Rawal et al., received a moderate risk rating, suggesting potential inconsistencies in outcome measurement (40, 41, 45, 49). Lastly, in the bias of selection of reported results, Rawal et al., Hu et al., Singh et al., Sidonio Jr et al., and Aleksić et al., were rated as moderate risk, indicating possible selective reporting that could exaggerate findings or omit key outcomes (40, 44–47). This analysis underscores the varying degrees of bias present across studies, emphasizing the need for careful interpretation, particularly for those with high overall risk ratings.

The risk of bias assessment for the randomized controlled trial (RCT) included in this review was conducted using the Cochrane Risk of Bias 2 (RoB 2) tool, which evaluates the quality of RCTs across five key domains: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in the measurement of outcomes, and bias in the selection of the reported result (38). The detailed results of this assessment are presented in Table 3.

The study demonstrates a moderate overall risk of bias (48). While baseline characteristics were comparable across groups, the randomization process and allocation concealment were not explicitly detailed, introducing potential bias. The trial’s open-label design further contributes to performance bias, as neither participants nor personnel were blinded; however, this is mitigated by the objective nature of the primary outcome, annualized bleeding rate (ABR). Missing outcome data were handled appropriately by scaling bleed counts during the available follow-up period, reducing the risk of attrition bias. Measurement of outcomes was objective and unlikely to be influenced by assessors, given the nature of the data collected. Lastly, the study appears to have reported all predefined outcomes without evidence of selective reporting. Despite these strengths, the lack of blinding and insufficient detail on randomization contribute to a moderate risk of bias overall.

This systematic review encompassed 12 studies investigating the application of Artificial Intelligence (AI) in bleeding disorders. The studies employed diverse methodologies, including retrospective and prospective cohort studies, randomized controlled trials, and feasibility studies. Conducted in geographically diverse settings, including the United States, United Kingdom, Spain, Germany, Brazil, Japan, Serbia, India, and China, the research provided a global understanding of AI’s role in bleeding disorder management. Observation periods varied significantly, with retrospective studies analyzing years of electronic health record (EHR) data and prospective trials evaluating AI interventions over months. Refer to Table 4 for a brief summary of all included studies.

The studies reviewed included diverse populations with varying demographic and clinical characteristics, reflecting the heterogeneous nature of bleeding disorders. The sample sizes ranged widely, from small cohorts with 96 participants (46) to extensive cohorts with over 23,000 individuals (40, 47, 50). Age distribution varied significantly across studies, with some focusing on pediatric populations (45) and others targeting older adults. Gender representation was often influenced by the disorder under investigation. Studies focusing on hemophilia A predominantly included males, given the X-linked inheritance of the condition. For example, the Chowdary et al., study consisted entirely of males (48), while studies on von Willebrand disease (VWD) and ITP included a more balanced gender distribution, such as Miah et al., which reported 53% male and 47% female participants among ITP patients (43).

Ethnic diversity was considered in some studies, particularly those conducted in multiethnic settings like the United States. For instance, Hu et al., analyzed data from the ATHN dataset, capturing Hispanic and non-Hispanic populations (40), while studies like Rawal et al., included racial categories such as White, Black, and Asian participants (48).

The studies employed a wide variety of machine learning algorithms, reflecting the complexity of bleeding disorders and their diverse datasets. Supervised learning models were the most commonly used, including Random Forest (39–41, 43–45, 50), XGBoost (40, 41), Gradient Boosting Machines (45, 50), CatBoost (40, 45), and Support Vector Machines (SVM) (41, 43, 44). These algorithms excelled in predictive and classification tasks, such as forecasting bleeding risks, identifying high-risk mutations, and optimizing prophylactic treatment regimens (39, 40). Logistic regression models were employed in several studies, particularly when the datasets had fewer predictors or were focused on well-defined clinical outcomes (42, 43, 48). In contrast to the above-mentioned studies, the study conducted by Sidonio Jr et al. in utilized a unary predictive model based on positive-unlabeled learning, which compared the characteristics of the diagnosed patient population to a potential undiagnosed population using a set of 12 key predictive variables (47).

Artificial intelligence (AI) has emerged as a valuable tool in the study of bleeding disorders, facilitating risk stratification, early diagnosis, and treatment optimization. Machine learning models have been employed across various bleeding disorders, including hemophilia, von Willebrand disease (VWD), immune thrombocytopenia (ITP), and cirrhosis-related variceal bleeding, to identify predictive patterns that enhance clinical decision-making.

Across bleeding disorders, key predictors consistently included:

Integration of these predictors into ML algorithms enhanced disease classification, early diagnosis, and personalized treatment strategies (Figure 2).

Across studies, Random Forest, XGBoost, and LightGBM consistently demonstrated strong predictive performance and interpretability. For immune thrombocytopenia (ITP), Random Forest achieved high accuracy, particularly in distinguishing ITP from other thrombocytopenic conditions. In hemophilia A prediction, LightGBM and Random Forest were most effective, with LightGBM achieving an F1-score of ∼0.99. Gradient Boosting Machines and Random Forest also performed well in von Willebrand disease (VWD) classification, identifying undiagnosed cases with high accuracy. For cirrhosis-related variceal bleeding, LogitBoost proved highly accurate, while broad population-level analysis favored Random Forest and CatBoost for target joint prediction in hemophilia patients.

The success of ensemble learning models underscores their ability to handle complex, multidimensional datasets. Techniques such as SHAP values and Lasso regression improved interpretability, while hyperparameter tuning enhanced predictive performance. Nonetheless, several limitations were evident across the literature. Common issues included class imbalance, missing data, and confounding, all of which may have influenced reported outcomes. External validation was limited, constraining generalizability. Deep learning models, although promising, demonstrated only moderate accuracy and remain difficult to implement clinically due to their “black box” nature. Addressing these limitations will require improved dataset quality, broader external validation, and the development of interpretable models suited for real-world settings.

The outcomes predicted by AI models also varied considerably. Short-term predictions included critical bleeding within weeks (An et al.) and treatment response over 12 weeks (Chowdary et al.). Long-term predictions encompassed inhibitor development, target joint formation, and severity classification. This range suggests that some models are better suited to acute decision-making, while others inform longitudinal management. Most studies relied on baseline predictors–demographics, genetics, or initial laboratory values–while only a few incorporated time-varying variables such as serial platelet counts, evolving comorbidities, or treatment adherence. Incorporating longitudinal data could substantially enhance clinical relevance.

Although a meta-analysis could have pooled diagnostic accuracy estimates, the heterogeneity of model architectures, predictors, outcomes, and metrics made statistical synthesis inappropriate. Instead, standardized reporting of AUROC, PPV, calibration, and consistent outcome definitions will be essential for future meta-analyses. Importantly, several studies were rated as moderate to high risk of bias–particularly due to missing outcome data, inadequate adjustment for confounders, or selective reporting–likely inflating reported accuracy. Conversely, findings from low-risk studies (Sidonio Jr et al.) more likely reflect true clinical performance. Overall, while results are promising, confidence is stronger in models with rigorous methodology and external validation.

Despite encouraging technical results, adoption remains slow. Challenges include fragmented datasets, poor interoperability across IT systems, regulatory hurdles, limited clinician familiarity, and concerns over transparency. Access inequities persist, especially in low-resource settings. Addressing these barriers will require coordinated efforts–multi-center registries, integration of interpretable AI into EHRs, and clinician training.

Heterogeneity was evident across disorders and study designs. In hemophilia, genetic and structural features dominated; in VWD, healthcare utilization and bleeding patterns were key; in ITP, platelet counts and comorbidities were critical; and in cirrhosis, clinical and endoscopic data were central. Shared predictors such as platelet count and factor VIII activity emerged across disorders. However, most models were developed in high-income countries, limiting insights for low-resource contexts where diagnostic gaps are greatest. Future work should test generalizability in these settings.

Machine learning approaches show promise for early diagnosis, risk stratification, and treatment optimization in bleeding disorders. Identifying undiagnosed cases–particularly in VWD and hemophilia–highlights potential for improved outcomes. Yet, most studies remain confined to retrospective design and internal validation. Prospective validation, randomized controlled trials of AI-assisted decision-making, and cost-effectiveness analyses are needed to guide real-world adoption.

Implementation should emphasize metrics with direct clinical meaning. For instance, PPV reflects the proportion of positive predictions that are correct and is critical in bleeding disorders, where false positives may lead to unnecessary factor replacement, invasive procedures, or prolonged monitoring. Sidonio Jr et al., reported PPVs of 75%–83%, which are encouraging but still insufficient for routine practice.

Practical integration requires embedding tools into clinical workflows. Examples include EHR-based risk alerts in ITP or dosing support in hemophilia prophylaxis. Barriers include interoperability, trust, regulation, and cost. Solutions may involve federated learning for data privacy, decision-support dashboards for clinicians, and structured training.

Future research should prioritize prospective, multi-center validation, registry-based adaptive trials, and integration of longitudinal and environmental data. Wearables and remote monitoring could enable personalized, real-time management. At the policy level, regulators must establish frameworks for AI adoption in rare hematologic disorders, ensuring safety while encouraging innovation.