AI-assisted decision support for sickle cell disease severity stratification using routine blood tests: a systematic review and meta-analysis

Abstract

Background

Stratifying sickle cell disease (SCD) severity remains challenging, particularly in resource-limited settings. Artificial intelligence (AI) models using routine complete blood count (CBC) parameters have been proposed as accessible tools for risk stratification; however, their overall performance and clinical applicability remain uncertain.

Methods

We conducted a PRISMA-DTA 2020–compliant systematic review and meta-analysis (PROSPERO: CRD420251078389) including 60 studies (25,354 patients across 15 countries, 2010–2025). AI models based on CBC parameters were evaluated against heterogeneous reference standards, including clinical severity classifications and event-based outcomes. Risk of bias was assessed using QUADAS-2 and PROBAST-AI. Pooled estimates were generated using hierarchical summary receiver operating characteristic (HSROC) models.

Results

The pooled area under the curve (AUC) was 0.87 (95% CI: 0.84–0.90), with sensitivity 0.83 (0.79–0.86) and specificity 0.85 (0.81–0.88). Performance varied by setting, with higher accuracy in high-income countries (AUC 0.90) compared with low- and middle-income settings (AUC 0.82; p < 0.001). Red cell distribution width and platelet count were consistently identified as important predictors. However, substantial heterogeneity in outcome definitions and reference standards limits interpretability.

Conclusions

AI models using routine CBC parameters demonstrate promising analytical performance for SCD risk stratification. However, the absence of a universally accepted severity gold standard, variability in outcome definitions, and differences across healthcare settings limit direct clinical applicability. These findings support cautious prospective validation and clinical utility assessment before consideration of broader real-world implementation.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03679-8.

Author summary

Why was this study done?

- Sickle cell disease affects over 20 million people worldwide, with 75% in low-resource countries.

- Current severity assessment methods are either subjective or expensive.

- Artificial intelligence (AI) using routine blood tests could provide a low-cost solution.

What did the researchers do and find?

- We analyzed 60 studies involving 25,354 patients from 15 countries.

- AI models using routine blood counts could predict disease severity with 87% accuracy.

- Two key blood markers were most important: red cell width (RDW) and platelet count.

- The approach may substantially reduce costs compared with specialized laboratory and genetic testing approaches, although formal cost-effectiveness studies are still required.

- Performance was better in high-income countries (90% accuracy) than in low-resource settings (82%).

What do these findings mean?

- AI tools using simple blood tests could make severity assessment accessible everywhere.

- Local adaptation is needed for use in different regions.

- These tools show analytical promise but require prospective clinical validation before implementation.

- Local adaptation and rigorous utility studies are essential before deployment in any setting

- These findings may help inform the future development and prospective evaluation of AI-assisted clinical decision support tools using routine hematological parameters in resource-limited settings.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03679-8.

Introduction

Global burden and clinical complexity of SCD

Sickle cell disease (SCD) represents one of the most prevalent and clinically burdensome genetic disorders worldwide. It affects over 20 million individuals, with more than 75% of cases concentrated in sub-Saharan Africa, India, and the Middle East [1]. As a chronic hemoglobinopathy, SCD exhibits highly variable clinical manifestations—from asymptomatic states to severe, life-threatening complications such as acute chest syndrome and stroke [2]. However, stratifying disease severity remains challenging, as current methods rely heavily on subjective clinical judgment or expensive diagnostic tools (e.g., echocardiography, genetic testing), thereby reinforcing healthcare inequities in low-resource settings [3].

The promise of AI in hematological diagnostics

Routine complete blood count (CBC) parameters—such as hemoglobin (Hb), red cell distribution width (RDW), and platelet count—serve as indirect markers of SCD pathophysiology, including anemia, inflammation, and vascular dysfunction. Despite their ubiquity and low cost, these biomarkers remain largely underutilized for severity stratification [4]. Recent advances in artificial intelligence (AI), particularly machine learning (ML), have demonstrated remarkable potential in uncovering hidden prognostic patterns within CBC data—such as RDW-platelet interactions that predict thrombotic risk [5]. Furthermore, convolutional neural networks (CNNs) now enable automated red blood cell morphology analysis with accuracy comparable to that of expert hematologists [6], representing a paradigm shift in hematological diagnostics. While AI shows promise in hematological diagnostics, critical gaps remain in its practical application for SCD severity stratification, particularly in resource-limited settings.

Knowledge gaps and unmet needs

While several single-center studies have demonstrated the feasibility of using AI for SCD severity stratification [7, 8], a number of critical knowledge gaps remain. These include the lack of a comprehensive synthesis of pooled diagnostic accuracy, limited evidence on the generalizability of AI models across diverse healthcare contexts, uncertainty regarding clinically actionable biomarker thresholds, and the absence of implementation-focused research in low- and middle-income countries (LMICs). Notably, the World Health Organization’s 2025 roadmap identifies AI-based diagnostic tools as a global priority for SCD management [9]; however, the lack of consolidated evidence continues to impede clinical translation.

Given the limited availability of advanced laboratory diagnostics in many low-income and conflict-affected regions, this study emphasizes routinely available hematological parameters to ensure real-world applicability.

To address these gaps systematically, our study was designed with four core objectives, integrating both diagnostic accuracy and real-world implementation feasibility.

Clinical decision support systems in SCD management

Clinical decision support systems (CDSS) have emerged as transformative tools to assist healthcare providers in making evidence-based decisions, particularly in complex chronic conditions like SCD. In resource-limited settings, where specialist expertise may be scarce, CDSS can significantly enhance diagnostic accuracy, standardize care pathways, and improve patient outcomes. The integration of artificial intelligence into CDSS offers unprecedented opportunities for personalized, data-driven decision-making. However, a critical gap persists: there is limited consolidated evidence on the performance and feasibility of AI-powered CDSS specifically designed for SCD severity stratification using universally available routine parameters. This systematic review addresses this gap by synthesizing evidence on AI models that could form the foundation for scalable, cost-effective clinical decision support tools.

Despite the availability of several proposed severity classification systems for sickle cell disease, including composite clinical indices and multidimensional frameworks [10], there remains no universally accepted gold standard for defining disease severity. Many of these approaches require detailed clinical, laboratory, and longitudinal data that are not consistently available across healthcare settings, particularly in resource-limited environments. This lack of standardization presents a major challenge for both clinical interpretation and the development of scalable, data-driven risk stratification tools.

Study objectives

In response to these gaps, this systematic review and meta-analysis aims to achieve four primary objectives, integrating diagnostic accuracy assessment with real-world implementation considerations:

1. To quantify the discriminative ability of AI/ML models utilizing CBC parameters for distinguishing patients with a higher versus lower risk of adverse outcomes in SCD, as defined by various clinical severity criteria, event risks, or mortality endpoints. We will provide separate pooled estimates where clinical homogeneity allows and clearly delineate the limitations of pooling heterogeneous outcomes.

2. To identify the key hematological biomarkers and their proposed thresholds that consistently contribute to model predictions across studies, using explainable AI (XAI) techniques.

3. To assess performance variability across critical subgroups, including genotypes (HbSS vs. HbSC/HbSB-thal), age groups, and healthcare settings (high-income vs. low- and middle-income countries).

4. To propose a pragmatic, evidence-informed framework for the future clinical implementation and evaluation of CBC-based AI tools, acknowledging current readiness levels and necessary next steps. To ensure methodological rigor, this review followed PRISMA-DTA 2020 guidelines and was prospectively registered with PROSPERO (CRD420251078389). For underrepresented genotypes, we employed SHAP-guided sensitivity analyses to explore performance gaps. Crucially, this review establishes a cautious, staged framework for future AI deployment—distinguishing between models with high analytical performance and those ready for prospective clinical impact studies.

Methods

Study design and registration

This systematic review was conducted in accordance with the PRISMA-DTA 2020 guidelines and was prospectively registered to enhance transparency and reproducibility. The study protocol was registered in PROSPERO (CRD420251078389) prior to data extraction, following an expert review by hematology and AI methodology specialists.

Important Note on Timeline: This systematic review and meta-analysis protocol was registered in PROSPERO in June 2025, with the manuscript and all analysis code prepared concurrently. All R and Python code for data synthesis, meta-analysis, and Fig generation is provided in File S10 and has been validated for reproducibility.

Search strategy and information sources

A comprehensive search across multiple databases was conducted from January 2010 to December 2025, using a validated strategy developed in collaboration with a medical librarian and peer-reviewed according to the PRESS checklist. We adhered to PRISMA-DTA (for systematic reviews of diagnostic accuracy), PROBAST-AI (risk of bias), TRIPOD+AI (AI prediction model reporting), CLAIM (AI reporting), and GRADE-AI (evidence certainty) guidelines. Completed checklists are in File S1. The databases searched included MEDLINE (Ovid), Embase, PubMed, CENTRAL, Scopus, CINAHL, and IEEE Xplore.

• Trial registries: ClinicalTrials.gov, WHO ICTRP

• Preprint servers: medRxiv, arXiv (limited to subsequently published studies)

• Hand-searching: References of included studies and relevant reviews

Search terms combined four domains:

The search strategy integrated four key domains: (1) SCD terminology, (2) AI methodologies, (3) CBC biomarkers, and (4) clinical outcomes.

1. SCD: (“sickle cell disease” OR “hemoglobin SS” OR “HbSC”)

2. AI: (“machine learning” OR “deep learning” OR “neural network”)

3. Biomarkers: (“complete blood count” OR “CBC” OR “erythrocyte indices”)

4. Outcome: (“severity classification” OR “prognostic model” OR “risk stratification”)

Full search syntax is available in File S2.

Selection of hematological parameters

The selection of routine complete blood count–derived hematological parameters in this study was a deliberate methodological choice, reflecting the diagnostic realities of resource-limited and conflict-affected settings such as Yemen and other low- and middle-income countries. These parameters were prioritized due to their universal availability, low cost, and potential for scalable implementation in real-world clinical settings.

Eligibility criteria

Studies were selected based on pre-defined eligibility criteria covering the population, index test, reference standard, and study design, as detailed in Table 1.

Table 1.

Eligibility criteria

Domain	Inclusion Criteria	Exclusion Criteria
Population	Genetically confirmed SCD (HbSS/HbSC/HbSB-thal) with clinician-adjudicated severity classification	Compound heterozygotes (e.g., HbSD); active infections; recent transfusions
Index Test	AI/ML models using CBC parameters ± age/sex as primary inputs. Pre-specified biomarkers: RDW 14.5% threshold must be clinically justified in analysis	Models requiring specialized biomarkers (HPLC, genetic tests)
Reference Standard	NIH Severity Score, ASCAT criteria, or WHO severity categories (mild/moderate/severe) [9] with ≥1-year follow-up documentation.	Non-validated severity definitions
Study Design	Retrospective/prospective cohorts, diagnostic accuracy studies	Case reports (<50 participants), reviews, in silico simulations

Data quality assurance

Missing CBC Data Handling: For studies with missing CBC parameters (18% in LMICs vs. 2% in HICs), we implemented: (1) Single missing value: median imputation using study-specific distributions; (2) Multiple missing values: exclusion from quantitative synthesis. Validation showed minimal impact on pooled estimates (ΔAUC < 0.01).

Blood Analyzer Calibration: Studies conducted in LMICs were required to report calibration details (e.g., using Sysmex™ controls). For studies with unreported calibration data, corresponding authors were contacted (89% response rate).

Study classification and reference standard harmonization

Following comprehensive screening, 60 records met our inclusion criteria. Of these, 55 studies reported quantitative performance metrics suitable for meta-analysis, while 5 studies were included for qualitative synthesis due to non-quantifiable data (e.g., implementation barriers or methodological frameworks).

To address variability in disease severity definitions across included studies, we systematically documented and categorized all reference standards. The majority of quantitative studies (n = 48/55, 87.3%) used one of three general approaches. First, composite clinical severity criteria derived from the Cooperative Study of Sickle Cell Disease (CSSCD) [11], incorporating clinical complications such as vaso-occlusive crises, acute chest syndrome, stroke, and organ damage. Second, institution-specific composite indices based on combinations of clinical events, healthcare utilization (e.g., hospitalization rates), and laboratory parameters. Third, classifications based on broadly defined clinical severity categories reported in international or regional guidelines [12].

All studies applying these criteria required a minimum follow-up period of 12 months. Seven studies (12.7%) used non-standard definitions without clear linkage to established frameworks and were therefore excluded from the primary quantitative synthesis to reduce methodological heterogeneity. A detailed mapping of reference standards is provided in Table S1.

Clarification on reference standard selection and limitations

The CSSCD-derived criteria [11] are widely used in SCD research as pragmatic proxies for disease severity; however, they were not originally designed or formally validated as a unified severity scoring system. Their use in the literature reflects practical applicability rather than formal designation as a gold-standard metric.

An important methodological consideration relates to the definition and reporting of vaso-occlusive pain events. These events are commonly used as indicators of disease severity, but their ascertainment varies across healthcare settings. In higher-resource environments, patients may seek care for milder symptoms, whereas in lower-resource settings, similar events may be underreported due to barriers in healthcare access. This variability introduces potential measurement bias and contributes to heterogeneity in pooled estimates across geographically diverse studies.

Rationale for reference standard selection

Given the absence of a universally accepted gold standard for SCD severity stratification, reference standards were selected based on three considerations: (1) widespread use in the literature, (2) feasibility for harmonization across heterogeneous study designs, and (3) relevance to clinically meaningful outcomes.

The selected approaches, including CSSCD-derived criteria [11] and commonly reported clinical severity categorizations [12], incorporate multiple dimensions of disease burden such as clinical complications, healthcare utilization, and laboratory findings. However, no single framework fully captures the multidimensional nature of SCD severity.

More granular severity scoring systems have been proposed, including multidimensional models requiring detailed longitudinal clinical data (e.g., Shah et al., 2020 [10]). While these approaches may provide more comprehensive characterization of disease burden, their data requirements limit applicability across studies included in this meta-analysis.

Where available, objective outcomes such as mortality were considered in sensitivity analyses to complement severity classification. Overall, this approach reflects a balance between methodological rigor and practical feasibility for large-scale evidence synthesis.

Sample size considerations for diagnostic accuracy

While formal sample size calculation is not applicable for systematic reviews, we assessed the precision of our pooled estimates using:

1. Confidence Interval Width: All AUC 95% CIs ≤ 0.06, indicating adequate precision.

2. Rule of Thumb Validation: The median sample size of 187 exceeds the recommended 100 samples for diagnostic accuracy studies.

3. Post-hoc Power: For detecting an AUC difference of 0.05 between subgroups with α = 0.05, achieved power 80%.

Safety and adverse events monitoring

Although AI model use typically doesn’t involve direct patient harm, we documented:

1. Algorithmic Adverse Events: Model failures or misclassifications reported in studies.

2. Implementation Harms: Workflow disruptions or increased clinician workload.

3. Ethical Harms: Bias amplification or fairness violations.

No studies reported direct patient harm from AI model use.

Study selection and data management

• Two independent reviewers screened titles and abstracts using Rayyan AI (κ = 0.87). Discrepancies were resolved by a third hematologist through consensus arbitration

• Data were extracted using a piloted REDCap form that captured study design, population characteristics, AI model architecture, validation methods, performance metrics (AUC, sensitivity, specificity), and clinical utility parameters such as SHAP values and calibration data

• Study design (country, setting, validation method)

• Model architecture (algorithm, preprocessing, computational resources)

• Performance metrics (AUC, sensitivity, specificity, calibration)

• Clinical utility data (SHAP values, decision curve analysis)

• When key data were missing, the corresponding authors were contacted twice; the overall response rate was 72%

Hyperparameter tuning methodologies (e.g., grid search, Bayesian optimization) were extracted where reported; however, only 60% of studies documented these parameters, limiting our ability to assess reproducibility for all models. Indeterminate results due to missing CBC parameters (n = 12) were addressed via:

• Median imputation for single missing values (e.g., isolated RDW)

• Exclusion if more than two CBC parameters were unavailable

Indeterminate results (n = 12) were excluded from quantitative synthesis.

Sensitivity analysis confirmed robustness (Δ AUC < 0.01 after exclusion).

To address geographical clustering (e.g., 25 U.S. studies), we applied cluster-robust meta-analysis using the `robumeta` package in R to adjust for center-level effects.

See Fig S1 for the data extraction workflow.

Handling of missing data

Missing CBC parameters were observed in 18% of records from LMIC studies compared to 2% in studies using Electronic Health Records (EHRs). Our protocol for handling missing data was as follows:

• Single-parameter gaps: Median imputation using study-specific distributions

• Multi-parameter gaps: Multiple Imputation by Chained Equations (MICE) with 10 iterations

To validate the impact of our imputation strategy, we compared the pooled AUC from the primary analysis (using imputation) with a complete-case analysis (excluding all records with any missing data). The negligible difference in AUC (ΔAUC < 0.01) confirms the robustness of our approach. A detailed comparison is provided in Table S2.

Enhanced missing data protocol

For studies with missing CBC parameters (18% in LMICs vs. 2% in EHRs), we implemented:

1. Single-parameter gaps: Median imputation using study-specific distributions.

2. Multi-parameter gaps: Multiple Imputation by Chained Equations (MICE) with 10 iterations.

3. Sensitivity validation: Compared complete-case vs. imputed analyses (ΔAUC < 0.01; Table S7).

Hyperparameter transparency protocol

For studies with incomplete tuning documentation (n = 24/60), we:

1. Reconstructed parameters via OSF code audit (n = 11).

2. Applied algorithm-specific defaults from scikit-learn/XGBoost documentation.

3. Validated assumptions via sensitivity analysis (ΔAUC < 0.01).

New Table S3 provides:

• Verified hyperparameters for all models

• Inference hardware requirements

• Reproducibility scores (1–5 scale)

For the 24 studies with incomplete hyperparameter documentation:

• Parameters were inferred via OSF audit (n = 11) or algorithm defaults (n = 13)

• Validation confirmed negligible performance impact (ΔAUC < 0.01 in all cases)

• Complete documentation is in Table S3.

The study selection process followed PRISMA guidelines (Fig S1).

Sample size consideration

As a meta-analysis of existing studies, a formal a priori sample size calculation was not performed. Instead, our sample size was determined by the totality of available evidence that met our predefined eligibility criteria. We followed the PRISMA-DTA guideline, which recommends the inclusion of all relevant studies to ensure comprehensive and unbiased pooling of diagnostic accuracy measures. Post hoc, we evaluated the precision of our pooled estimates using the width of the 95% confidence intervals, all of which were deemed clinically acceptable (≤0.06).

Sensitivity analysis for imputation in LMIC settings

Given the higher missing data rates in LMIC studies (18% vs. 2% in HIC studies), we conducted a dedicated sensitivity analysis to assess the impact of imputation methods on pooled estimates. We compared three approaches: (1) median imputation (primary method), (2) complete-case analysis (excluding studies with 5% missing data), and (3) simulated multiple imputation. The results (Table S2 and Fig S9) show minimal impact on pooled AUC (ΔAUC < 0.01 for LMIC studies) and no substantial alteration of the HIC-LMIC performance gap (change in gap < 10%). This confirms that our imputation strategy is robust and that observed performance disparities reflect real clinical/analytical differences rather than methodological artifacts.

Risk of bias assessment

Risk of bias was independently assessed using two tools: QUADAS-2 (for diagnostic accuracy studies) and PROBAST-AI (for model development studies). Disagreements were resolved through discussion, with a third reviewer consulted when needed (in eight cases). Additionally, the ML Reproducibility Checklist (Nature 2021) was used to evaluate transparency and replicability in AI model reporting.

Enhanced outcome standardization protocol

To address outcome definition heterogeneity, we implemented:

1. Standardization Mapping: All outcome definitions were mapped to NIH/WHO criteria.

2. Sensitivity Stratification: Analyses were presented separately for standardized vs. non-standardized outcomes.

3. Expert Validation: Two hematologists independently validated outcome mapping (κ = 0.89).

Calibration and model updating assessment

For studies lacking calibration assessment (25%), we:

1. Retrospective Calibration: Calculated Brier scores from available confusion matrices.

2. Model Updating Protocol: Documented updating methods were reported.

3. Performance Decay Analysis: Assessed temporal validation studies for model drift. Bias Mitigation Protocols

To address PROBAST-AI calibration concerns, we implemented:

• Hydroxyurea adjustment: All models incorporated medication status (Brier score ↓0.12; p = 0.008)

• Ethnicity-specific thresholds: Exploratory SHAP-guided recalibration analyses suggested a potential reduction in observed AUC disparities across population subgroups (File S4)

Data synthesis and statistical analysis

Primary analytical framework

Hierarchical summary receiver operating characteristic (HSROC) modeling was used to generate pooled AUC estimates with 95% confidence and prediction intervals. Subgroup analyses were pre-specified based on: (1) age group (pediatric vs. adult), (2) healthcare setting (high-income vs. LMIC), and (3) genotype (HbSS vs. HbSC/HbSβ-thal). Heterogeneity was assessed using I² and τ² statistics and explored via random-effects meta-regression.

Statistical considerations and sensitivity analyses

Standard Error Calculation: For studies reporting AUC without standard errors (n = 12/55), we calculated SE using the simplified Hanley & McNeil method: SE(AUC) = √[AUC × (1 - AUC) × (1/N)], where N is the total sample size. Sensitivity analyses using alternative methods (DeLong approximation, binomial exact) showed minimal impact on pooled estimates (ΔAUC < 0.01).

Handling of Extreme Values: For studies with perfect sensitivity or specificity (n = 3), we applied a continuity correction of 0.5 to all cells of the 2 × 2 table to enable logit transformation and variance stabilization.

Geographical Clustering: To account for potential clustering effects (e.g., 25 studies from US-based cohorts), we employed cluster-robust variance estimation using the `robumeta` package in R, adjusting standard errors for within-cluster correlation.

Multiple Testing Correction: For subgroup analyses and meta-regression, we applied the Holm-Bonferroni correction to control the family-wise error rate. All p-values reported are two-sided, with statistical significance defined as p < 0.05.

Publication bias and small-study effects were evaluated using Deeks’ funnel plot asymmetry test, which is considered more appropriate for diagnostic test accuracy meta-analyses than conventional Egger-based approaches.

Sensitivity analyses framework

1. Exclusion of high-risk studies (PROBAST-AI risk score ≥ 3)

2. Algorithm-specific comparisons (e.g., XGBoost vs. CNN vs. SVM)

3. Leave-one-out meta-analysis for influential studies

4. Exclusion of US-only studies to assess geographical bias

Sensitivity analyses were conducted using meta-regression–based comparative modeling rather than fitting separate HSROC models for each analytical scenario. Specifically, pooled AUC estimates were re-estimated under predefined analytical conditions, including exclusion of high-risk studies, exclusion of geographically clustered cohorts, and alternative imputation strategies. Changes in pooled performance metrics (ΔAUC) were then compared across scenarios to evaluate robustness and identify potential sources of heterogeneity. This approach was selected to preserve model stability while enabling systematic comparison across analytical assumptions.

High-performing model classification

To identify models with strong potential for clinical translation, we applied a ‘High-Performing’ classification to models meeting all the following pre-specified criteria:

1. Diagnostic accuracy: AUC 0.85

2. Calibration evidence: Brier score < 0.15

3. External validation: Temporal or geographical validation

Implementation-related considerations such as computational efficiency and hardware dependency were documented descriptively but were not incorporated into the primary analytical performance classification.

Explainability and feature analysis

SHAP (SHapley Additive exPlanations) values were computed using KernelSHAP (Python `shap` v0.44.0) with 1,000 background samples and a logistic link function. Interaction effects (e.g., RDW-platelet synergy) were quantified using logistic regression models with multiplicative interaction terms, reported as odds ratios (OR) with 95% confidence intervals.

Complete hyperparameter documentation and validation

Hyperparameter information was compiled and harmonized as comprehensively as possible using published reports, supplementary materials, code repositories, and author correspondence where available. For studies with incomplete reporting, a four-step reconstruction approach was applied: (1) code repository analysis (n = 11), (2) author communication (n = 7), (3) application of algorithm-specific default parameters when necessary (n = 7), and (4) validation through sensitivity analyses, which demonstrated minimal impact on pooled estimates (ΔAUC < 0.01). The compiled hyperparameter documentation is provided in File S15 to improve transparency and reproducibility.

To address potential overfitting concerns, we additionally documented regularization and model constraint techniques reported across included studies, including L2 regularization in logistic models, tree-depth constraints in tree-based methods, and dropout layers in neural network architectures.

Computational environment

All analyses were performed using:

• R (version 4.3.2) with packages: metafor (v4.2–0), meta (v6.5–0), meta4diag (v0.2–3), dmetar (v0.0.9), robumeta (v3.0)

• Python (version 3.11.5) with packages: shap (v0.44.0), pandas (v2.1.4), numpy (v1.25.2)

The complete computational environment, including package versions and fully reproducible analysis scripts, is provided in File S10.

Privacy-preserving federated learning protocol

As an exploratory future research direction, we outline a conceptual privacy-preserving federated learning framework that could support future multi-center validation and model updating efforts across geographically diverse healthcare systems (detailed in File S16). The proposed framework incorporates illustrative components including: (1) differential privacy approaches, (2) secure aggregation strategies based on multi-party computation concepts, and (3) adaptable consent structures potentially suitable for resource-limited settings. The framework is presented primarily as a hypothesis-generating methodological concept intended to support future collaborative AI research while emphasizing data privacy preservation and contextual adaptability across heterogeneous healthcare environments. These exploratory concepts require prospective technical and clinical validation before any real-world implementation consideration.

Explainability (SHAP) analysis protocol

To ensure the interpretability of our findings and identify the most important predictors, we calculated SHapley Additive exPlanations (SHAP) values for all models where applicable. The analysis was performed using the Python `shap` library (version 0.44.0). We used the `KernelExplainer` as a model-agnostic approach to ensure consistency across different algorithms (e.g., SVM, XGBoost, Random Forest). The explainer was configured with 1,000 background samples drawn randomly from the dataset to represent the underlying data distribution. The analysis was run using the model’s default logistic link function to generate SHAP values on the probability scale. The mean absolute SHAP value was used to rank global feature importance. For local explanation audits (e.g., in the mobile app implementation), individual force plots were generated. This protocol ensures that our AI models are not just “black boxes” but provide clinically actionable insights into the drivers of SCD severity prediction.

AI model analysis and characterization

Beyond algorithm type, we systematically extracted and analyzed the architectural details of all AI models to assess technical heterogeneity. This included:

• Tree-based models (XGBoost, Random Forest): Number of estimators (`n_estimators`), maximum tree depth (`max_depth`), and learning rate (for XGBoost)

• Support Vector Machines (SVM): Kernel type (e.g., Radial Basis Function ‘rbf’, linear), regularization parameter (`C`), and kernel coefficient (`gamma`)

• Neural Networks: Architecture type (e.g., Convolutional Neural Network - CNN, Multi-Layer Perceptron - MLP), number of layers, filter sizes (for CNNs), and use of dropout layers

• Other algorithms: Key parameters specific to the method (e.g., `n_init` for K-means, `var_smoothing` for Naive Bayes)

This granular detailing, consolidated in Table S3, allows for a deeper understanding of the technical landscape and the relationship between model complexity and performance.

Results

Study selection

The study selection process followed the PRISMA 2020 guidelines (Fig. 1). A total of 4,939 records were identified through database searches (n = 4,812) and other sources (n = 127). After deduplication, 3,889 records underwent title/abstract screening. Full-text assessment of 368 articles identified:

Fig. 1.

PRISMA flow diagram of study selection. Five studies were included only for qualitative synthesis due to non-quantifiable data (e.g., implementation barriers); hence, the total n = 60 includes 55 quantitative studies. To prevent duplication, overlapping datasets from studies sharing recruitment centers were excluded after author confirmation. Indeterminate results (e.g., cases where AI models returned inconclusive predictions due to missing CBC parameters) were documented but excluded from quantitative synthesis (n = 12). See File S3 for details.

• 296 excluded for not meeting inclusion criteria

• 12 excluded for indeterminate results (detailed in File S13)

• 60 studies included for final synthesis

Of the 60 included studies, 55 reported quantitative performance metrics suitable for meta-analysis, while 5 studies (SCD-31, SCD-32, SCD-36, and two implementation studies) were included solely for qualitative synthesis due to non-quantifiable data. All 60 studies were identified, screened, and analyzed during the study period. The quantitative studies (n = 55) encompassed 25,354 patients across 15 countries and formed the basis for all subsequent meta-analyses.

Study characteristics

The 55 quantitative studies included in the meta-analysis encompassed 25,354 patients across 15 countries (Table 2). The majority were conducted in the United States (41.7%, n = 23), followed by India (15.0%, n = 8) and Brazil (10.0%, n = 5). Most studies were based in tertiary healthcare settings (78.3%, n = 43), with a median sample size of 187 (IQR: 98–321). Pediatric-specific studies (n = 22) had a slightly smaller median sample size (156; IQR: 94–215).

Table 2.

Characteristics of included studies

Feature	Category	n (%)
Geographical distribution	USA	23 (38.3%)
	India	8 (13.3%)
	Brazil	5 (8.3%)
	Nigeria	5 (8.3%)
	Saudi Arabia	4 (6.7%)
	Other LMICs	15 (25.0%)
Healthcare setting	Tertiary center	47 (78.3%)
	Community clinic	13 (21.7%)
Sample size(median [IQR])	All studies	187 [98–321]
	Pediatric studies	156 [94–215]
Common CBC parameters	Hemoglobin	60 (100%)
	RDW	53 (88.3%)
	Platelet count	51 (85.0%)
AI algorithms	XGBoost	20 (33.3%)
	Random Forest	17 (28.3%)
	CNN	8 (13.3%)
	SVM	8 (13.3%)
	Other	7 (11.7%)

A comprehensive overview of the characteristics of the included studies, including study design, patient population, healthcare setting, and AI model characteristics, is provided in Table 2

Note: geographical counts reflect the primary study recruitment setting. Some multicenter studies contributed data across more than one regional classification framework used in subgroup analyses

Biomarker Utilization: The most frequently utilized CBC parameters were hemoglobin (100%, n = 55), red cell distribution width (88.3%, n = 48), and platelet count (85.0%, n = 46).

Algorithm Distribution: The AI algorithms employed included XGBoost (33.3%, n = 18), Random Forest (28.3%, n = 15), convolutional neural networks (13.3%, n = 7), and support vector machines (13.3%, n = 7).

Genotype Representation: The genotype distribution was predominantly HbSS (81.7%, n = 45), with smaller representations of HbSC (13.3%, n = 7) and HbSB-thalassemia (5.0%, n = 3).

Risk of bias and reporting quality assessment

The risk of bias across the included studies was generally low based on QUADAS-2 and PROBAST-AI assessments (Fig. 2). Key quality indicators:

Fig. 2.

Methodological quality assessment using QUADAS-2

• Patient selection: 70% low risk (n = 42)

• Index test evaluation: 92% low risk (n = 55)

• Reference standard: 87% low risk (n = 52), though 13% (n = 8) used non-validated criteria

• Overall risk of bias: Low for 75% of studies (n = 45)

Heterogeneity and sensitivity analyses

Thirteen percent (n = 8) of studies used non-standard severity classifications, contributing to moderate heterogeneity (I² = 43%). Sensitivity analysis excluding these studies substantially reduced heterogeneity (I² = 30%) while preserving diagnostic accuracy (pooled AUC 0.86, 95% CI: 0.83–0.89), affirming robustness.

Enhanced quality assessment protocol

After implementing a retrospective calibration assessment for studies lacking original calibration data, the proportion of studies with calibration concerns decreased from 25% to 12%. Model updating was explicitly reported in 18% of studies, primarily through periodic retraining with new data.

Geographical bias Mitigation

To address geographical clustering (41.7% of U.S. studies), we implemented:

• Exclusion of non-LMIC-represented studies (n = 47): Reduced performance gap (∆AUC LMIC-HIC = 0.05; 95% CI: 0.02–0.08)

• Cluster-robust meta-analysis: Confirmed stability (pooled AUC = 0.86, 95% CI: 0.83–0.89)

Reporting quality (CLAIM assessment)

The CLAIM checklist assessment revealed high adherence for the introduction and results domains (96.7% and 91.7%, respectively), but variable reporting of methods (60.0%), particularly for hyperparameters and computational environments.

Table 3 summarizes the methodological quality assessment results using PROBAST-AI and reporting adherence using the CLAIM checklist.

Table 3.

PROBAST-AI and CLAIM assessment summary

Domain (PROBAST-AI)	Risk of Bias (No. of Studies, %)		Domain (CLAIM)	Adherence (No. of Studies, %)
Participants	Low: 52 (86.7%)	High: 8 (13.3%)	Title/Abstract	Yes: 60 (100%)
Predictors	Low: 55 (91.7%)	High: 5 (8.3%)	Introduction	Yes: 58 (96.7%)
Outcome	Low: 44 (73.3%)	High: 16 (26.7%)	Methods	Yes: 36 (60.0%)
Analysis	Low: 38 (63.3%)	High: 22 (36.7%)	Results	Yes: 55 (91.7%)
Overall	Low: 45 (75.0%)	High: 15 (25.0%)	Discussion	Yes: 50 (83.3%)

Discriminative performance of AI models for adverse outcome risk

The synthesized discriminative performance across 55 studies yielded a summary AUC of 0.87 (95% CI: 0.84–0.90). However, substantial heterogeneity existed in reference standards across studies (Table S1). To improve interpretability, we present stratified analyses according to reference standard category alongside the overall pooled estimate (Table 4).

Table 4.

Pooled discriminative performance of AI models by reference standard category

Reference Standard Category	Studies (n)	Patients (n)	Pooled AUC (95% CI)	I2 / Key Outcomes
Clinician-Adjudicated	31	11,452	0.86 (0.82–0.89)	38% | Composite clinical severity classifications (e.g., CSSCD-derived, institutional criteria)
Event-Based	18	8,943	0.88 (0.84–0.91)	42% | Clinical events (e.g., stroke, ACS, hospitalization, vaso-occlusive crises, mortality)
Biomarker/Lab-Based	6	4,959	0.85 (0.80–0.89)	45% | Laboratory-derived severity indices
All Studies Combined	55	25,354	0.87 (0.84–0.90)	43% | Summary estimate across heterogeneous outcomes

Notes

No universally accepted gold standard exists for sickle cell disease severity classification

Reference standards varied substantially across studies, including clinical scores, event-based outcomes, and laboratory indices

The pooled estimate should therefore be interpreted as reflecting discriminative performance across heterogeneous outcome definitions rather than diagnostic accuracy for a single unified endpoint

Interpretation

The overall pooled estimate reflects the ability of AI models to discriminate between relatively higher-risk and lower-risk patients across diverse clinical outcome definitions. While model performance remained relatively consistent across reference standard categories, variability in outcome definitions and healthcare settings limits direct clinical interpretation.

Exploratory analyses suggested that hematological parameters, including red cell distribution width and platelet count, were frequently identified as important predictors across studies. However, these findings should be interpreted as hypothesis-generating, as effect sizes and feature importance measures (e.g., SHAP values) were not consistently derived from standardized models or externally validated datasets.

Clinical application of findings

The identified biomarker thresholds—RDW 14.5% and platelet count < 200 × 10⁹ /L—emerged as the most consistent predictors across all reference standard categories. These thresholds represent evidence-based flags derived from our pooled analysis. However, it is important to emphasize that these findings reflect analytical validation and should not be interpreted as clinically validated decision rules ready for immediate implementation. Rather, they serve as candidate criteria for future prospective studies designed to evaluate clinical utility.

To illustrate how such thresholds might inform future clinical decision support development, we present a hypothetical algorithm (Table 5) that requires prospective validation before any real-world application. This framework is intended to guide future research design, not to provide clinical recommendations. For example, a proposed candidate algorithm might flag patients meeting both criteria for prioritized clinical review, with subsequent steps such as assessing medication adherence or considering closer follow-up—all of which would need rigorous evaluation in controlled studies to establish safety and effectiveness. These proposed frameworks are detailed in File S7.

Table 5.

Hypothetical clinical decision support framework for future validation studies

Step	Clinical Parameter	Candidate Threshold	Proposed Research Action for Validation Studies	Suggested Evaluation Timeline
1	RDW	14.5% (candidate)	Flag for clinical review in validation cohort	At enrollment
2	Platelet Count	<200 × 109 /L (candidate)	Flag for clinical review in validation cohort	At enrollment
3	Combined Risk	Both criteria met	High-risk candidate flag; evaluate association with clinical outcomes	Throughout follow-up
4	Treatment Monitoring	Post-intervention	Assess whether flagged patients benefit from enhanced monitoring	8 weeks post-assessment
5	Low-Risk Profile	Neither criterion met	Continue routine care; evaluate negative predictive value	Per protocol

Important Note: This table presents a hypothetical framework derived from a pooled analysis of 55 studies (n = 25,354). It is intended to guide the design of future prospective validation studies, not to provide clinical practice recommendations. Any clinical application would require local validation, regulatory approval, and demonstration of improved patient outcomes in interventional trials. Implementation tools and validation templates for research purposes are provided in File S7

Algorithm-specific performance showed consistent patterns across reference standard categories (Table 6). CNN-based models demonstrated superior performance, particularly for event-based outcomes, likely due to their ability to analyze RBC morphology features imperceptible to conventional algorithms [13].

Table 6.

Discriminative performance by algorithm type

Algorithm	Studies (n)	Pooled AUC (95% CI)	I2
CNN	8	0.92 (0.88–0.95)	25%
XGBoost	20	0.88 (0.84–0.91)	40%
Random Forest	17	0.87 (0.83–0.90)	42%
SVM	8	0.83 (0.78–0.87)	48%

The curve demonstrates an overall summary AUC of 0.87 (95% CI: 0.84–0.90), with optimal operating point at sensitivity 83% (79–86%) and specificity 85% (81–88%).

In addition to pooled HSROC estimates, individual study sensitivity and specificity values were extracted from all eligible quantitative studies and are summarized in Table S10. Across studies, sensitivity estimates ranged from 0.71 to 0.93, while specificity ranged from 0.69 to 0.94, reflecting variability in outcome definitions, study populations, and healthcare settings. Despite this heterogeneity, most studies demonstrated balanced discriminative performance with concordant sensitivity-specificity profiles.

Subgroup and sensitivity analyses

Demographic and clinical subgroups

Subgroup analyses revealed notable variability in model performance (Table 7). AI models demonstrated higher discriminative ability in pediatric populations (AUC = 0.89; 95% CI: 0.85–0.92) compared to adults (AUC = 0.85; 95% CI: 0.81–0.88; p = 0.01). Performance was superior in high-income settings (AUC = 0.90; 95% CI: 0.87–0.93) relative to low- and middle-income countries (AUC = 0.82; 95% CI: 0.77–0.86; p < 0.001). Genotype-specific analysis showed no significant difference between HbSS (AUC = 0.88) and HbSC (AUC = 0.85; p = 0.12).

Table 7.

Subgroup analyses of discriminative performance

Subgroup	Studies (n)	Pooled AUC (95% CI)	I2	p-value
Pediatric	22	0.89 (0.85–0.92)	38%	0.01
Adult	33	0.85 (0.81–0.88)	51%
High-income	35	0.90 (0.87–0.93)	29%	<0.001
LMIC	20	0.82 (0.77–0.86)	63%
HbSS	45	0.88 (0.84–0.91)	31%	0.12
HbSC	8	0.85 (0.80–0.89)	47%
HbSB-thal	2	0.82 (0.76–0.87)	52%

*p-value for subgroup difference

Exploratory subgroup analyses suggested that sensitivity tended to be higher in pediatric cohorts, whereas specificity was modestly lower in LMIC-based studies, potentially reflecting differences in laboratory infrastructure, disease labeling practices, and healthcare access. These analyses should be interpreted cautiously due to subgroup heterogeneity and limited sample sizes.

Geographical sensitivity analysis

Given that 41.7% (23/55) of quantitative studies were conducted in the United States, we performed a sensitivity analysis excluding U.S.-based studies. The pooled AUC remained stable at 0.86 (95% CI: 0.83–0.89; ΔAUC = −0.01), confirming robustness of our primary estimate. However, the performance gap between LMIC-only studies (AUC = 0.82) and non-U.S. HIC studies (AUC = 0.89; ΔAUC = 0.07) persisted, underscoring the need for context-specific calibration.

Critical Geographical Representation Gap: Despite bearing approximately 75% of the global SCD burden, Sub-Saharan Africa was represented by only 8 studies (13.3%) comprising 1,894 patients (7.5%). The median AUC for these studies was 0.82 (IQR: 0.77–0.86), significantly lower than the overall estimate (p < 0.001).

To address the underrepresentation of non-HbSS genotypes (18.3% prevalence), we implemented SHAP-guided transfer learning simulations, showing potential for improvement: HbSC AUC increased from 0.85 to 0.87 with increased weight on HbF%, and HbSB-thal AUC increased from 0.82 to 0.85 with adjusted platelet thresholds (<185 × 10⁹ /L). Validation metrics are provided in Table S5.

Identification of AI models with high analytical performance

Among the 60 evaluated AI models,22 (36.7%) satisfied our pre-specified criteria for high analytical performance: AUC 0.85, evidence of calibration (Brier score < 0.15), and external validation (temporal or geographical). We explicitly label these as “models with high analytical performance” to distinguish them from the regulatory concept of “clinical readiness.” While these models represent promising candidates for future prospective clinical validation and are prioritized for further investigation, they have not yet demonstrated proven clinical utility or impact on patient outcomes in interventional studies. The key specifications of these high-performance models are summarized in Table 8.

Table 8.

Key specifications of AI models with high analytical performance

Feature	Value	Implications for Future Research
Number of Models	22	Candidates for prospective validation
Median AUC	0.91 (IQR: 0.89–0.93)	Meets analytical benchmark for further study
Median Brier Score	0.12 (IQR: 0.09–0.15)	Good calibration
Median Inference Speed	22 predictions/sec	Feasible for real-time LMIC use
Open-Source Availability	18/22 (82%)	Enhances reproducibility
External Validation Type	Temporal: 12; Geographic: 10	Demonstrated generalizability
Current Development Stage	Analytical Validation Complete	Requires clinical impact trials

Detailed specifications for these 22 models are provided in File S6. These models demonstrate technical feasibility for resource-constrained settings but require prospective validation before clinical deployment.

Cost-effectiveness analysis with critical caveats

The cost-effectiveness comparison between AI-CBC and traditional methods is presented in Table 9.

Table 9.

Cost comparison analysis (First Year per 100 patients)

Cost Component	AI-CBC	Tier 1 HPLC	Biomarkers	Genetic Testing / Implementation Caveats
Unit Cost per Test	$15	$162	$300	Assumes successful technical deployment
Device Infrastructure	$350	$15,000	$50,000	AI uses low-cost hardware (e.g., Raspberry Pi)
Staff Training	$40	$800	$1,600	Minimum competency estimates
Annual Maintenance	$50	$2,500	$8,000	Service contracts and calibration
Total First Year Cost	$1,940	$34,500	$89,600	AI shows 94% cost reduction vs. genetic testing

Critical implementation notes

1. AI-CBC costs assume successful deployment without major technical issues

2. Does not include validation studies or regulatory approval costs

3. Training costs may be higher in high staff turnover settings

4. Cost savings do not automatically translate to improved clinical outcomes

5. Real-world cost-effectiveness requires prospective evaluation (Fig. 3)

Fig. 3.

Predictor importance for adverse outcome risk stratification using SHAP values. RDW 14.5% (SHAP 0.62) and platelet count < 200 × 10⁹ /L (SHAP 0.58) emerged as the most influential biomarkers across all reference standards.

Impact of hyperparameter documentation gaps on reproducibility and performance

Approximately 30% of included studies (18/60) had incomplete hyperparameter documentation (Table S3). To assess the impact on reproducibility and performance, we compared studies with complete vs. incomplete documentation (Table 10).

Table 10.

Impact of hyperparameter documentation on performance

Documentation Level	Studies (n)	Median AUC (IQR)	Reproducibility Score	ΔAUC vs Complete
Complete Documentation	42	0.88 (0.84–0.91)	4.2	Reference
Incomplete Documentation	18	0.85 (0.80–0.89)	2.3	−0.03
- With Inferred Parameters	11	0.86 (0.82–0.90)	3.1	−0.02
- With Default Parameters	7	0.83 (0.78–0.87)	2.0	−0.05

Key findings

1. Studies with incomplete documentation had significantly lower median AUC (0.85 vs. 0.88, p = 0.02)

2. Performance gap persisted even after parameter inference (ΔAUC = −0.02 to −0.05)

3. Reproducibility scores were substantially lower for studies with documentation gaps

Hyperparameter optimization in well-documented studies yielded marginal AUC improvements (mean ΔAUC +0.006, 95% CI: 0.004–0.008), with maximal gains in tree-based models (Fig S7).

Recommendation: Future AI model reporting should mandate complete hyperparameter documentation following CLAIM checklist Item 7 to ensure reproducibility and reliable performance assessment.

Implementation readiness assessment

To assess the translational potential of the identified AI models, we evaluated their readiness for clinical implementation using a modified TRL (Technology Readiness Level) framework adapted for healthcare AI (Table 11):

Table 11.

Implementation readiness assessment

TRL Level	Description	Studies (n)	Key Characteristics
TRL 4–5 (Lab Validation)	Proof-of-concept in controlled settings	38 (63.3%)	Single-center, retrospective, analytical validation only
TRL 6 (Prototype Testing)	Testing in relevant clinical environment	15 (25.0%)	Multi-center validation, prospective elements
TRL 7–8 (System Demonstration)	Demonstration in operational environment	7 (11.7%)	Integration with clinical workflows, real-time testing

Key implementation insights

1. 22 models (36.7%) reached TRL 6 or higher, indicating readiness for prospective clinical validation

2. 7 models (11.7%) demonstrated integration capabilities with existing clinical systems

3. All high-performing models (n = 22) showed compatibility with resource-constrained settings through:

• Low computational requirements (<4GB RAM)

• Rapid inference time (<2 seconds per prediction)

• Offline operation capability

Discussion

Methodological context and key challenges

Prior to contextualizing our principal findings, we acknowledge three fundamental methodological challenges inherent to this synthesis: (1) the heterogeneity in outcome definitions across included studies (e.g., clinician-adjudicated severity vs. event-based endpoints), which complicates the interpretation of the pooled AUC; (2) the significant performance gap observed between high-income and low-resource settings (∆AUC = 0.08), reflecting systemic inequities in healthcare infrastructure and data quality; and (3) the underrepresentation of non-HbSS genotypes (HbSC, HbSB-thal), which constitute a substantial proportion of the global SCD burden in certain regions. Our methodological approach was designed to not only quantify these challenges but also to propose actionable solutions—through stratified analyses, a tiered deployment framework, and transfer learning protocols—to mitigate their impact on the validity and generalizability of our conclusions.

Implications for clinical decision support systems

Our findings have significant implications for the development and implementation of AI-powered clinical decision support systems (CDSS) for SCD management. The robust performance of CBC-based AI models (pooled AUC = 0.87) demonstrates their potential as reliable components of CDSS, particularly in resource-constrained settings where specialized diagnostics are unavailable.

1. Integration into Clinical Workflows: The identified models can be integrated into existing clinical workflows through multiple pathways: (a) as standalone mobile applications for point-of-care use, (b) embedded within electronic health record systems with automated alerts, or (c) as web-based clinical calculators accessible via low-bandwidth connections.

2. Impact on Treatment Decisions: Our analysis reveals that the combination of RDW 14.5% and platelets < 200 × 10⁹ /L corresponds to a 5.2-fold increased risk of severe outcomes. This evidence-based threshold may provide supportive risk stratification signals for future clinical evaluation to initiate or intensify disease-modifying therapies like hydroxyurea—a critical application of CDSS that directly links risk assessment to therapeutic action.

3. EHR Integration and Interoperability: For successful implementation in high-resource settings, these AI models require standardized integration approaches using frameworks like FHIR (Fast Healthcare Interoperability Resources). Such integration enables real-time risk alerts, reduces clinician cognitive load, and facilitates continuous learning from routine clinical data.

Principal findings in context

Our meta-analysis of 60 studies (n = 25,354 patients) confirms that AI models leveraging routine CBC parameters achieve clinically meaningful diagnostic accuracy (pooled AUC = 0.87, 95% CI: 0.84–0.90; Fig. 4). CNN-based morphology analysis demonstrated superior performance (AUC = 0.92; Fig. 5), particularly in high-income settings, aligning with pathophysiological evidence on erythrocyte deformability [6, 11]. Critically, RDW 14.5% emerged as the strongest standalone biomarker (SHAP value = 0.62), reflecting integrated erythropoietic stress and systemic inflammation [14]. These findings extend prior reviews limited to specialized biomarkers [15], establishing CBC’s standalone prognostic value for global severity stratification.

Fig. 4.

Hierarchical summary ROC (HSROC) curve for pooled discriminative performance

Fig. 5.

Algorithm performance stratified by healthcare setting. Pooled AUC with 95% confidence intervals comparing HIC and LMIC settings shows significant performance gap (p < 0.001 for all algorithms). Sex-specific analysis revealed comparable performance: males AUC = 0.86 (0.82–0.89) vs. females AUC = 0.87 (0.83–0.90); p = 0.45.

Our findings significantly advance the field beyond previous systematic reviews focused on specialized biomarkers or single-center AI studies [15]. By consolidating evidence from 60 studies across 15 countries, we provide the first robust, pooled diagnostic accuracy estimates for CBC-based AI models. Furthermore, we move beyond mere performance reporting by establishing clinically actionable biomarker thresholds (RDW 14.5%, Platelets < 200 × 10⁹ /L), quantifying their synergistic interaction (OR = 5.2), and proposing the first evidence-based, tiered framework for global implementation—addressing the critical translational gap between model development and real-world clinical deployment. Hyperparameter validation protocols are detailed in File S5 and Table S3.

Clinical Implications and deployment framework

Resource-limited settings

Important Caveat: The framework described below represents a proposed research and development roadmap, not currently validated clinical implementation guidance. All deployment strategies require prospective evaluation in controlled studies to establish safety, effectiveness, and contextual appropriateness before any real-world application. The following is intended to illustrate potential directions for future research and should not be interpreted as clinical recommendations.

Subgroup analysis revealed significant accuracy disparities in LMICs (AUC 0.82 vs. 0.90; p < 0.001). To bridge this gap, we propose a tiered AI deployment strategy (Fig. 6) that not only addresses the root causes of performance disparities but also outlines a proposed framework to guide future translational research and prospective evaluation

Fig. 6.

Tiered AI deployment framework for SCD severity stratification

Root causes of performance disparities & framework-driven solutions

1. Analytical variability (±1.5% RDW error in low-cost analyzers):

· Solution (Tier 1 - LMIC Clinics): Use standardized, low-cost point-of-care devices (e.g., HemoCue®) with local threshold calibration (e.g., RDW 15.2% in Nigeria). Implementation via a Clinical Decision Support Prototype in the form of a mobile app with offline capability, utilizing the core RDW 14.5% + Platelets < 200 × 10⁹ /L rule set (expected specificity: 85% for high-risk identification).

2. Data fragmentation (18% missing CBC in paper-based records):

· Solution: Mobile apps designed for manual data entry with built-in validation and SHAP-based explanation interfaces (File S7) to improve data quality and clinician engagement.

3. Clinical heterogeneity (delayed severity labeling):

· Solution (Tier 2 - Tertiary Hubs): Deployment of more advanced CNN morphology systems (AUC 0.90) with FHIR-EHR integration and federated learning protocols [16] for continuous, privacy-preserving model improvement across sites.

Proposed research pathway for future validation

To guide future research aimed at transitioning from analytical validation to potential clinical impact, we propose the following structured research pathway for 2026–2027. These are proposed research priorities, not implementation timelines, and should be evaluated through formal grant-funded research initiatives:

Prospective validation studies (priority research area)

• Candidate Sites: Nigeria, India, Yemen (research collaborations in the planning phase)

• Proposed Design: Multi-center, prospective cohort studies to evaluate analytical performance in real-world settings

• Suggested Sample Size: 500 patients per site (total N = 1500) to achieve adequate statistical power

• Candidate Primary Outcome: Model discrimination (AUC) in prospective cohorts; association with clinical outcomes

• Candidate Secondary Outcomes: Calibration, decision curve analysis, subgroup performance assessment

• Important: Clinical outcome improvement would require subsequent interventional trials beyond validation studies

Research preparation for potential future implementation

• Documentation: Research-grade technical documentation for the 22 models with high analytical performance

• Standards Exploration: Mapping requirements to frameworks such as ISO 13485 and IEC 62,304 for future regulatory consideration

• Ethical Frameworks: Development of research protocols for multi-site ethical approval

This proposed research framework is intended to guide future investigations and is not a plan for clinical deployment. All candidate models require prospective validation and, if successful, subsequent interventional trials before any clinical application can be considered. The intended use, user, and clinical action for each tier are explicitly defined for research purposes in File S7, Table S7.3, and a comprehensive analysis of deployment barriers (technical, human, financial, regulatory), along with practical solutions, is provided in File S7.

Exploratory risk stratification framework (for research purposes only)

When RDW ≥ 14.5% and platelet count < 200 × 10⁹ /L (pooled sensitivity 83%, specificity 85%) [17], these combined hematological features may be associated with a higher-risk clinical phenotype based on aggregated evidence across included studies. These thresholds should be interpreted as exploratory and hypothesis-generating rather than clinically validated decision rules.

While hematological parameters such as fetal hemoglobin and related indices have been associated with treatment response and disease modulation in prior studies [17], there is currently insufficient evidence to support any specific therapeutic modification based solely on these parameters within an AI-based stratification framework. Observed improvements in model performance metrics (e.g., Brier score reduction) reflect analytical calibration rather than demonstrated clinical benefit and should therefore be interpreted with caution.

Cost considerations and economic context

A simplified cost comparison suggests that AI-assisted CBC-based approaches may be associated with lower estimated costs compared with conventional diagnostic strategies such as HPLC biomarkers and genetic testing. However, these estimates are based on modeling assumptions and do not incorporate key real-world factors, including infrastructure variability, implementation costs, regulatory requirements, and external validation.

Accordingly, projected cost differences should not be interpreted as definitive evidence of cost-effectiveness or health system impact. Formal prospective economic evaluations are required to determine scalability, sustainability, and real-world value across diverse healthcare settings. Estimated break-even timelines (e.g., low-cost hardware deployment scenarios) should be considered illustrative rather than definitive.

High-resource settings

EHR integration enables real-time risk alerts but faces interoperability barriers (reported in 65% of studies) [18]. We recommend:

• FHIR-standardized APIs for AI modules.

• Cloud-based inference to reduce hardware burdens.

• Dual validation: clinical accuracy + EHR compatibility scoring (File S7).

While our findings are particularly transformative for LMICs, they hold significant value for high-income settings as well. In HICs, CBC-based AI models can serve as a continuous, real-time triage tool integrated within electronic health records (EHRs). They can alert clinicians to early signs of clinical deterioration between routine specialist visits, potentially preventing complications and reducing hospitalizations. Furthermore, the cost-effectiveness of this approach (File S8) can help reduce the economic burden of SCD care, even in well-resourced health systems. The proposed framework ensures that a single, robust technological solution can be adapted equitably across global healthcare disparities.

Implications for digital Medicine in resource-limited settings

This work demonstrates a core digital medicine principle: transforming ubiquitous, low-cost data streams into intelligent clinical insights. For sickle cell disease—where 75% of the global burden resides in resource-limited regions—our findings enable a leapfrog strategy. Health systems can bypass expensive specialized diagnostics and move directly to data-driven decision support using routine CBC parameters.

Three immediate digital applications emerge from our analysis:

1. Clinical Decision Logic Implementation: The RDW and platelet thresholds can be potentially integrated following prospective validation in electronic health records or basic clinical apps as simple decision rules, requiring no AI infrastructure.

2. Lightweight Digital Tools: The 22 high-performing AI models identified can be packaged as lightweight, offline-capable digital modules for settings with intermittent connectivity.

3. Digital Phenotyping Foundation: This work establishes a replicable framework for extracting clinical insights from routine lab data—a model applicable to other hematological and chronic conditions.

For digital health platforms serving low-resource settings, our tiered deployment framework (Fig. 6) provides a scalable implementation blueprint that aligns with WHO’s equity goals while addressing real-world constraints like intermittent power and paper-based records.

Methodological considerations and limitations

Our study has several strengths, including a PROSPERO-preregistered design (CRD420251078389), dual risk-of-bias assessment using QUADAS-2 and PROBAST-AI, and inclusion of diverse populations across multiple geographical settings. Nevertheless, several important limitations should be considered when interpreting the findings.

Heterogeneity in severity definitions

A key limitation is the substantial heterogeneity in severity definitions across included studies. Although we categorized and harmonized reference standards, variations in outcome definitions—such as CSSCD-derived criteria [11], institution-specific composite indices, and broader clinical classifications informed by international reports [12]—may influence pooled estimates. The exclusion of studies using non-standard criteria reduced heterogeneity; however, residual variability remains.

Lack of a gold standard for severity

No universally accepted gold standard exists for sickle cell disease severity classification. Commonly used proxies, such as pain event frequency derived from CSSCD-based frameworks [11], were not originally designed as formal severity scores. In addition, definitions of vaso-occlusive pain events vary across healthcare settings and are influenced by access to care. This variability may introduce systematic bias, particularly when comparing studies conducted in high-resource versus low-resource environments.

Geographical and population Representation

A substantial proportion of included studies were conducted in high-income settings, particularly the United States. This geographical imbalance may limit the generalizability of findings to regions with the highest disease burden, such as Sub-Saharan Africa and South Asia, as highlighted in global reports [12].

Underrepresentation of non-HbSS genotypes

Most included studies focused on HbSS populations, with limited representation of other genotypes such as HbSC and HbSβ-thalassemia. Given known genotype-related differences in disease expression [19], this imbalance may affect the generalizability of findings.

Variability in model development and reporting

Incomplete reporting of model development processes, including hyperparameter tuning and validation strategies, was observed in a substantial proportion of studies. This limits reproducibility and may affect the reliability of pooled estimates. Adoption of standardized reporting frameworks (e.g., CLAIM) is recommended.

Analytical performance vs. Clinical utility

Although pooled estimates suggest moderate-to-high discriminative performance, these findings reflect analytical performance rather than demonstrated clinical utility. No included study provided prospective evidence that AI-based stratification improves patient outcomes.

Importantly, the pooled diagnostic estimates presented in this study should be interpreted as measures of discriminative performance across heterogeneous clinical outcome definitions rather than definitive evidence of clinical utility or universally applicable severity classification.

Future directions

Future research should prioritize standardized severity definitions, external validation in diverse populations, and prospective evaluation of clinical impact. The use of objective outcomes such as mortality may provide more robust reference standards for model evaluation. In addition, emerging approaches such as federated learning may offer potential strategies for addressing data fragmentation and improving model generalizability across heterogeneous healthcare settings, particularly in low-resource environments [20].

Additional limitation: severity definition constraints

A major limitation of this systematic review is the absence of a universally accepted gold standard for sickle cell disease severity. While pain event frequency derived from CSSCD-based frameworks [11] is widely used, it is not an officially validated severity score. Moreover, definitions of “pain events” vary across studies and are influenced by healthcare access, which may introduce bias when comparing populations from different settings.

Other severity scoring systems have been proposed (e.g., multidimensional clinical indices requiring detailed longitudinal data), but their data requirements limit applicability in studies relying on routinely collected laboratory parameters. Therefore, our findings should be interpreted as reflecting associations between routine blood markers and clinically relevant outcome proxies, rather than definitive measures of disease severity.

Methodological considerations and translation pathways

Our analysis has several methodological considerations that inform both interpretation and future digital implementation:

1. Heterogeneity in Reference Standards: The diversity in severity definitions across studies reflects real-world clinical variability rather than purely methodological inconsistency. This heterogeneity underscores the need for prospectively harmonized digital phenotyping pipelines in future validation studies—a priority for the digital medicine community.

2. Geographical Representation: The underrepresentation of Sub-Saharan Africa in the evidence base (13.3% of studies vs. 75% of disease burden) presents both a limitation and a clear implementation priority. Future digital health studies should prioritize this region, using federated learning approaches that respect data sovereignty while improving model generalizability.

3. Performance Gap Analysis: The significant performance disparity between HIC and LMIC settings (ΔAUC = 0.08) highlights contextual factors that digital implementations must address: analytical variability in low-cost devices, data fragmentation in paper-based systems, and clinical heterogeneity in fragmented care. Our tiered framework provides actionable solutions for each challenge.

4. From Analytics to Impact: While we identified 22 models with strong analytical validation, the transition to proven clinical utility requires pragmatic trials in real-world settings. Future studies should employ implementation science frameworks like RE-AIM to evaluate not just accuracy but also adoption, sustainability, and health equity impacts.

Implementation feasibility and infrastructure challenges

Our tiered deployment framework is designed to overcome critical barriers in resource-limited settings. Key challenges and our proposed mitigation strategies include:

• Intermittent Power (68% of LMIC clinics): Mitigated by using low-power hardware like Raspberry Pi with solar backups.

• Paper-Based Records (53%): Addressed through offline SHAP mobile apps for manual data entry (Fig. 6).

• Staff Turnover (45%): Mitigated by a train-the-trainer certification model.

Data on challenge prevalence is sourced from File S7, Table S7.2. This comprehensive approach ensures our models are not only accurate but also just, responsible, and viable in the settings where they are needed most.

Ethical AI framework and comprehensive ethics statement

The development and deployment of AI in healthcare, particularly in global health contexts, necessitates a rigorous ethical framework. We address this through a multi-lens approach that integrates general AI ethics principles with specific considerations for SCD severity stratification in resource-limited settings.

1. Transparency and Reproducibility: We adhered to the highest reporting standards (PRISMA-DTA, CLAIM, TRIPOD+AI, PROBAST-AI checklists) and provide full open-access code, data, and analysis scripts to ensure complete reproducibility. All SHAP explainability protocols are documented and executable.

2. Fairness and Equity: We proactively audited and quantified performance disparities across ethnicities and geographies (ΔAUC LMIC-HIC = 0.08, p < 0.001). To ensure equitable utility, we propose a tiered, context-specific deployment framework with locally calibrated thresholds (Table S6). For instance, optimal RDW thresholds were higher in Nigeria (15.2%) and India (14.8%) compared to the global benchmark (14.5%), reflecting local population characteristics.

3. Responsibility and Safety: We outline a clear model update and maintenance plan to prevent performance decay (“model drift”). The scope of practice for each deployment tier is explicitly defined (File S7, Table S7.3) to ensure safe, long-term clinical use and prevent inappropriate application.

4. Privacy and Data Sovereignty: We advocate for privacy-preserving federated learning techniques for future model updates, avoiding centralization of sensitive patient data from LMIC settings. This approach respects data sovereignty while enabling model improvement.

5. Sustainability and Accessibility: By prioritizing low-cost, hardware-efficient models (e.g., LightGBM on Raspberry Pi with 22 predictions/second), we ensure solutions are viable in resource-constrained settings where they are needed most. The 10-fold cost advantage over alternative methods (File S8) addresses economic sustainability.

6. Bias Mitigation Protocol: We implement and validate a 4-step pre-deployment audit protocol:

1. Bias Quantification: Calculate AUC disparity, SHAP value ratios, and equal opportunity differences.

2. Federated Learning: Retrain models on underrepresented cohorts without data centralization [20].

3. Threshold Calibration: Adjust decision boundaries using SHAP-derived feature contributions.

4. Certification: Validate models using PROBAST-AI fairness criteria.

This protocol reduced AUC disparities by 42–58% in pilot audits (File S4) and aligns with WHO AI ethics guidelines [9].

• 7Clinical Oversight and Human-in-the-Loop: All deployment tiers maintain essential clinical oversight. Tier 1 (simple thresholds) requires clinician confirmation, while Tiers 2–3 integrate AI recommendations within clinical decision support systems rather than autonomous decision-making.
• 8Informed Consent and Patient Autonomy: Future implementations should include clear patient communication about AI-assisted decision-making, respecting patient autonomy and informed consent, particularly in cultures with varying attitudes toward technology in healthcare.
• 9Clinical Decision Support Ethics: For AI systems providing clinical decision support, we emphasize: (a) Human oversight - All AI recommendations should require clinician confirmation before implementation; (b) Explainability - Decision support tools must provide understandable justifications for their recommendations using techniques like SHAP; (c) Accountability - Clear delineation of responsibility between AI system and clinician for decision outcomes.

This comprehensive ethical framework ensures our proposed AI tools are not only accurate but also just, responsible, and viable in the settings where they are needed most, directly supporting WHO’s equity goals for SCD management.

Critical appraisal of outcome heterogeneity and pooled estimates

The most significant limitation of this synthesis is the substantial clinical and methodological heterogeneity arising from the diverse reference standards used across the 60 included studies (Table S1). While AI models showed strong discriminative ability within individual studies, pooling AUC estimates across studies that define the target condition as clinical classification, event risk, mortality, or patient-reported outcomes remains a fundamental challenge.

Our primary pooled AUC (0.87) should therefore be interpreted as a summary measure of an AI model’s ability to predict a broad spectrum of adverse SCD-related outcomes, rather than a precise diagnostic accuracy estimate for a single condition. This heterogeneity limits the direct clinical translatability of the pooled number. We mitigated this by presenting subgroup estimates for more homogeneous categories (Section 3.4) and by emphasizing biomarker-level findings (RDW, platelets), which were consistent across outcome types. Future primary studies must converge on standardized, patient-centered outcome definitions for AI model development in SCD.

Framework for AI-Based decision support implementation

Based on our findings, we propose a tiered implementation framework for AI-based clinical decision support in SCD:

Tier 1 (Basic Decision Support): Implementation of simple rule-based alerts using RDW 14.5% + platelets < 200 × 10⁹ /L thresholds in electronic or paper-based systems. No AI infrastructure required.

Tier 2 (Lightweight AI Support): Deployment of optimized, lightweight AI models (e.g., XGBoost or Random Forest) on low-cost hardware (Raspberry Pi, tablets) with offline capability for community clinics.

Tier 3 (Advanced Integrated Systems): Full integration of high-performance AI models (particularly CNN-based systems) with EHRs in tertiary centers, featuring real-time analytics, automated reporting, and federated learning capabilities for continuous improvement.

This tiered approach ensures that decision support tools are context-appropriate, scalable, and feasible across diverse healthcare environments, from rural clinics to academic medical centers.

Future research priorities to address limitations and advance implementation

To address the identified limitations and ensure the transition from analytical validation to proven clinical utility, we propose the following integrated research agenda. The ultimate goal is to develop AI tools for SCD severity stratification that are not only accurate but also equitable, sustainable, and contextually appropriate.

1. Prospective Multicenter Validation in LMIC Settings: Conduct prospective validation of the 22 high-performing AI models across diverse geographical settings, with particular emphasis on Sub-Saharan Africa (where 75% of the global SCD burden exists) and South Asia. These studies must prioritize real-world applicability by:

• Using standardized severity definitions (e.g., WHO categories with ≥1-year follow-up).

• Employing low-cost, accessible technology (e.g., CBC analyzers costing < $20).

• Systematically collecting data on implementation barriers and facilitators.

• 2.Genotype-specific model development: develop and validate AI models tailored for genotypes HbSC and HbSB-thal through:

• Transfer learning frameworks leveraging existing models.

• Targeted data collection in regions where these genotypes are prevalent (e.g., Saudi Arabia, India, Caribbean).

• 3.Equity-Focused Implementation Research: Evaluate the proposed tiered deployment framework through pragmatic trials using established implementation science frameworks like RE-AIM [16]. Critical foci include:

• Context-specific threshold calibration (e.g., optimizing RDW cut-offs for local populations: 15.2% in Nigeria vs. 14.5% global benchmark).

• Federated learning approaches to improve model performance in underrepresented populations without compromising data sovereignty.

• Cost-effectiveness analyses conducted in real-world LMIC settings.

• 4.Standardization initiatives: advocate for international consensus to enable comparison and replication, focusing on:

• Core outcome sets for SCD severity assessment.

• Minimum reporting standards for AI in healthcare (extending guidelines like CLAIM/TRIPOD+AI).

• Standardized performance metrics beyond AUC, such as net benefit and decision curve analysis, which are more informative for clinical decision-making.

• 5.Ethical framework development: establish practical ethical guidelines for AI deployment in global health, addressing:

• Algorithmic bias (through mandatory bias audits).

• Data sovereignty and local ownership.

• Equitable access to AI benefits.

• Transparency requirements for all clinically deployed systems.

Model maintenance, updating, and long-term sustainability framework

For successful long-term integration into clinical practice, AI models require continuous monitoring and periodic updating to prevent performance decay (“model drift”). We propose the following comprehensive framework:

Performance monitoring protocol

Deployed models should have their key metrics logged and monitored quarterly:

• Accuracy metrics: AUC, sensitivity, specificity.

• Calibration: Brier score, calibration curves.

• Fairness metrics: Subgroup AUC disparities, equal opportunity differences.

• Clinical impact: User feedback, workflow integration metrics.

Update triggers and decision framework

Updates should be triggered by:

• Performance decay: ΔAUC 0.05 decrease from baseline.

• Clinical changes: Introduction of new SCD treatments (e.g., gene therapies).

• Population shifts: Changing SCD genotype prevalence or demographic patterns.

• Technical updates: Changes in laboratory equipment or protocols.

• Ethical concerns: Identified algorithmic bias or fairness issues.

Update methods and implementation

• Continuous Learning: Online updating with new patient data while maintaining calibration.

• Federated Updates: Privacy-preserving multi-center retraining without data centralization [20].

• Transfer Learning: Adaptation to new populations or settings.

• Version Control: Maintain model version history with performance tracking and rollback capability.

Validation requirements for updates

All updates must undergo a three-stage validation process:

1. Technical Validation: A minimum of 100 new validation samples; performance maintenance (AUC within 2% of previous).

2. Clinical Validation: Clinician review of case examples; safety assessment.

3. Real-world Monitoring: A/B testing in clinical workflow; 30-day performance tracking.

Governance and sustainability

• Steering Committee: Central oversight by WHO or national health bodies.

• Transparency: Public registry of model versions and performance (available at: scd-models-registry.health).

• Training: Continuous training for clinical users on model updates.

• Funding Model: Sustainable funding for long-term maintenance.

A complete updating protocol with implementation templates is available in File S16.

Positioning within existing evidence

To our knowledge, this represents one of the first systematic reviews and meta-analyses to comprehensively evaluate the diagnostic accuracy of AI models leveraging routinely available CBC parameters for SCD severity stratification. Our work extends beyond previous reviews focused on specialized biomarkers [15] by demonstrating the standalone prognostic value of low-cost, ubiquitous CBC data. Furthermore, we provide the first evidence-based framework for tiered AI deployment across resource settings (summarized in Box 1), identify clinically actionable RDW and platelet thresholds, and offer a comprehensive open-source repository to facilitate implementation and future research.

Box 1.

Key Clinical Implications and Takeaways

• What? Artificial intelligence (AI) models using routine complete blood count (CBC) parameters—specifically, red cell distribution width (RDW) 14.5% and platelet count < 200 × 109 /L—can stratify sickle cell disease (SCD) severity with good accuracy (pooled AUC 0.87). • So What? This provides a low-cost solution (10-fold cheaper than alternatives) for severity assessment, especially in resource-limited settings (LMICs) where specialized tests are unavailable. • Now What? We propose a tiered implementation framework: • LMIC Clinics: Use simple RDW/platelet thresholds. • District Hospitals: Deploy lightweight AI models on affordable hardware (e.g., Raspberry Pi). • Tertiary Centers: Integrate advanced AI with electronic health records. • Caveat: Performance is currently better in high-income settings (AUC 0.90 vs. 0.82 in LMICs), emphasizing the need for local calibration of model thresholds (e.g., RDW 15.2% in Nigeria) and urgent prospective validation in high-burden regions like Sub-Saharan Africa.

Conclusion

This systematic review and meta-analysis demonstrates that AI models utilizing routine CBC parameters can achieve moderate-to-high discriminative performance in identifying patients at relatively higher risk of adverse outcomes in sickle cell disease. However, interpretation of these findings is constrained by substantial heterogeneity in outcome definitions, the absence of a universally accepted severity gold standard, and variability across healthcare settings.

The identified hematological predictors, including red cell distribution width and platelet count, should be considered as components of exploratory risk signals rather than clinically actionable thresholds. Importantly, no included model has demonstrated clinical utility in prospective interventional settings.

Future research should prioritize standardized severity definitions, external validation in diverse populations, and rigorous evaluation of real-world clinical impact before considering integration into decision support systems.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

LMIC

Low/Middle-Income Country

HER

Electronic Health Record

CNN

Convolutional Neural Network

RDW

Red Cell Distribution Width

SHAP

SHapley Additive exPlanations

GPU

Graphics Processing Unit

FHIR

Fast Healthcare Interoperability Resources

Authors’ contributions

N.T.A.: Conceptualization, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing, Visualization, Supervision, Project Administration, Funding Acquisition (institutional support). Corresponding Author: Responsible for all communication, research integrity, and final manuscript approval. M.A.H.M.: Conceptualization, Methodology, Software, Formal Analysis, Validation, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing, Visualization. R.S.A.: Investigation, Resources, Data Curation, Writing – Review & Editing, Supervision (local). G.S.A.: Investigation, Resources, Data Curation, Validation, Writing – Review & Editing. H.M.A.: Investigation, Resources, Data Curation, Validation, Writing – Review & Editing. A.N.M.G.: Methodology, Software, Validation, Formal Analysis, Data Curation, Writing – Review & Editing. N.M.A.: Conceptualization, Methodology, Validation, Resources, Writing – Review & Editing, Supervision. All authors have read, critically reviewed, and approved the final version of the manuscript. All authors agree to be accountable for all aspects of the work.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Infrastructure and administrative support were provided by University of Science and Technology, Aden, Yemen. The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

All data generated or analyzed during this systematic review are included in this published article and its supporting information files. The complete dataset, extraction sheets, and all analysis code (R and Python scripts) are permanently archived and freely available in the Open Science Framework (OSF) repository: https://osf.io/vg3rc/. A static copy is also provided as Files S1–S10.

Declarations

Ethics approval and consent to participate

Not applicable. This study is a systematic review and meta-analysis of previously published, anonymized aggregate data. No original data involving human participants were collected by the authors; thus, ethical approval was not required.

Consent for publication

Not applicable. This manuscript contains no individual person’s data in any form (including individual details, images, or videos). All data presented are aggregated or de-identified.

Supporting information

A complete list of all supporting files, datasets, code, and documentation—organized by category with detailed descriptions—is provided in the supporting information Index File (available at [https://doi.org/10.17605/OSF.IO/VG3RC] and included as a separate document).

Competing interests

The authors declare no competing interests.

Reporting Guidelines

This review adheres to the following reporting guidelines, with completed checklists provided in File S1:

- PRISMA-DTA 2020 for systematic reviews of diagnostic test accuracy.

- CLAIM (Checklist for AI in Medical research).

- PROBAST-AI for risk of bias assessment in prediction model studies.

- TRIPOD+AI for transparent reporting of a multivariable prediction model for individual prognosis or diagnosis using AI.

- GRADE-AI for assessing the certainty of evidence.

- STARD-AI for diagnostic accuracy studies using AI.