The diagnostic accuracy of artificial intelligence in detecting foot and ankle fractures: a systematic review and meta-analysis of diagnostic test accuracy

Abstract

Introduction

Foot and ankle fractures, including radiographically subtle or occult injuries, present a diagnostic challenge in emergency settings, with missed diagnoses causing severe complications. Artificial intelligence (AI), specifically deep learning, offers a promising adjunct for radiographic interpretation. This systematic review and meta-analysis evaluates the diagnostic test accuracy of AI in detecting foot and ankle fractures and appraises the available evidence for occult fracture detection where reported.

Method

Adhering to PRISMA-DTA guidelines, a systematic search was conducted across PubMed, Embase, and Scopus from inception to 10 May 2026. Studies evaluating AI algorithms for identifying foot and ankle fractures on radiographs or CT were included, with occult-fracture data extracted separately when available. Extracted data populated 2 × 2 contingency tables. We utilized a bivariate random-effects model to calculate pooled sensitivity, specificity, and the summary receiver operating characteristic (SROC) curve. Quality was assessed via QUADAS-2.

Results

Sixteen studies encompassing over 37,000 radiographs or images were included. Pooled sensitivity was 92.4% (95% CrI 82.5–96.1%) and pooled specificity was 95.1% (95% CrI 83.3–98.3%). The diagnostic odds ratio was 236.64 (95% CrI 33.56-1087.11). At a 25% pre-test probability, a positive AI result increased post-test probability to 86%, while a negative result reduced it to 3%. Substantial between-study heterogeneity was observed. Only two studies explicitly reported occult fracture cases; therefore, the pooled estimates primarily reflect AI performance for overall foot and ankle fracture detection rather than occult fractures alone.

Conclusion

AI algorithms demonstrate high diagnostic accuracy for foot and ankle fracture detection and show meaningful clinical utility as adjunctive tools. However, current evidence is insufficient to support robust occult-fracture-specific conclusions, and all included studies were retrospective. Prospective validation, standardized reporting of occult fracture subgroups, and real-world impact studies are needed before widespread clinical implementation.

Introduction

Foot and ankle fractures are among the most common musculoskeletal injuries encountered in emergency departments, accounting for approximately 10–15% of all fractures presenting to acute care settings [1]. Ankle fractures alone occur at an estimated incidence of 100–180 per 100,000 person-years in Western populations, while foot fractures, including metatarsal and calcaneal injuries, contribute an additional substantial burden [2, 3]. Among these, occult fractures represent a particularly challenging diagnostic subset, with reported miss rates on conventional radiography ranging from 3% to as high as 30% depending on the anatomical location and reader experience [4]. Delayed or missed diagnoses of occult foot and ankle fractures are associated with complications including malunion, avascular necrosis, post-traumatic osteoarthritis, chronic pain, and prolonged functional disability, which collectively increase healthcare utilisation costs by an estimated 20–40% compared to timely diagnoses [2–4]. Consequently, there is a pressing clinical need for adjunctive diagnostic tools capable of improving the detection sensitivity for these radiographically subtle injuries [4].

Artificial intelligence (AI), particularly deep learning (DL) and machine learning (ML) algorithms, has emerged as a promising technology for medical image analysis over the past decade. Convolutional neural networks (CNNs) and their variants, including ResNet, Inception, EfficientNet, and ensemble architectures, have demonstrated remarkable performance in automated fracture detection across various anatomical regions [5]. These AI systems can be trained on large datasets of radiographic images to identify patterns indicative of fractures that may escape the human eye, potentially serving as a second reader or triage tool in high-volume clinical settings [5]. Several studies have evaluated the diagnostic accuracy of AI algorithms specifically for foot and ankle fracture detection, with reported sensitivities and specificities varying considerably across different models, datasets, and clinical contexts [6–8]. However, the available literature often evaluates overall fracture detection rather than truly radiographically occult fractures, creating a need to distinguish pooled general fracture-detection performance from occult-fracture-specific evidence.

Systematic reviews with diagnostic test accuracy (DTA) meta-analyses provide the highest level of evidence for evaluating the performance of diagnostic technologies by pooling data across multiple studies. A rigorous synthesis of the available evidence is essential to determine whether AI algorithms have achieved sufficient diagnostic accuracy to be considered for clinical implementation. Furthermore, understanding the sources of heterogeneity across studies, including variations in AI architectures, imaging modalities, reference standards, validation settings, and patient populations, is critical for contextualizing the results and guiding future research. This information is particularly relevant as healthcare systems increasingly explore the integration of AI-assisted radiographic interpretation into routine clinical workflows. Therefore, the objective of this systematic review and diagnostic test accuracy meta-analysis was to systematically evaluate and synthesize the pooled diagnostic accuracy of AI algorithms in detecting foot and ankle fractures on radiographic imaging, while separately describing the limited evidence for occult fracture detection when reported.

Method

Study registration and reporting guidelines

This systematic review and meta-analysis was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy (PRISMA-DTA) 2020 guidelines [9]. No funding was received for this study.

Search strategy

A comprehensive and systematic literature search was conducted across three electronic databases: PubMed, Embase, and Scopus, from inception until 10 May 2026. The search strategy was developed using a combination of Medical Subject Headings (MeSH) terms and free-text keywords related to three core concepts: (1) artificial intelligence, deep learning, or machine learning; (2) fracture detection or diagnosis; and (3) foot and/or ankle. Boolean operators (AND, OR) were used to combine search terms, and no restrictions were applied on publication date or language. The reference lists of all included studies and relevant review articles were manually screened to identify any additional eligible studies not captured by the electronic search.

Eligibility criteria

Studies were eligible for inclusion if they met the following criteria: (1) human studies with a case-control, cohort, cross-sectional, or randomized controlled trial design; (2) application of an AI, DL, or ML algorithm for the purpose of fracture detection; (3) evaluation of foot and/or ankle fractures, with occult-fracture data extracted as a distinct subgroup when explicitly reported; (4) reporting of diagnostic accuracy metrics, including true positives (TP), false positives (FP), true negatives (TN), false negatives (FN), sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), or data sufficient to reconstruct the 2 × 2 contingency table. Studies were excluded if they were: (1) case reports or case series; (2) study protocols without results; (3) conference or meeting abstracts without a corresponding full-text publication; (4) review articles of any type (systematic, scoping, narrative, or literature reviews); (5) non-human, phantom, or cadaveric studies; (6) not focused on foot and/or ankle fractures; or (7) lacking an AI/DL/ML component.

Study selection and data extraction

Two independent reviewers performed the study selection process in a stepwise manner. After removal of duplicates, titles and abstracts were screened for relevance against the predefined eligibility criteria. Potentially eligible studies were then retrieved in full text and assessed for final inclusion. Any discrepancies between the two reviewers at any stage of the screening or data extraction process were resolved through discussion and consensus. Data extraction was performed independently by both reviewers using a standardized data extraction form that captured the following variables: first author, year of publication, country, study design, sample size, number of fracture and control cases, number of occult fractures, anatomical site, imaging modality, AI/DL model architecture(s), commercial or research algorithm status, internal or external validation status, reference standard (ground truth), training/validation/test set sizes, radiographic views used, and diagnostic accuracy metrics (TP, FP, TN, FN, sensitivity, specificity, AUC-ROC, accuracy, F1-score, positive predictive value, and negative predictive value). Where diagnostic accuracy metrics were not directly reported, the 2 × 2 contingency table (TP, FP, FN, TN) was back-calculated from available data including sensitivity, specificity, and the total number of positive and negative cases.

Risk of bias assessment

The methodological quality and risk of bias of each included study were assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool, which evaluates four key domains: patient selection, index test, reference standard, and flow and timing. Each domain was rated as having low risk of bias or some concerns (flagged with concerns). The applicability of each study to the review question was also assessed across the first three domains. Two reviewers independently completed the QUADAS-2 assessment, and any disagreements were resolved through discussion.

Statistical analysis

All statistical analyses were performed using RStudio (version 2024.12.1 + 563, R version 4.4.2; R Foundation for Statistical Computing, Vienna, Austria). A bivariate random-effects meta-analysis model was used to jointly estimate pooled sensitivity and specificity, as recommended for diagnostic test accuracy meta-analyses. This approach was selected because DTA studies generate paired sensitivity and specificity estimates that may be correlated through diagnostic threshold effects; therefore, separate univariate pooling of sensitivity and specificity would ignore this correlation. The analysis was implemented using the “mada” and “meta” packages in R, employing a Bayesian bivariate normal model on the logit scale with posterior estimates obtained via Markov chain Monte Carlo (MCMC) methods. Pooled estimates were reported with 95% credible intervals (CrI).

The following summary measures were calculated: pooled sensitivity, pooled specificity, positive likelihood ratio (LR+), negative likelihood ratio (LR-), and diagnostic odds ratio (DOR). A summary receiver operating characteristic (SROC) curve was constructed to illustrate the overall diagnostic performance across studies. Between-study heterogeneity was assessed using the between-study standard deviations for logit-transformed sensitivity and specificity, as well as the between-study correlation parameter (rho), rather than relying on a single univariate I2 statistic. This modelling strategy was considered appropriate because recent meta-epidemiological evidence shows that heterogeneity estimators and the use or non-use of Hartung-Knapp adjustment can materially influence pooled results and statistical significance in conventional random-effects meta-analyses [10]. In the present Bayesian bivariate DTA framework, uncertainty in between-study heterogeneity was incorporated through the posterior distribution of the random-effects parameters. A Fagan nomogram was generated to illustrate the clinical utility of the AI algorithms by depicting the relationship between pre-test probability, likelihood ratios, and post-test probability. Sensitivity analyses were performed to evaluate the robustness of the pooled estimates. A p-value of < 0.05 was considered statistically significant for all analyses.

Exploratory subgroup analyses were planned for anatomical site (ankle versus foot/calcaneus versus mixed foot-and-ankle studies), algorithm type (commercial versus research algorithms), validation setting (internal versus external validation), imaging modality (plain radiographs versus CT), and occult-fracture reporting status. Quantitative subgroup pooling was performed only when at least three studies within a subgroup reported extractable 2 × 2 diagnostic tables. When this threshold was not met, subgroup domains were summarized descriptively to avoid overinterpreting sparse or non-comparable data.

Results

Study selection

Database searching identified a total of 652 records across PubMed (n = 266), EMBASE (n = 204), and Scopus (n = 182). Following removal of 34 duplicate records, 618 records proceeded to title and abstract screening, of which 571 were excluded. The remaining 47 reports were sought for full-text retrieval, all of which were successfully retrieved (reports not retrieved = 0). Full-text assessment of all 47 reports led to exclusion of 31 studies due to reasons including conference abstracts without full-text publication (n = 3), mixed anatomical regions with no extractable foot/ankle-specific data (n = 11), AI used for segmentation only (n = 2), AI used for treatment planning (n = 5), AI used for prognosis (n = 3), and non-derivable diagnostic accuracy metrics (n = 7). Ultimately, 16 studies met all eligibility criteria and were included in the final systematic review (Fig. 1) [11–26].

Fig. 1.

PRISMA flow diagram showing study selection process

Risk of bias analysis

Risk of bias was assessed across 16 included studies using the QUADAS-2 tool, evaluating four domains: patient selection (D1), index test (D2), reference standard (D3), and flow and timing (D4) (Fig. 2). The majority of studies demonstrated low risk of bias across all four domains, including Ashkani-Esfahani et al. [19], Wang et al. [22], Kim et al. [25], Cheng et al. [20], Guo et al. [16], Guo et al. [21], Hendrickx et al. [11], Bousson et al. [15], Ruitenbeek et al. [17], and Pranata et al. [26]. Overall, the preponderance of low-risk judgments across studies suggests an acceptable quality of evidence base, with only a minority of studies flagged for methodological concerns in specific domains.

Fig. 2.

Risk of bias using QUADAS-2

Demographic characteristics

A total of 16 studies published between 2019 and 2025 met the eligibility criteria and were included in this systematic review (Table 1). All included studies employed a retrospective design, with the majority being retrospective cohort studies and one study utilizing a retrospective case-control design. The studies originated from diverse geographical regions, including the United States, South Korea, China, Taiwan, France, the Netherlands, Sweden, and Australia, with two studies being multicenter in design. Sample sizes ranged from 171 to 14,311 subjects, reflecting considerable variability in the scale of the datasets used for model development and validation. The majority of studies focused on ankle fractures (n = 9), followed by foot fractures including calcaneal fractures (n = 4), and studies evaluating multiple anatomical regions inclusive of the foot and ankle (n = 3).

Table 1.

Demographic and methodological characteristics of the 16 included studies

Author (Reference)	Country	Sample size	Fracture (n)	Control (n)	Occult (n)	Site	Imaging	AI model(s)	Reference standard
Ashkani-Esfahani et al. [19]	USA/Netherlands	2100	1050	1050	72	Ankle (malleolar)	Plain radiographs	Inception V3; ResNet-50	CT + expert consensus
Kim et al. [18]	South Korea	1226	1040	186	NR	Ankle	Plain radiographs	Inception V3	CT-confirmed labeling
Kitamura et al. [14]	USA	596	298	298	NR	Ankle	Plain radiographs	Inception V3; ResNet; Xception (ensemble)	Radiology reports + expert consensus
Wang et al. [22]	China	1891	NR	NR	NR	Ankle	Plain radiographs	ResNet50; SENet; ResNet50-SE	Expert radiologist consensus
Kim et al. [25]	South Korea	4861	3274	1587	NR	Foot	Plain radiographs	VGG16; ResNet50; DenseNet121; EfficientNetB0	Expert orthopedic surgeon
Cheng et al. [20]	Taiwan	3493	2295	1198	NR	Ankle	Plain radiographs	EfficientNet-based ensemble	CT/expert consensus
Guo et al. [16]	China	961	524	437	NR	Calcaneus	Plain radiographs	Custom CNN	Expert diagnosis
Guo et al. [21]	China	2211	957	1254	NR	Foot	Plain radiographs	U-Net-CH (EfficientNet-b3)	Expert orthopedist annotation
Wu et al. [12]	China	NR	NR	NR	NR	Ankle	Plain radiographs	ML-based classification	Expert consensus
Yu et al. [24]	China	171	Variable	Variable	NR	Calcaneus	Plain radiographs	SURF-BoVW + XGBoost; SVM	Expert orthopedic surgeon
Hendrickx et al. [11]	Multicenter	446	253	193	74	Posterior malleolus	Radiographs + clinical data	ML (clinical + imaging)	Expert orthopedic diagnosis
Bousson et al. [15]	France	480	Variable	Variable	NR	Ankle and foot	Plain radiographs	BoneView; SmartUrgence; Rayvolve	Expert radiologist consensus
Guermazi et al. [23]	France	1500	356	1144	NR	Ankle and foot	Plain radiographs	SmartUrgence; BoneView; Rayvolve	MSK radiologist consensus
Ruitenbeek et al. [17]	Netherlands	14,311	4156	10,155	NR	Ankle and foot	Plain radiographs	qXR Fracture (Qure.ai)	Radiology reports + expert
Pranata et al. [26]	Taiwan	14,311	683	4156	10,155	15 body parts incl. ankle, foot and toe	CT images	Multitask deep neural network	Physician-labeled DICOM
Olczak et al. [13]	Sweden	1931	683	683	1248	Ankle	Plain radiographs	VGG-16; ResNet-50 (classification); SURF + Canny edge detection + contour tracing (fracture localization)	Expert consensus + registry

NR, not separately reported; NS in the previous table was standardized to NR. Highlighted cells indicate corrected or clarified entries in the revised table.

Plain radiographs were the predominant imaging modality evaluated across the included studies, with only one study [26] utilizing CT images. A wide variety of AI architectures were employed, including Inception V3, ResNet-50, VGG-16, DenseNet-121, EfficientNet, ensemble models, and traditional machine learning classifiers such as XGBoost and support vector machines (SVM). Transfer learning from pre-trained models was the most common training strategy, although some studies employed de novo training or combined approaches. Reference standards varied across studies and included expert radiologist or orthopedic surgeon consensus, CT-confirmed labeling, radiology reports verified by senior specialists, and clinical or surgical follow-up. Only two studies [11, 19] explicitly reported the number of occult fractures within their datasets, with 72 and 74 occult fractures identified, respectively; occult-fracture-specific 2 × 2 diagnostic tables were not consistently available across studies.

The training, validation, and test set partitioning strategies varied across studies. The most common data split was an 80/10/10 ratio for training, validation, and testing, although several studies employed k-fold cross-validation rather than a fixed holdout test set. External validation was performed in four studies that evaluated commercial AI algorithms (BoneView, SmartUrgence, Rayvolve, and qXR Fracture) on independent datasets not used during model training. Radiographic views assessed included anteroposterior (AP), lateral, and mortise views, with several studies specifically investigating the incremental value of multi-view inputs over single-view analysis. The main heterogeneity domains considered for subgroup exploration were anatomical site, commercial versus research algorithms, internal versus external validation, and imaging modality.

Meta-analysis of DTA

The bivariate random-effects meta-analysis (Figs. 3 and 4) yielded a pooled sensitivity of 92.4% (95% CrI 82.5–96.1%) and a pooled specificity of 95.1% (95% CrI 83.3–98.3%). The pooled diagnostic odds ratio (DOR) was 236.64 (95% CrI 33.56–1087.11), indicating a high overall discriminative capacity of AI algorithms for detecting foot and ankle fractures. The positive likelihood ratio (LR+) was 18.79 (95% CrI 5.28–55.48), suggesting that a positive AI result substantially increases the post-test probability of fracture. The negative likelihood ratio (LR-) was 0.080 (95% CrI 0.040–0.196), indicating that a negative AI result meaningfully decreases the post-test probability. The Fagan nomogram (Fig. 5) demonstrated that at a pre-test probability of 25%, a positive AI result increased the post-test probability to approximately 86%, while a negative result reduced it to approximately 3%, underscoring the potential clinical utility of these algorithms as adjunctive diagnostic tools.

Fig. 3.

HSROC for AI in diagnosing foot and ankle fractures

Fig. 4.

Forest plot for sensitivity and specificity analysis

Fig. 5.

Fagan nomogram analysis

Substantial between-study heterogeneity was observed, with between-study standard deviations for logit-transformed sensitivity and specificity of 0.715 (95% CrI 0.325–1.683) and 1.144 (95% CrI 0.654–2.119), respectively. The between-study correlation parameter (rho) was 0.579 (95% CrI − 0.206 to 0.946), suggesting a moderate positive correlation between sensitivity and specificity across studies, which may reflect the influence of threshold effects and differences in AI model calibration. The accuracy parameter (Lambda) was 5.652 (95% CrI 3.615–7.256), and the shape parameter (beta) was 0.471 (95% CrI − 0.493 to 1.308), indicating slight asymmetry in the SROC curve.

Exploratory subgroup assessment was therefore reported descriptively. Ankle-focused studies, foot/calcaneus-focused studies, and mixed foot-and-ankle studies were all represented, and four studies evaluated commercial AI tools using external validation datasets. However, subgroup-specific 2 × 2 diagnostic tables were not consistently extractable across these categories, and only two studies reported occult fracture counts. As a result, quantitative subgroup meta-analysis for occult fractures, anatomical site, commercial versus research algorithms, and internal versus external validation was not considered statistically robust.

Discussion

This systematic review and diagnostic test accuracy meta-analysis provides a comprehensive evaluation of AI for foot and ankle fracture detection, analyzing 16 studies with more than 37,000 radiographs or images. The pooled sensitivity (92.4%) and specificity (95.1%) indicate that modern AI models show high diagnostic performance in the available retrospective evidence base. Importantly, these pooled estimates should be interpreted as reflecting overall foot and ankle fracture detection rather than occult-fracture-specific diagnostic accuracy, because only two included studies explicitly reported occult fracture cases. AI models using ensemble and multi-view architectures frequently achieved high performance in individual studies, supporting the hypothesis that multi-view and ensemble approaches may improve fracture detection accuracy. The Fagan nomogram further suggested potential clinical utility, with a positive AI result increasing post-test fracture probability from 25 to 86%, while a negative result reduced it to 3%. However, substantial heterogeneity across studies and the absence of prospective impact studies require cautious interpretation.

The foot alone comprises 26 bones, 33 joints, and over 100 ligamentous structures, creating a radiographic environment where overlapping osseous densities, accessory ossicles, and normal anatomical variants frequently mimic or obscure fracture lines [27]. Occult fractures in this region carry substantial clinical consequences when missed, including malunion rates of 10–15%, avascular necrosis risk exceeding 20% for certain talar injuries, and progression to post-traumatic osteoarthritis in up to 40% of inadequately treated ankle fractures [28]. Although AI algorithms may be capable of detecting subtle cortical disruptions, trabecular pattern irregularities, and soft tissue swelling patterns that serve as indirect fracture signs, the present review cannot confirm this performance specifically for occult fractures because occult cases were inconsistently reported and not accompanied by extractable subgroup 2 × 2 data in most studies [29, 30]. Therefore, the occult-fracture implication of our findings should be considered hypothesis-generating rather than definitive.

From a clinical applicability standpoint, these findings may be relevant for emergency department and musculoskeletal radiology workflows, but implementation should remain cautious. Missed foot and ankle fractures are an important clinical problem, and AI could potentially function as a concurrent reader, triage aid, or safety-net tool in settings where radiographs are initially interpreted by non-specialists [31]. Nevertheless, all included studies were retrospective, and no prospective impact studies evaluated whether AI deployment reduces missed diagnoses, shortens time to treatment, improves patient outcomes, or changes resource use in real-world practice. The performance gap observed between commercial AI tools evaluated in emergency department settings and research-grade models on curated datasets also warrants caution, because radiographic quality, patient positioning, local prevalence, and scanner or acquisition protocols may influence deployment-phase performance [32–37]. Healthcare systems considering implementation should therefore prioritize prospective local validation and workflow monitoring before relying on AI outputs in clinical decision-making [38].

This study has several limitations that must be acknowledged. First, all 16 included studies were retrospective in design, which introduces inherent selection bias and limits the generalizability of findings to prospective clinical workflows where image quality, patient acuity, and prevalence vary more widely. Second, although the initial rationale emphasized occult fractures, only two studies explicitly reported the number of occult fractures within their datasets, making it impossible to perform a robust subgroup analysis specifically isolating AI performance for truly radiographically occult injuries as opposed to overall fracture detection. Accordingly, the title, objective, interpretation, and conclusion of the revised manuscript have been aligned with the broader evidence base of foot and ankle fracture detection. Third, substantial methodological heterogeneity existed across studies in terms of AI architectures, training strategies, dataset sizes ranging from 171 to over 14,000, reference standards, and data partitioning approaches, which complicates direct comparison and contributes to the observed between-study variability. Fourth, exploratory subgroup analyses by ankle versus foot fractures, commercial versus research algorithms, internal versus external validation, and imaging modality were limited by inconsistent reporting of subgroup-specific 2 × 2 data, so these factors were summarized descriptively rather than pooled quantitatively. Fifth, the predominance of plain radiographs as the evaluated imaging modality, with only one study using CT, limits conclusions about AI performance across the full spectrum of imaging used in clinical fracture evaluation. Sixth, geographic and institutional concentration of training data raises concerns about algorithmic bias, as models trained predominantly on East Asian or Western European populations may not generalize to other demographic groups with differing skeletal morphology or fracture patterns. Finally, publication bias cannot be excluded, as studies demonstrating poor AI performance are less likely to be published, potentially inflating the pooled diagnostic accuracy estimates.

Conclusion

This systematic review and diagnostic test accuracy meta-analysis demonstrates that artificial intelligence algorithms achieve high pooled sensitivity and specificity for detecting foot and ankle fractures on radiographic imaging, with a strong overall discriminative performance. The clinical utility analysis suggests that AI may function as an effective adjunctive tool, substantially increasing post-test probability when positive and meaningfully reducing the probability of fracture when negative. However, the evidence base remains exclusively retrospective, heterogeneous in methodology, and insufficiently granular regarding truly occult fractures; therefore, the pooled results should not be interpreted as definitive evidence for occult-fracture-specific detection. Future research should prioritize prospective, multi-center trials evaluating AI as a real-time clinical decision support tool, with pre-specified subgroup analyses for occult fractures, standardized reporting of diagnostic accuracy metrics, and evaluation of patient-centered outcomes including time to diagnosis, treatment appropriateness, and functional recovery. Until such evidence is available, AI algorithms should be regarded as promising adjuncts to, rather than replacements for, expert clinical and radiological judgment in the evaluation of suspected foot and ankle fractures.

Author contributions

GTP and JHW contributed to the conception and design of the study. GTP and JHW performed the literature search, study selection, and data extraction. GTP and JHW performed the statistical analysis and interpreted the results. GTP and JHW drafted the manuscript. Both authors reviewed and approved the final version of the manuscript.

Funding

No funding was received for this study.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable. This study is a systematic review and meta-analysis of previously published studies and did not involve direct human participants or animal subjects. Therefore, ethics approval and informed consent were not required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.