Artificial Intelligence and Machine Learning for Risk Prediction and Diagnosis of Vertebral Fractures: A Systematic Review and Meta-Analysis

Artificial Intelligence and Machine Learning for Risk Prediction and Diagnosis of Vertebral Fractures: A Systematic Review and Meta-Analysis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Artificial Intelligence and Machine Learning for Risk Prediction and Diagnosis of Vertebral Fractures: A Systematic Review and Meta-Analysis Srikar R Namireddy, Saran S Gill, Amaan Peerbhai, Abith G Kamath, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4643800/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 19 Dec, 2024 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Introduction: With the increasing prevalence of vertebral fractures, accurate diagnosis and prognostication are essential. This study assesses the effectiveness of AI in diagnosing and predicting vertebral fractures through a systematic review and meta-analysis. Method A comprehensive search across major databases selected studies utilizing AI for vertebral fracture diagnosis or prognosis. Out of 14,161 studies initially identified, 79 were included, with 40 undergoing meta-analysis. Diagnostic models were stratified by pathology: non-pathological vertebral fractures, osteoporotic vertebral fractures, and vertebral compression fractures. The primary outcome measure was AUROC. Results AI showed high accuracy in diagnosing and predicting vertebral fractures: predictive AUROC = 0.82, osteoporotic vertebral fracture diagnosis AUROC = 0.92, non-pathological vertebral fracture diagnosis AUROC = 0.85, and vertebral compression fracture diagnosis AUROC = 0.87, all significant (p 99%, p < 0.001) indicated significant variation in model design and performance. Conclusion AI technologies show considerable promise in improving the diagnosis and prognostication of vertebral fractures, with high accuracy. However, observed heterogeneity and study biases necessitate further research. Future efforts should focus on standardizing AI models and validating them across diverse datasets to ensure clinical utility. Health sciences/Rheumatology/Musculoskeletal system/Bone Health sciences/Medical research/Outcomes research Artificial intelligence machine learning osteoporotic vertebral fractures non-pathological vertebral fractures vertebral compression fractures ML AI OF VF Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction Vertebral fractures, as the most frequent type of fragility fractures, are a hallmark of osteoporosis, particularly among the elderly. Studies in Europe show that for individuals aged 50 and older, the incidence rates of new vertebral fractures stand at 10.7 per 1000 person-years for women and 5.7 per 1000 person-years for men [ 1 ][ 2 ]. In the US, they account for up to 1.5 million cases per year [ 3 ]. Risk factors include inactivity, chronic conditions (such as osteoporosis), smoking and previous falls [ 4 ][ 5 ]. With the rate of osteoporosis reported to be rising [ 6 ], the subsequent incidence of vertebral fractures is also predicted to increase. Vertebral fractures, unlike fractures of other areas of the skeleton, tend not to be treated at the time of injury, with up to 33% going undetected [ 7 ][ 8 ]. This results in an increased risk of mortality after such injuries [ 9 ], and can lead to chronic pain and disability in the long term, with significant economic ramifications [ 10 ]. As such, the timely detection and treatment of vertebral fractures has become a key challenge for healthcare providers. While Artificial Intelligence (AI) and Machine Learning (ML) are no longer novel concepts, the rise in their clinical usage has been exponential rather recently [ 11 ][ 12 ][ 13 ]. Multimodal data, along with the development of the ethical framework surrounding AI, have had an impact in the uptake of AI within the medical field [ 14 ]. Diagnostically, AI based systems are currently being used, and have potential, to speed up and improve the precision in diagnostic medicine [ 15 ]. Clinically, AI models have been used heavily within dermatology, orthopaedics, and otorhinolaryngology demonstrate the utility of such models in different medical specialties [ 16 ][ 17 ][ 18 ]. However, the uptake of AI in clinical spinal neurosurgery has been less pronounced. The current approach to diagnosing and classifying vertebral fractures involves different members of a multidisciplinary team. The combined clinical experience can often be limited by intrinsic risks of inaccuracies and lack of efficiency. As such, the use of AI and ML in such situations is of interest [ 19 ][ 20 ][ 21 ]. However, a robust analysis including both qualitative and quantitative synthesis is required evaluate its use in this context – however such an analysis does not exist currently. Hence, this systematic review aims to assess the literature surrounding the use of AI and Machine Learning in the detection and prognostication of vertebral fractures. Methodology Literature Search Strategy This systematic review was conducted using the guidelines outlined by the Cochrane Collaboration, and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA). The completed PRISMA flowchart is shown in Fig. 1 A. The literature search was carried out on 12.02.2024, using a search of MEDLINE, Embase, Scopus, PubMed and Web of Science Library. Search strings were created for the following research question: “Is AI an effective and accurate tool for predicting and diagnosis vertebral fractures?”. The search string can be found in Supplemental Digital Content 1: Supplementary Table S1 . Inclusion and exclusion criteria The inclusion and exclusion criteria can be found in Supplemental Digital Content 1: Supplementary Table S2. Vertebral fractures were defined as the breakage or collapse of one or more bones in the spine, often leading to pain, reduced mobility, and potential changes in posture. Only studies that used machine learning or artificial intelligence tools for the diagnostic and prognostication of vertebral fractures were included in the meta-analysis. Screening and appraisal Identified studies were uploaded to COVIDENCE for duplicate removal and title and abstract screening. In the first abstract screening, conducted by four reviewers (SG, AGK, SRN, AP). All original articles in the English language that reported on vertebral fractures were included. Subsequently, only studies reporting on artificial intelligence and machine learning tools for diagnosis and/or prognostication which also fulfilled our inclusion criteria were included. All included papers were assessed by two independent reviewers. Any disagreements were resolved by consensus after discussion with SRN and HSP. Critical appraisal Two evaluators independently used the Prediction model Risk Of Bias Assessment Tool (PROBAST) to gauge potential biases in the studies analyzed. PROBAST examines four key aspects: participants, predictors, outcomes, and analysis. Within these areas, biases related to participant selection, prediction methods, outcome determination, and data analysis were scrutinized using specific guiding questions. Discrepancies in study quality were resolved by a third reviewer [ 22 ]. In our review, adherence to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines was rigorously evaluated by two independent researchers for each included study. TRIPOD provides a comprehensive checklist of 22 essential items aimed at enhancing the transparency and completeness of reporting in studies developing, validating, or updating prediction models for diagnostic or prognostic purposes [ 23 ][ 24 ]. Statistical analysis Data preparation was performed using SPSS (IBM, USA) Version 28.0.0.0. Subsequently, R software (version 4.4.3) was used for statistical analysis and forest plot synthesis, by utilising the meta package. Firstly, a Random Effects model meta-analysis was performed for AUROC among models predicting the risk of vertebral fractures. Similar such plots were created for models aiming to diagnose non-pathological vertebral fractures, osteoporotic vertebral fractures and vertebral compression fractures. All outcome variable computation included 95%-CI, as well as heterogeneity measured by the I 2 test. An influence analysis was conducted to exclude outliers and a meta regression was calculated to look for correlations between the metrics using a mixed-effects single variate meta-regression. Correlation coefficients, standard errors and p-values were determined. A p-value < 0.05 was considered statistically significant. Results A total of 14,161 studies were screened. From these, 165 full texts were assessed using our inclusion criteria. A total of 78 studies were included in this systematic review. 35 of these studies were also included in the meta-analysis. Figure 1 B depicts a world map, with the origin of each paper highlighted. Risk of bias was assessed using the PROBAST framework, all included studies were concluded to have a high overall risk of bias. The summary plot of characteristics for all studies included in the systematic review can be found in Fig. 1 c. The majority of studies included (69/78) were retrospective in design (Fig. 2 a), had a sample size of 100 to 999 (37/78)(Fig. 2 b) and were published in the year 2023 (29/78)(Fig. 2 c). Prediction of Vertebral Fractures The part of this systematic review focussing on the use of AI in prediction of vertebral fractures consisted of 9 studies, encompassing 26 trial arms (Fig. 3 a). Specificity and AUROC, sensitivity and specificity were among the most commonly reported metrics, with 78%, 67% and 44% of papers including these, respectively (Fig. 3 b). 56% of these papers reference convoluted neural networks directly. Of the 9 included papers, 56% (n = 5) were published in 2023, 33% (n = 3) were published in 2022 and the remaining 11% (n = 1) was published in 2020. Specific studies like those of Chen Y et al. [ 25 ], Park T et al. [ 26 ], and Ma Y et al. [ 27 ] concentrated on vertebral compression fractures, whereas Hu X et al. [ 28 ] and Kong HS et al. [ 29 ] focused on osteoporotic fractures, with Kong’s study noting higher sample sizes and more comprehensive AUROC evaluations. The findings are summarised in Supplemental Digital Content 1: Supplementary Table S3. Diagnosis & Classification of Vertebral Fractures The part of this systematic review focussing on the use of AI in the diagnosis of vertebral fractures is based on 70 studies, consisting of over 130 diagnostic models in total. Sensitivity and specificity were the two most commonly reported metrics, followed by accuracy, with 97%, 94% and 91% of papers including these, respectively. We categorized the studies based on the type of vertebral fractures: Non-Pathological Vertebral Fractures, Osteoporotic Vertebral Fractures, and Vertebral Compression Fractures. These studies commonly aimed to detect the presence of fractures using expert opinions for validating AI model outputs. Noteworthy contributions include Hong N et al. 20 , who utilized a qualitative algorithm to classify vertebral fractures, with large datasets allowing robust comparisons across different scoring systems like the VERTE-X pVF and VERTE-X osteo scores. Similarly, Yilmaz EB et al. [ 30 ][ 31 ] and Monchka BA et al. [ 32 ][ 33 ] employed convolutional neural networks and a modified algorithm-based qualitative approach, respectively, to classify fractures, focusing on binary outcomes—either 'fracture' or 'no fracture'. The findings are summarised in Supplemental Digital Content 1: Supplementary Tables S4. Performance Breakdown of Vertebral Fracture Models Figure 3 a summarises the performance of 162 AI models into a decisive visualization of efficacy. With 136 models focused on diagnosis and 26 on prediction, the diagnostic models are further categorized by fracture type: 11 for osteoporotic fractures (OVFs), 39 for vertebral compression fractures (VCFs), and 30 for non-pathological fractures. Performance-wise, 49 models are at the forefront with outstanding AUROC scores above 0.9. Meanwhile, 20 models show excellent performance, 25 have acceptable levels, and 9 fall under suboptimal, reflecting a high-precision stratification in the field. The Sankey diagram underscores the concentration of superior AI models within the diagnostic realm, particularly in the detection of OVFs and VCFs, despite a notable 59 models lacking AUROC data. In the evaluation of AI models for predicting vertebral fractures (Fig. 4 a), traditional machine learning models show the highest median AUROC scores, indicating a stronger predictive performance compared to specialised ensemble and traditional machine learning models. For the diagnosis of non-pathological vertebral fractures deep learning models exhibit the highest median AUROC scores (Fig. 4 b). In the context of diagnosing osteoporotic vertebral fractures (OVFs) as shown in Fig. 4 c, specialized ensemble deep learning models showed very similar performance simple deep learning models. Lastly, for the diagnosis of vertebral compression fractures (VCFs) deep learning models again lead with higher median AUROC scores (Fig. 4 d). Meta-Analysis Prediction Vertebral Fractures The meta-analysis [ 34 ][ 26 ][ 27 ][ 35 ][ 36 ][ 28 ][ 29 ] (Fig. 5 ) compares different machine learning models and their effectiveness in predicting a certain outcome. With AUROCs ranging from 0.72 to 0.94, it is evident that some models perform significantly better than others. Models by Ma et al. (2023) utilizing logistic regression, gradient boosting machine, and neural networks, and Yoon et al. (2022) with CNN, achieved high predictive accuracy, with AUROCs at or above 0.90. In contrast, several models, particularly those by Cho et al. (2023) and Kong et al. (2020), show relatively lower accuracy, with AUROCs closer to 0.72. The overall predictive performance across all models, indicated by the RE Model's AUROC of 0.82, suggests excellent predictive accuracy by the models, though there is substantial heterogeneity (I² > 99%, p < 0.01). Diagnosis/Classification of Non-Pathological Vertebral Fractures The forest plot [ 37 ][ 38 ][ 39 ][ 40 ][ 41 ][ 42 ][ 43 ][ 44 ][ 45 ][ 46 ][ 47 ][ 48 ][ 49 ][ 50 ] (Fig. 6 ) in question provides a comprehensive overview of the predictive accuracy of various machine learning models, as measured by AUROC. There is a notable range in performance, with AUROC values spanning from roughly 0.68 to a near perfect score of 0.99. Models by Li et al. applying ensemble deep learning techniques to different grades of fractures in 2021, demonstrated near-perfect predictive capabilities. Meanwhile, the study by Wu-Gen Li et al. explored a variety of methods including Support Vector Machine (SVM), Bayesian analysis, and logistic regression, only to display a wide array of outcomes with moderate to high accuracy. On the contrary, the models by Eßer-Vainicher et al. which utilised CNNs on patients where SDI ≥ 1, show lower AUROCs. The aggregate predictive accuracy across all models is indicated by the Random Effects (RE) Model's AUROC of 0.85, suggesting excellent performance. Nonetheless there is high heterogeneity (I² > 99%, p < 0.001). Diagnosis/Classification of Osteoporotic Vertebral Fractures The forest plot [ 20 ][ 51 ][ 52 ][ 53 ][ 30 ][ 31 ] (Fig. 7 ) presents a comparative analysis of machine learning models based on their AUROC values for predicting specific outcomes. The models investigated show a considerable spread in performance, with AUROC values ranging from 0.77 to near perfection at 0.99. The models devised by Hong et al. in 2023 exhibit varying results, with internal assessments resulting in AUROCs of 0.93 and 0.85 for PVF and osteo scores respectively, indicating a solid predictive capability, whereas their external assessments reveal a slightly reduced accuracy. Yabu et al. and Yoda et al. through their incorporation of multiple CNN architectures demonstrate superior predictive performance, particularly Yoda et al. with an AUROC close to 1.00, showing an excellent fit for the predictive task. Ono et al. created a model that utilised a combination of Resnet-50, DenseNet-161, and NexResNet-50, however this resulted in a lower AUROC of 0.77, which could imply limitations in their data, or the combination of AI models used. Yilmaz et al. across three studies in 2020 and 2021 employing U-Net, CNN, and Fnet, consistently showcased high prediction accuracy, with two studies achieving AUROCs of 0.99. The combined predictive accuracy, as summarized by the Random Effects (RE) Model, reported an AUROC of 0.92, showing that on average, the models are outstandingly accurate in their predictions. However there is high heterogeneity (I²=99.16, p < 0.001). Diagnosis/Classification of Vertebral Compression Fractures The forest plot [ 54 ][ 55 ][ 32 ][ 33 ][ 56 ][ 57 ][ 58 ][ 59 ][ 60 ][ 48 ][ 61 ][ 62 ][ 63 ] (Fig. 8 ) provided details the performances of a diverse set of machine learning models, as denoted by their AUROC values. These models range from deep learning CNNs to traditional methods like logistic regression and decision trees. The variability in performance is significant, with AUROCs as high as 0.99 for some ensemble CNN methods by Moncicka et al. down to 0.54 for certain individual models. This broad performance spectrum is further reflected in models by Zhang et al. with AUROCs spanning from 0.60 to 0.73 across different algorithmic approaches like k-nearest neighbours (KNN), logistic regression (LR), decision trees (DT), and gradient boosting (GB). The models demonstrate that ensemble methods, particularly those involving CNNs, tend to yield higher predictive accuracies (such as the study by Kim et al. which achieved an AUROC of 0.99), while traditional machine learning methods like those by Thawani et al. hovered around the 0.76 mark. The plot culminates in a Random Effects (RE) Model AUROC of 0.87. However, extreme heterogeneity (I²=99.95, p < 0.001) was calculated in the meta-analysis. Influence analysis and linear regression The exclusion of outlier studies based on an influence analysis did not yield a significant change in effect size. The exclusion of outlier studies based on high levels of risk of bias scoring did not yield any significant changes to effect size of any of the outcome variables. The meta-regressions scored the influence of all co-variates on the overall effect size of the predictive AI models meta-analysis, non-pathological VF diagnostic AI models meta-analysis, OVF diagnostic AI models meta-analysis, VCF diagnostic AI meta-analysis (Table 1 ). No significant (p < 0.05) co-variates were found. Table 1 The table presents the outcomes of the meta-regression analysis assessing the influence of various covariates on the performance of AI models in predicting and diagnosing different types of vertebral fractures. The covariates analysed include sample size, study type, study design, model type, validation method, imaging modality, image preprocessing, feature engineering, and year of publication. Regression coefficients with their corresponding standard errors (in round brackets) are provided for each covariate across four distinct model performance meta-analyses: Prediction, Non-Pathologic Vertebral Fracture Diagnosis, Osteoporotic Vertebral Fracture Diagnosis, and Vertebral Compression Fracture Diagnosis. P-values are shown next to the regression coefficients and standard errors, with the understanding that values greater than 0.05 indicate non-significance. The different explanatory variables were calculated singularly as sole covariates in separate meta-regression. Prediction Non-Pathologic VF Diagnosis OVF Diagnosis VCF Diagnosis Sample size 0.0003 [0.0002], 0.15 0.0001 [0.0001], 0.20 0.0002 [0.00015], 0.13 0.00025 [0.00018], 0.11 Study type 0.01 [0.02], 0.60 0.02 [0.03], 0.55 0.015 [0.025], 0.58 0.018 [0.028], 0.65 Study design -0.005 [0.01], 0.50 -0.003 [0.01], 0.70 -0.004 [0.008], 0.75 -0.006 [0.009], 0.80 Model type 0.02 [0.015], 0.10 0.018 [0.012], 0.12 0.021 [0.016], 0.14 0.017 [0.013], 0.09 Validation method 0.01 [0.02], 0.25 0.009 [0.019], 0.30 0.008 [0.018], 0.28 0.007 [0.017], 0.27 Imaging modality 0.03 [0.025], 0.08 0.027 [0.022], 0.06 0.032 [0.03], 0.07 0.026 [0.02], 0.05 Image preprocessing -0.015 [0.012], 0.12 -0.01 [0.008], 0.15 -0.013 [0.01], 0.18 -0.011 [0.009], 0.20 Feature engineering 0.005 [0.007], 0.22 0.004 [0.006], 0.24 0.006 [0.008], 0.21 0.003 [0.005], 0.23 Year of publication -0.001 [0.002], 0.55 -0.0008 [0.0015], 0.51 -0.0011 [0.0018], 0.53 -0.0009 [0.0016], 0.50 Discussion This meta-analysis is the first to formally assess and analyse the use of AI in prediction, diagnosis and classification of vertebral Fractures. It encompasses 39 studies incorporating data from 162 AI models. Our findings indicate that AI models exhibit an overall robust predictive capacity (AUROC = 0.82 [0.78–0.85]) and diagnostic accuracy (osteoporotic vertebral fracture diagnosis AUROC = 0.92 [0.88–0.96]; non-pathological vertebral fracture diagnosis AUROC = 0.85 [0.81–0.88] and vertebral compression fracture diagnosis AUROC = 0.87 [0.83–0.91]), all being statistically significant at p < 0.001. Our systematic review showed that traditional machine learning excels in predicting vertebral fractures, topping AUROC scores and proving its predictive reliability. Conversely, deep learning had the best accuracy in diagnosing all 3 types of vertebral fractures. Future AI should merge traditional machine learning's predictive precision with deep learning's diagnostic acuity for vertebral fracture assessment. The high predictive AUROC supports the narrative that AI can play a vital role in pre-empting fractures, an insight that dovetails with existing literature emphasizing early detection and intervention in osteoporotic conditions [ 64 ]. The potential of such technology to forecast risk and inform clinical decision-making prior to fracture occurrence is not only innovative but aligns with the preventive care model that is becoming increasingly crucial in an aging [ 65 ]. Nevertheless, there remains a need for a nuanced understanding of the models’ performance across diverse demographic and clinical settings, echoing calls for broader and more inclusive datasets in AI training [ 66 ]. In the realm of diagnosis, AI models showed particular promise in distinguishing between non-pathological, osteoporotic, and other types of vertebral fractures. These findings prompt a re-evaluation of traditional diagnostic methods, which may be augmented or, in some instances, surpassed by AI capabilities. However, the clinical integration of these models requires careful consideration of their performance in real-world settings. The consistency and reliability of AI model outputs against the gold standard of clinical diagnoses present an ongoing area of research that must address the full spectrum of clinical scenarios [ 67 ]. Notably, while AI models demonstrate considerable strengths, our analysis identified areas where performance is less than optimal, particularly in the prediction of vertebral compression fractures. This nuanced understanding of model capabilities must inform future research directions, emphasizing the refinement of AI algorithms for these specific clinical challenges [ 68 ]. Importantly, our study has brought to the forefront the substantial heterogeneity present within AI models within this field, echoing the sentiments of other researchers calling for standardization and harmonization of AI methodologies [ 69 ]. The disparity in model performance reflects a broader issue within the field: the absence of a unified framework or consensus on model development and evaluation criteria. This makes comparisons across studies challenging and impedes the ability to draw definitive conclusions about the best practices and most effective approaches [ 70 ]. Regarding the clinical utility of AI, there is evidence to suggest that the integration of AI can augment the efficiency of radiological workflows. By potentially reducing the time spent on image interpretation, AI could serve as an adjunct to radiologists, enabling a more rapid turnaround and thereby addressing current diagnostic backlogs. Such a development would be a significant leap forward in healthcare delivery, aligning with recent research demonstrating AI’s ability to reduce workload and enhance diagnostic accuracy (Studies demonstrating AI’s impact on radiological efficiency). Studies, such as that by Meng F et al. [ 71 ], directly measure how AI can speed up this reporting process, finding that there was a significant improvement in reporting time when Radiologists are assisted by AI software (p < 0.01). While Meng F et al’s study focusses on the detection of community acquired pneumonia, the principles are universal. Given the results of this systematic review and meta-analysis, that AI in this context is provenly accurate and apt for use in clinical practise. However, financial and certification requirements are restricting the uptake. In 2024, Pauling C et al. [ 72 ] evaluated several commercially available AI models used to detect fractures, and found variations in pricing strategies for such models from a pay-per-use framework to an annual fee. This study highlighted the scarcity of models that are externally validated for clinical use and commercially available, in the United Kingdom post-Brexit. Pauling C et al emphasized need to develop models that are ready for use and certified by the Medical Devices Directive, the United Kingdom Conformity Assessed marking or similar bodies and certifications. Given the epidemiological burden of vertebral fractures, and the increasing constraints of healthcare systems globally, a cost-efficiency analysis is warranted to assess whether funding for AI technologies in spinal neurosurgery would have a significant positive impact at large. We undertook an exhaustive search of the literature, resulting in a study with a very large and high-powered pooled analysis. However, our findings must also be viewed in the context of the limitations of this study. Less than half of the studies included in the meta-analysis provided AUC data in the requisite format for analysis, with metrics such as specificity and sensitivity being more prevalent. The assessment of articles, in line with the PROBAST framework, revealed a general lack of information concerning missing data handling and overall data analysis procedures. Moreover, substantial variance in sample sizes was observed, with some studies having as few as 15 data points available for analysis. Additionally, confidence intervals were not consistently reported across the papers, necessitating our calculation of these intervals. Each study utilized a different AI model, each with its own parameters and methodologies. We aimed to account for the intrinsic weaknesses of the existing literature using a robust analytical approach, nonetheless it necessitates cautious interpretation of the results. Conclusion This meta-analysis, included 162 AI models suggests that AI based programmes can accurately diagnose and predict the risk of vertebral fractures, (predictive AUROC = 0.82 [0.78–0.85]; osteoporotic vertebral fracture diagnosis AUROC = 0.92 [0.88–0.96]; non-pathological vertebral fracture diagnosis AUROC = 0.85 [0.81–0.88] and vertebral compression fracture diagnosis AUROC = 0.87 [0.83–0.91]) at a significant level (p < 0.001). Traditional AI models accounted for the most successful predictive tools and deep learning models contributed to the most successful diagnostic tools. As such future development should be centred around this. However, given the high risk of bias in the papers included, likely including some level of selection and sampling bias, our findings should be interpreted with caution. We recognise the potential benefit of the widespread use of AI both predictively and diagnostically and highlight the need for a well-designed large multicentric study to further explore the benefits of AI in spine surgery, and answer questions on the practicality, efficacy, and cost-efficiency of the AI models in clinical practice. Abbreviations AI Artificial Intelligence ML Machine Learning US United States of America AUROC Area under the receiver operating characteristic PPV Positive predictive value NPV Negative predictive value FN False negative FP False positive TN True negative TP True positive AURPC Area Under Precision Recall Curve OVF Osteoporotic vertebral fracture VCF Vertebral compression fracture VF Vertebral fracture SDI Spinal deformity index CNN Convoluted neural network LSTM Long short-term memory networks MLP A multilayer perceptron Declarations Author contributions SRN was involved in conceptualisation, data curation, formal analysis, investigation, methodology, project administration, software, supervision, validation, visualisation, writing – original draft, and writing – review & editing. SSG, AP, AGK were involved in data curation, formal analysis, investigation, validation and writing – original draft. DSCR and HSP were involved in conceptualisation, writing – original draft, and writing – review & editing. SR, AS, AK, JN, DJ and DK were involved in conceptualisation, writing – review & editing. SGT was involved in conceptualisation, methodology, formal analysis, investigation, supervision, validation, visualisation, writing – original draft, and writing – review & editing. Competing Interests The authors declare no competing interests. Data availability statement All relevant data supporting the findings of this study can be accessed within the Supplementary Digital Content attached to the article. Additional Information The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. The authors have no relevant financial or non-financial interests to disclose. All data and materials as well as software application support their published claims and comply with field standards. Consent to publish has been received from all participants. References Ensrud, K. E. Epidemiology of fracture risk with advancing age. J. Gerontol. A Biol. Sci. Med. Sci. 68, 1236–1242 (2013). Donnally, C. J., DiPompeo, C. M. & Varacallo, M. Vertebral compression fractures. PubMed. Treasure Island (FL): StatPearls Publishing; 2024. The European Prospective Osteoporosis Study (EPOS) Group. Incidence of vertebral fracture in Europe: results from the European prospective osteoporosis study (EPOS). J. Bone Miner. Res. 17, 716–724 (2002). Freitas, S. S. et al. Rate and circumstances of clinical vertebral fractures in older men. Osteoporos. Int. 19, 615–623 (2007). Nevitt, M. C. et al. Risk factors for a first-incident radiographic vertebral fracture in women ≥ 65 years of age: the study of osteoporotic fractures. J. Bone Miner. Res. 20, 131–140 (2004). Savage, J. W., Schroeder, G. D. & Anderson, P. A. Vertebroplasty and kyphoplasty for the treatment of osteoporotic vertebral compression fractures. J. Am. Acad. Orthop. Surg. 22, 653–664 (2014). Cooper, C., Atkinson, E. J., O’Fallon, W. M. & Melton, J. L. Incidence of clinically diagnosed vertebral fractures: a population-based study in Rochester, Minnesota, 1985–1989. J. Bone Miner. Res. 7, 221–227 (2009). Fink, H. A. et al. What proportion of incident radiographic vertebral deformities is clinically diagnosed and vice versa? J. Bone Miner. Res. 20, 1216–1222 (2005). Ensrud, K. E. et al. Prevalent vertebral deformities predict mortality and hospitalization in older women with low bone mass. J. Am. Geriatr. Soc. 48, 241–249 (2000). Ross, P. D. Clinical consequences of vertebral fractures. Am. J. Med. 103, 30S-42S; discussion 42S-43S (1997). Thomas, B. Artificial intelligence: review of current and future applications in medicine. Fed. Pract. 38, (2021). Davenport, T. & Kalakota, R. The potential for artificial intelligence in healthcare. Future Healthc. J. 6, 94–98 (2019). Al-Antari, M. A. Artificial intelligence for medical diagnostics—existing and future AI technology! Diagnostics. 13, 688 (2023). Bajwa, J., Munir, U., Nori, A. & Williams, B. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc. J. 8, e188-e194 (2021). Hardy, M. & Harvey, H. Artificial intelligence in diagnostic imaging: impact on the radiography profession. Br. J. Radiol. 93, 20190840 (2019). Kurmis, A. P. & Ianunzio, J. R. Artificial intelligence in orthopedic surgery: evolution, current state and future directions. Arthroplasty. 4, (2022). De, A., Sarda, A., Gupta, S. & Das, S. Use of artificial intelligence in dermatology. Indian J. Dermatol. 65, 352 (2020). Tama, B. A., Kim, D. H., Kim, G., Kim, S. W. & Lee, S. Recent advances in the application of artificial intelligence in otorhinolaryngology-head and neck surgery. Clin. Exp. Otorhinolaryngol. 13, 326–339 (2020). Shen, L. et al. Using artificial intelligence to diagnose osteoporotic vertebral fractures on plain radiographs. J. Bone Miner. Res. (2023). Hong, N. et al. Deep-learning-based detection of vertebral fracture and osteoporosis using lateral spine X-ray radiography. J. Bone Miner. Res. 38, 887–895 (2023). Del Lama, R. S., Candido, R. M., Chiari-Correia, N. S., Nogueira-Barbosa, M. H. & de Azevedo-Marques, P. M. Computer-aided diagnosis of vertebral compression fractures using convolutional neural networks and radiomics. J. Digit. Imaging. 35, 446–458 (2022). Wolff, R. F. et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann. Intern. Med. 170, 51–58 (2019). Collins, G. S., Reitsma, J. B., Altman, D. G. & Moons, K. G. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 350, g7594 (2015). Heus, P. et al. Uniformity in measuring adherence to reporting guidelines: the example of TRIPOD for assessing completeness of reporting of prediction model studies. BMJ Open. 9, e025611 (2019). Chen, Y., Sun, X., Sui, X., Li, Y. & Wang, Z. Application of bone alkaline phosphatase and 25-oxhydryl-vitamin D in diagnosis and prediction of osteoporotic vertebral compression fractures. J. Orthop. Surg. Res. 18, 739 (2023). Yoon, M. A. et al. Automated segmentation of the fractured vertebrae on CT and its applicability in a radiomics model to predict fracture malignancy. Sci. Rep. 12, (2022). Ma, Y., Lu, Q., Yuan, F. & Chen, H. Comparison of the effectiveness of different machine learning algorithms in predicting new fractures after PKP for osteoporotic vertebral compression fractures. J. Orthop. Surg. Res. 18, (2023). Hu, X. et al. Prediction of subsequent osteoporotic vertebral compression fracture on CT radiography via deep learning. View (Beijing, China) 3, (2022). Kong, S. H. et al. Development of a spine X-ray-based fracture prediction model using a deep learning algorithm. Endocrinol. Metab. 37, 674–683 (2022). Yilmaz, E. B. et al. Assessing attribution maps for explaining CNN-based vertebral fracture classifiers. Lect. Notes Comput. Sci. 3, 3–12 (2020). Yilmaz, E. B. et al. Automated deep learning-based detection of osteoporotic fractures in CT images. Lect. Notes Comput. Sci. 376, 376–385 (2021). Monchka, B. A., Kimelman, D., Lix, L. M. & Leslie, W. D. Feasibility of a generalized convolutional neural network for automated identification of vertebral compression fractures: the Manitoba Bone Mineral Density Registry. Bone. 150, 116017 (2021). Monchka, B. A. et al. Development of a manufacturer-independent convolutional neural network for the automated identification of vertebral compression fractures in vertebral fracture assessment images using active learning. Bone. 161, 116427 (2022). Cho, S. T. et al. Prediction of progressive collapse in osteoporotic vertebral fractures using conventional statistics and machine learning. Spine. 48, 1535 (2023). Gui, C. et al. Radiomic modeling to predict risk of vertebral compression fracture after stereotactic body radiation therapy for spinal metastases. J. Neurosurg. Spine. 36, 294–302 (2022). Seol, Y. et al. Predicting vertebral compression fracture prior to spinal SBRT using radiomics from planning CT. Eur. Spine J. (2023). Murata, K. et al. Artificial intelligence for the detection of vertebral fractures on plain spinal radiography. Sci. Rep. 10, (2020). Chen, H. Y. et al. Application of deep learning algorithm to detect and visualize vertebral fractures on plain frontal radiographs. PLOS ONE. 16, e0245992 (2021). Biamonte, E. et al. Artificial intelligence-based radiomics on computed tomography of lumbar spine in subjects with fragility vertebral fractures. J. Endocrinol. Invest. 45, 2007–2017 (2022). Tomita, N., Cheung, Y. Y. & Hassanpour, S. Deep neural networks for automatic detection of osteoporotic vertebral fractures on CT scans. Comput. Biol. Med. 98, 8–15 (2018). Li, Y. C. et al. Can a deep-learning model for the automated detection of vertebral fractures approach the performance level of human subspecialists? Clin. Orthop. Relat. Res. (2021). Zhang, J. et al. Automated detection and classification of acute vertebral body fractures using a convolutional neural network on computed tomography. J. Bone Miner. Res. 14, (2023). Li, W. G. et al. The value of radiomics-based CT combined with machine learning in the diagnosis of occult vertebral fractures. BMC Musculoskelet. Disord. 24, 819 (2023). Eller-Vainicher, C. et al. Recognition of morphometric vertebral fractures by artificial neural networks: analysis from GISMO Lombardia database. PLoS ONE. 6, e27277 (2011). Muehlematter, U. J. et al. Vertebral body insufficiency fractures: detection of vertebrae at risk on standard CT images using texture analysis and machine learning. Eur. Radiol. 29, 2207–2217 (2018). Nicolaes, J. et al. Towards improved identification of vertebral fractures in routine computed tomography (CT) scans: development and external validation of a machine learning algorithm. J. Bone Miner. Res. 38, 1856–1866 (2023). Nicolaes, J. et al. External validation of a convolutional neural network algorithm for opportunistically detecting vertebral fractures in routine CT scans. Osteoporos. Int. 35, 143–152 (2024). Wang, X. et al. Value of 18F-FDG-PET/CT radiomics combined with clinical variables in the differential diagnosis of malignant and benign vertebral compression fractures. EJNMMI Res. 13, 89 (2023). Nicolaes, J. et al. Detection of vertebral fractures in CT using 3D convolutional neural networks. Lect. Notes Comput. Sci. 3, 3–14 (2020). Nicolaes, J. et al. External validation of a convolutional neural network algorithm for opportunistically detecting vertebral fractures in routine CT scans. Osteoporos. Int. 35, 143–152 (2024). Yabu, A. et al. Using artificial intelligence to diagnose fresh osteoporotic vertebral fractures on magnetic resonance images. Spine J. (2021). Yoda, T. et al. Automated differentiation between osteoporotic vertebral fracture and malignant vertebral fracture on MRI using a deep convolutional neural network. Spine. 47, E347-E352 (2022). Ono, Y. et al. A deep learning-based model for classifying osteoporotic lumbar vertebral fractures on radiographs: a retrospective model development and validation study. J. Imaging. 9, 187 (2023). Chen, W. et al. A deep-learning model for identifying fresh vertebral compression fractures on digital radiography. Bone. 32, 1496–1505 (2021). Kim, D. H. et al. Automated vertebral segmentation and measurement of vertebral compression ratio based on deep learning in X-ray images. J. Digit. Imaging. (2021). Chiari-Correia, N. S. et al. A 3D radiomics-based artificial neural network model for benign versus malignant vertebral compression fracture classification in MRI. J. Digit. Imaging. 36, 1565–1577 (2023). Liu, B. et al. Benign vs malignant vertebral compression fractures with MRI: a comparison between automatic deep learning network and radiologist’s assessment. Eur. Radiol. 33, 5060–5068 (2023). Duan, S. et al. Differential diagnosis of benign and malignant vertebral compression fractures: comparison and correlation of radiomics and deep learning frameworks based on spinal CT and clinical characteristics. Eur. J. Radiol. 165, 110899 (2023). Dong, Q. et al. Generalizability of deep learning classification of spinal osteoporotic compression fractures on radiographs using an adaptation of the modified-2 algorithm-based qualitative criteria. Acad. Radiol. 30, 2973–2987 (2023). Thawait, S. K. et al. Comparison of four prediction models to discriminate benign from malignant vertebral compression fractures according to MRI feature analysis. AJR Am. J. Roentgenol. 200, 493–502 (2013). Zhang, H. et al. Differentiation of benign versus malignant indistinguishable vertebral compression fractures by different machine learning with MRI-based radiomic features. Eur. Radiol. 33, 5069–5076 (2023). Ryu, S. M. et al. Diagnosis of osteoporotic vertebral compression fractures and fracture level detection using multitask learning with U-Net in lumbar spine lateral radiographs. Comput. Struct. Biotechnol. J. (2023). Feng, S. et al. Two-stream compare and contrast network for vertebral compression fracture diagnosis. IEEE Trans. Med. Imaging. 40, 2496–2506 (2021). Rinaldi, C. et al. The early detection of osteoporosis in a cohort of healthcare workers: is there room for a screening program? J. Clin. Endocrinol. Metab. 106, e485-e495 (2021). Sözen, T., Özışık, L. & Başaran, N. Ç. An overview and management of osteoporosis. Eur. J. Rheumatol. 4, 46–56 (2017). Shams, R. A., Zowghi, D. & Bano, M. AI and the quest for diversity and inclusion: a systematic literature review. AI Ethics. 4, 73–88 (2023). Yin, J., Ngiam, K. Y. & Teo, H. H. Role of artificial intelligence applications in real-life clinical practice: systematic review. J. Med. Internet Res. 23, e25743 (2021). Mittermaier, M., Raza, M. M. & Kvedar, J. C. Bias in AI-based models for medical applications: challenges and mitigation strategies. NPJ Digit. Med. 6, 27 (2023). Marwaha, J. S. & Kvedar, J. C. Crossing the chasm from model performance to clinical impact: the need to improve implementation and evaluation of AI. NPJ Digit. Med. 5, 25 (2022). Cimpeanu, T. et al. Artificial intelligence development races in heterogeneous settings. Sci. Rep. 12, 5729 (2022). Meng, F. et al. AI support for accurate and fast radiological diagnosis of COVID-19: an international multicenter, multivendor CT study. Eur. Radiol. 33, 4280–4291 (2022). Pauling, C. et al. Commercially available artificial intelligence tools for fracture detection: the evidence. BJR Open. 6, tzd005 (2023). Additional Declarations No competing interests reported. Supplementary Files AIinVFsupplementaryfile.docx Cite Share Download PDF Status: Published Journal Publication published 19 Dec, 2024 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 08 Aug, 2024 Reviews received at journal 28 Jul, 2024 Reviews received at journal 23 Jul, 2024 Reviews received at journal 20 Jul, 2024 Reviewers agreed at journal 17 Jul, 2024 Reviewers agreed at journal 15 Jul, 2024 Reviewers agreed at journal 14 Jul, 2024 Reviewers invited by journal 07 Jul, 2024 Editor assigned by journal 07 Jul, 2024 Editor invited by journal 07 Jul, 2024 Submission checks completed at journal 04 Jul, 2024 First submitted to journal 26 Jun, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4643800","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":330820051,"identity":"7a65186f-764f-4d61-a200-34c1dca4d4a9","order_by":0,"name":"Srikar R Namireddy","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Srikar","middleName":"R","lastName":"Namireddy","suffix":""},{"id":330820052,"identity":"19ce7f35-02bc-4be3-8391-7898b5ea16ae","order_by":1,"name":"Saran S Gill","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Saran","middleName":"S","lastName":"Gill","suffix":""},{"id":330820054,"identity":"078e2544-8ca5-435e-8d0b-1adb90b66aca","order_by":2,"name":"Amaan Peerbhai","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Amaan","middleName":"","lastName":"Peerbhai","suffix":""},{"id":330820056,"identity":"3631f314-e4d7-4d6c-913f-7e746d77676b","order_by":3,"name":"Abith G Kamath","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Abith","middleName":"G","lastName":"Kamath","suffix":""},{"id":330820058,"identity":"3a68b6a6-300f-46f2-a8fb-1dead7f4e079","order_by":4,"name":"Daniele S. C. Ramsay","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Daniele","middleName":"S. C.","lastName":"Ramsay","suffix":""},{"id":330820061,"identity":"ae611def-e01f-4db9-a5e7-e953de3ae7a7","order_by":5,"name":"Hariharan Subbiah Ponniah","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Hariharan","middleName":"Subbiah","lastName":"Ponniah","suffix":""},{"id":330820062,"identity":"0467ca51-b766-4027-91e3-fcf171502f99","order_by":6,"name":"Ahmed Salih","email":"","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":false,"prefix":"","firstName":"Ahmed","middleName":"","lastName":"Salih","suffix":""},{"id":330820065,"identity":"8306d709-4b64-45c7-9f1c-e35f5c7f7a04","order_by":7,"name":"Dragan Jankovic","email":"","orcid":"","institution":"University Medical Center Mainz","correspondingAuthor":false,"prefix":"","firstName":"Dragan","middleName":"","lastName":"Jankovic","suffix":""},{"id":330820069,"identity":"d0b654fc-7fdf-4295-8743-dfdefc464bd6","order_by":8,"name":"Darius Kalasauskas","email":"","orcid":"","institution":"University Medical Center Mainz","correspondingAuthor":false,"prefix":"","firstName":"Darius","middleName":"","lastName":"Kalasauskas","suffix":""},{"id":330820070,"identity":"18546ff6-a0bf-4f7d-82cb-5430ccbf3ecc","order_by":9,"name":"Jonathan Neuhoff","email":"","orcid":"","institution":"Berufsgenossenschaftliche Unfallklinik Frankfurt am Main","correspondingAuthor":false,"prefix":"","firstName":"Jonathan","middleName":"","lastName":"Neuhoff","suffix":""},{"id":330820073,"identity":"43562c33-90f3-442b-8aec-058cbee55ce0","order_by":10,"name":"Andreas Kramer","email":"","orcid":"","institution":"University Medical Center Mainz","correspondingAuthor":false,"prefix":"","firstName":"Andreas","middleName":"","lastName":"Kramer","suffix":""},{"id":330820076,"identity":"af26f987-81c9-42c2-9532-1ebd5588c284","order_by":11,"name":"Salvatore Russo","email":"","orcid":"","institution":"Imperial College Healthcare NHS Trust","correspondingAuthor":false,"prefix":"","firstName":"Salvatore","middleName":"","lastName":"Russo","suffix":""},{"id":330820077,"identity":"494d5f70-7f31-48ca-8daa-de26872a95b7","order_by":12,"name":"Santhosh G. Thavarajasingam","email":"data:image/png;base64,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","orcid":"","institution":"Imperial Brain \u0026 Spine Initiative","correspondingAuthor":true,"prefix":"","firstName":"Santhosh","middleName":"G.","lastName":"Thavarajasingam","suffix":""}],"badges":[],"createdAt":"2024-06-26 15:38:48","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4643800/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4643800/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-024-75628-2","type":"published","date":"2024-12-19T15:56:53+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":61328259,"identity":"c40d4c25-dbcd-4f16-87bc-b43e82d78884","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":473803,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea: \u003c/strong\u003eThe preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart outlining the study selection process is shown.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb: \u003c/strong\u003eA world map indicated the origin of publications included in this study (n = 78) (Superscript references). The countries are coloured according to whether n = 1, 2, 3, 4, 5, 6, 11 or 19 studies from these countries have been included in this systematic review. Following countries are coloured: Germany (n = 5), China (n = 19), South Korea (n = 11), United States of America (n = 11), Brazil (n = 2), Japan (n = 4), Taiwan (n = 4), Italy (n = 3), Canada (n = 3), Switzerland (n = 6), Denmark (n = 1), India (n = 3), Australia (n = 3), Belgium (n = 1), Israel (n = 2), Philippines (n = 1), Poland (n = 1), United Kingdom (n = 1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec: \u003c/strong\u003eA risk of bias summary plot for all included studies (n = 78) across the domains of the Prediction model Risk Of Bias Assessment Tool (PROBAST)\u003c/p\u003e","description":"","filename":"f1.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/267c931d31546ca83e533286.png"},{"id":61329012,"identity":"ccbc659a-5d27-44f6-a7f5-a32093b7bde9","added_by":"auto","created_at":"2024-07-29 14:30:49","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":129958,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea: \u003c/strong\u003eBar plot visualizes the number of prospective (n = 8), retrospective (n = 69) and ambispective (n = 1) studies included in the systematic review (n = 78) (Superscript references)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb:\u003c/strong\u003e Bar plot visualizes the number of studies with certain sample sizes: 0-99 (n = 11), 100-999 (n = 37), 1000-9999 (n = 20), 10000+ (n = 7), unsure (n = 3)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec:\u003c/strong\u003e Line plot displays the number of studies for the following years of publications: 2011 (n = 1), 2012 (n = 1), 2013 (n = 1), 2017 (n = 3), 2018 (n = 3), 2019 (n = 2), 2020 (n = 9), 2021 (n = 13), 2022 (n = 12), 2023 (n = 29), 2024 (n = 4). Each year is indicated as a blue circle, and the circles are connected by an interrupted line to visualise the trend more clearly.\u003c/p\u003e","description":"","filename":"f2.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/763140846d94331aeba5f0d4.png"},{"id":61328260,"identity":"8a4728fb-327d-4ec0-bfc4-7cce92e2d4d3","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":343563,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea:\u003c/strong\u003e This Sankey diagram represents the categorization of 162 total artificial intelligence (AI) models into various types and performance levels based on their AUROC scores. Of the total, 136 models are designated as diagnostic AI models, while 26 are predictive AI models. The performance levels, corresponding to different ranges of AUROC scores, are color-coded and flow from these categories into four distinct performance categories: Suboptimal performance (AUROC score: 0.5-0.7) includes 9 models, acceptable performance (AUROC score: 0.7-0.8) includes 25 models, excellent performance (AUROC score: 0.8-0.9) includes 20 models, and outstanding performance (AUROC score: 0.9+) includes 49 models. Additionally, there are 59 models for which no AUROC values are provided. Diagnostic AI models are further broken down into osteoporotic vertebral fractures (OVFs) with 11 models, vertebral compression fractures (VCFs) with 39 models, and non-pathological vertebral fractures (non-pathological VFs) with 30 models. Each subgroup of fractures feeds into the various performance levels, showing the distribution of models’ performance based on their diagnostic category.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb:\u003c/strong\u003e This radar chart provides a comparative visualization of the mean performance metrics for different groups of AI models. The chart is segmented into five performance metrics: AUROC, Accuracy, Precision, Sensitivity, and Specificity, with values ranging from 0 to 1. There are four groups of models compared: Predictive Models, Non-Pathological Vertebral Fracture (VF) Diagnostic Models, Osteoporotic Vertebral Fractures (OVF) Diagnostic Models, and Vertebral Compression Fractures (VCF) Diagnostic Models. Each group is represented by a different coloured line that traces the mean score for each performance metric. The lines create shapes that allow for an at-a-glance comparison of how each model group performs across these metrics. The closer the edge of a shape is to the outer perimeter of the radar chart, the higher the mean performance score for that metric. The chart facilitates a direct comparison of the model groups, indicating areas where some models excel or where there may be room for improvement.\u003c/p\u003e","description":"","filename":"f3.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/ac2776d69bc45d7bd7337137.png"},{"id":61328262,"identity":"8e335800-70a1-463e-8845-6e19b7016528","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":423673,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ea:\u003c/strong\u003e presents a violin plot comparing the performance of different AI models in predicting vertebral fractures. Three strata of models are displayed: 'Specialised ensemble models,' 'Deep learning models,' and 'Traditional machine learning models,' with their respective AUROC scores. The width of each violin represents the distribution density of the AUROC scores, with wider sections indicating a higher frequency of scores in that range. The box within each violin shows the interquartile range, and the line within denotes the median AUROC score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eb:\u003c/strong\u003e presents a violin plot comparing the performance of different AI models in diagnosing non-pathological vertebral fractures. Four strata of models are displayed: 'Probabilistic models,' 'Specialised ensemble models,' 'Deep learning models,' and 'Traditional machine learning models,' with their respective AUROC scores. The width of each violin represents the distribution density of the AUROC scores, with wider sections indicating a higher frequency of scores in that range. The box within each violin shows the interquartile range, and the line within denotes the median AUROC score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ec:\u003c/strong\u003e presents a violin plot comparing the performance of different AI models in diagnosing OVFs. Two strata of models are displayed: ‘Specialised/ensemble deep learning models,' and 'Deep learning models,' with their respective AUROC scores. The width of each violin represents the distribution density of the AUROC scores, with wider sections indicating a higher frequency of scores in that range. The box within each violin shows the interquartile range, and the line within denotes the median AUROC score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ed:\u003c/strong\u003e presents a violin plot comparing the performance of different AI models in diagnosing VCFs. Three strata of models are displayed: 'Specialised ensemble models,' 'Deep learning models,' and 'Traditional machine learning models,' with their respective AUROC scores. The width of each violin represents the distribution density of the AUROC scores, with wider sections indicating a higher frequency of scores in that range. The box within each violin shows the interquartile range, and the line within denotes the median AUROC score.\u003c/p\u003e","description":"","filename":"f4.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/96f1e5d8a109e30ddd8e2186.png"},{"id":61328264,"identity":"6b7d6990-b2b5-488d-982e-fa87e2e13b81","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":326299,"visible":true,"origin":"","legend":"\u003cp\u003eA forest plot displaying the predictive performance of various statistical models is presented, pooling the results from several studies conducted between 2020 and 2023. Each study is listed with the author, the year of publication, and the specific model used, such as neural networks, decision trees, or convolutional neural networks (CNNs). The predictive accuracy of each model is quantified by the AUROC (Area Under the Receiver Operating Characteristic curve), with the size of the grey square indicating the model's performance and correlating to the sample size of the study. The horizontal lines represent the 95% confidence intervals (CI) for the AUROC, and the overall pooled predictive accuracy across all studies is illustrated by the diamond at the bottom of the plot. This summary measure combines the strength of evidence from the individual studies. Heterogeneity in study outcomes is expressed through the I² statistic and its associated tau² (τ²) and p-value, providing insight into the variability among the different predictive models. A p-value less than 0.05 indicates statistically significant predictive accuracy. The weighting of each study, displayed as a percentage, is based on the inverse of the variance, granting more influence to studies with more precise effect estimates.\u003c/p\u003e","description":"","filename":"f5.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/6bef7ce1e80a161ce922b569.png"},{"id":61328263,"identity":"7f807e99-daed-499c-badd-45773cee3704","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":341028,"visible":true,"origin":"","legend":"\u003cp\u003eThis comprehensive forest plot aggregates the diagnostic accuracies of a multitude of studies, evaluating the AUROC (Area Under the Receiver Operating Characteristic curve) of various diagnostic models in the medical field tailored towards identifying non-pathological vertebral fractures. Each entry details the study by author, publication year, and utilized model or technique, ranging from advanced algorithms like CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory networks) to ensemble methods and radiomic analyses. The size of the grey squares reflects the study's sample size, directly influencing the visual weight of each study's AUROC result on the plot. The black horizontal lines spanning from each square represent the 95% confidence intervals, providing a graphical representation of the estimate's precision. At the plot's base, the black diamond summarizes the combined AUROC across all studies, indicating the overall predictive strength of these models. Heterogeneity among the studies' outcomes is quantified by an I² statistic, tau² (τ²), and p-value, signaling the extent of variability and its statistical significance. Studies with higher weights, denoted in percentages, suggest a greater impact on the pooled result due to their lower variance. This plot serves as a critical summary, enabling readers to visualize the efficacy of various predictive models in a specific medical domain.\u003c/p\u003e","description":"","filename":"f6.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/9c75dff3f41b95194495666e.png"},{"id":61328265,"identity":"5655da8b-9689-4083-b99a-e3e577a901a6","added_by":"auto","created_at":"2024-07-29 14:22:49","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":154962,"visible":true,"origin":"","legend":"\u003cp\u003eThis comprehensive forest plot aggregates the diagnostic accuracies of a multitude of studies, evaluating the AUROC (Area Under the Receiver Operating Characteristic curve) of various diagnostic models in the medical field tailored towards identifying osteoporotic vertebral fractures. Each entry details the study by author, publication year, and utilized model or technique, ranging from advanced algorithms like CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory networks) to ensemble methods and radiomic analyses. The size of the grey squares reflects the study's sample size, directly influencing the visual weight of each study's AUROC result on the plot. The black horizontal lines spanning from each square represent the 95% confidence intervals, providing a graphical representation of the estimate's precision. At the plot's base, the black diamond summarizes the combined AUROC across all studies, indicating the overall predictive strength of these models. Heterogeneity among the studies' outcomes is quantified by an I² statistic, tau² (τ²), and p-value, signalling the extent of variability and its statistical significance. Studies with higher weights, denoted in percentages, suggest a greater impact on the pooled result due to their lower variance. This plot serves as a critical summary, enabling readers to visualize the efficacy of various predictive models in a specific medical domain.\u003c/p\u003e","description":"","filename":"f7.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/fa55d4a01c5604ccec877f47.png"},{"id":61328266,"identity":"f21d7cc4-d972-4dc7-b01b-e75addc4b1d5","added_by":"auto","created_at":"2024-07-29 14:22:50","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":435533,"visible":true,"origin":"","legend":"\u003cp\u003eThis comprehensive forest plot aggregates the diagnostic accuracies of a multitude of studies, evaluating the AUROC (Area Under the Receiver Operating Characteristic curve) of various diagnostic models in the medical field tailored towards identifying vertebral compression fractures. Each entry details the study by author, publication year, and utilized model or technique, ranging from advanced algorithms like CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory networks) to ensemble methods and radiomic analyses. The size of the grey squares reflects the study's sample size, directly influencing the visual weight of each study's AUROC result on the plot. The black horizontal lines spanning from each square represent the 95% confidence intervals, providing a graphical representation of the estimate's precision. At the plot's base, the black diamond summarizes the combined AUROC across all studies, indicating the overall predictive strength of these models. Heterogeneity among the studies' outcomes is quantified by an I² statistic, tau² (τ²), and p-value, signalling the extent of variability and its statistical significance. Studies with higher weights, denoted in percentages, suggest a greater impact on the pooled result due to their lower variance. This plot serves as a critical summary, enabling readers to visualize the efficacy of various predictive models in a specific medical domain.\u003c/p\u003e","description":"","filename":"f8.png","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/a5add37c370e84efc66ead7d.png"},{"id":72201443,"identity":"df449eae-95e5-44a7-abfc-8a71bbf5afa1","added_by":"auto","created_at":"2024-12-23 16:05:04","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3026227,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/216e6a4d-2884-4b3a-9a18-a1a916894b67.pdf"},{"id":61328267,"identity":"7ec503fe-f74a-41de-b0c0-eaaf8f7765e0","added_by":"auto","created_at":"2024-07-29 14:22:51","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":72494,"visible":true,"origin":"","legend":"","description":"","filename":"AIinVFsupplementaryfile.docx","url":"https://assets-eu.researchsquare.com/files/rs-4643800/v1/2291fcdbb76ee938fd3e6b42.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial Intelligence and Machine Learning for Risk Prediction and Diagnosis of Vertebral Fractures: A Systematic Review and Meta-Analysis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eVertebral fractures, as the most frequent type of fragility fractures, are a hallmark of osteoporosis, particularly among the elderly. Studies in Europe show that for individuals aged 50 and older, the incidence rates of new vertebral fractures stand at 10.7 per 1000 person-years for women and 5.7 per 1000 person-years for men [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e][\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. In the US, they account for up to 1.5\u0026nbsp;million cases per year [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Risk factors include inactivity, chronic conditions (such as osteoporosis), smoking and previous falls [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e][\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. With the rate of osteoporosis reported to be rising [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e], the subsequent incidence of vertebral fractures is also predicted to increase. Vertebral fractures, unlike fractures of other areas of the skeleton, tend not to be treated at the time of injury, with up to 33% going undetected [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e][\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. This results in an increased risk of mortality after such injuries [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], and can lead to chronic pain and disability in the long term, with significant economic ramifications [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. As such, the timely detection and treatment of vertebral fractures has become a key challenge for healthcare providers.\u003c/p\u003e \u003cp\u003eWhile Artificial Intelligence (AI) and Machine Learning (ML) are no longer novel concepts, the rise in their clinical usage has been exponential rather recently [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e][\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e][\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Multimodal data, along with the development of the ethical framework surrounding AI, have had an impact in the uptake of AI within the medical field [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Diagnostically, AI based systems are currently being used, and have potential, to speed up and improve the precision in diagnostic medicine [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Clinically, AI models have been used heavily within dermatology, orthopaedics, and otorhinolaryngology demonstrate the utility of such models in different medical specialties [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e][\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e][\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. However, the uptake of AI in clinical spinal neurosurgery has been less pronounced.\u003c/p\u003e \u003cp\u003eThe current approach to diagnosing and classifying vertebral fractures involves different members of a multidisciplinary team. The combined clinical experience can often be limited by intrinsic risks of inaccuracies and lack of efficiency. As such, the use of AI and ML in such situations is of interest [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e][\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e][\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. However, a robust analysis including both qualitative and quantitative synthesis is required evaluate its use in this context \u0026ndash; however such an analysis does not exist currently. Hence, this systematic review aims to assess the literature surrounding the use of AI and Machine Learning in the detection and prognostication of vertebral fractures.\u003c/p\u003e "},{"header":"Methodology","content":"\n\u003ch3\u003eLiterature Search Strategy\u003c/h3\u003e\n\u003cp\u003eThis systematic review was conducted using the guidelines outlined by the Cochrane Collaboration, and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA). The completed PRISMA flowchart is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eA. The literature search was carried out on 12.02.2024, using a search of MEDLINE, Embase, Scopus, PubMed and Web of Science Library. Search strings were created for the following research question: \u0026ldquo;Is AI an effective and accurate tool for predicting and diagnosis vertebral fractures?\u0026rdquo;. The search string can be found in Supplemental Digital Content 1: Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eInclusion and exclusion criteria\u003c/h2\u003e \u003cp\u003eThe inclusion and exclusion criteria can be found in Supplemental Digital Content 1: Supplementary Table S2. Vertebral fractures were defined as the breakage or collapse of one or more bones in the spine, often leading to pain, reduced mobility, and potential changes in posture. Only studies that used machine learning or artificial intelligence tools for the diagnostic and prognostication of vertebral fractures were included in the meta-analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eScreening and appraisal\u003c/h2\u003e \u003cp\u003eIdentified studies were uploaded to COVIDENCE for duplicate removal and title and abstract screening. In the first abstract screening, conducted by four reviewers (SG, AGK, SRN, AP). All original articles in the English language that reported on vertebral fractures were included. Subsequently, only studies reporting on artificial intelligence and machine learning tools for diagnosis and/or prognostication which also fulfilled our inclusion criteria were included. All included papers were assessed by two independent reviewers. Any disagreements were resolved by consensus after discussion with SRN and HSP.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eCritical appraisal\u003c/h2\u003e \u003cp\u003eTwo evaluators independently used the Prediction model Risk Of Bias Assessment Tool (PROBAST) to gauge potential biases in the studies analyzed. PROBAST examines four key aspects: participants, predictors, outcomes, and analysis. Within these areas, biases related to participant selection, prediction methods, outcome determination, and data analysis were scrutinized using specific guiding questions. Discrepancies in study quality were resolved by a third reviewer [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. In our review, adherence to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines was rigorously evaluated by two independent researchers for each included study. TRIPOD provides a comprehensive checklist of 22 essential items aimed at enhancing the transparency and completeness of reporting in studies developing, validating, or updating prediction models for diagnostic or prognostic purposes [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e][\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eData preparation was performed using SPSS (IBM, USA) Version 28.0.0.0. Subsequently, R software (version 4.4.3) was used for statistical analysis and forest plot synthesis, by utilising the meta package. Firstly, a Random Effects model meta-analysis was performed for AUROC among models predicting the risk of vertebral fractures. Similar such plots were created for models aiming to diagnose non-pathological vertebral fractures, osteoporotic vertebral fractures and vertebral compression fractures. All outcome variable computation included 95%-CI, as well as heterogeneity measured by the I\u003csup\u003e2\u003c/sup\u003e test. An influence analysis was conducted to exclude outliers and a meta regression was calculated to look for correlations between the metrics using a mixed-effects single variate meta-regression. Correlation coefficients, standard errors and p-values were determined. A p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eA total of 14,161 studies were screened. From these, 165 full texts were assessed using our inclusion criteria. A total of 78 studies were included in this systematic review. 35 of these studies were also included in the meta-analysis. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eB depicts a world map, with the origin of each paper highlighted. Risk of bias was assessed using the PROBAST framework, all included studies were concluded to have a high overall risk of bias. The summary plot of characteristics for all studies included in the systematic review can be found in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003ec. The majority of studies included (69/78) were retrospective in design (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e2\u003c/span\u003ea), had a sample size of 100 to 999 (37/78)(Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e2\u003c/span\u003eb) and were published in the year 2023 (29/78)(Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e2\u003c/span\u003ec).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePrediction of Vertebral Fractures\u003c/h2\u003e \u003cp\u003eThe part of this systematic review focussing on the use of AI in prediction of vertebral fractures consisted of 9 studies, encompassing 26 trial arms (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). Specificity and AUROC, sensitivity and specificity were among the most commonly reported metrics, with 78%, 67% and 44% of papers including these, respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). 56% of these papers reference convoluted neural networks directly. Of the 9 included papers, 56% (n\u0026thinsp;=\u0026thinsp;5) were published in 2023, 33% (n\u0026thinsp;=\u0026thinsp;3) were published in 2022 and the remaining 11% (n\u0026thinsp;=\u0026thinsp;1) was published in 2020. Specific studies like those of Chen Y et al. [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], Park T et al. [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], and Ma Y et al. [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] concentrated on vertebral compression fractures, whereas Hu X et al. [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] and Kong HS et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] focused on osteoporotic fractures, with Kong\u0026rsquo;s study noting higher sample sizes and more comprehensive AUROC evaluations. The findings are summarised in Supplemental Digital Content 1: Supplementary Table S3.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eDiagnosis \u0026amp; Classification of Vertebral Fractures\u003c/h2\u003e \u003cp\u003eThe part of this systematic review focussing on the use of AI in the diagnosis of vertebral fractures is based on 70 studies, consisting of over 130 diagnostic models in total. Sensitivity and specificity were the two most commonly reported metrics, followed by accuracy, with 97%, 94% and 91% of papers including these, respectively. We categorized the studies based on the type of vertebral fractures: Non-Pathological Vertebral Fractures, Osteoporotic Vertebral Fractures, and Vertebral Compression Fractures. These studies commonly aimed to detect the presence of fractures using expert opinions for validating AI model outputs. Noteworthy contributions include Hong N et al.\u003csup\u003e20\u003c/sup\u003e, who utilized a qualitative algorithm to classify vertebral fractures, with large datasets allowing robust comparisons across different scoring systems like the VERTE-X pVF and VERTE-X osteo scores. Similarly, Yilmaz EB et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e][\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] and Monchka BA et al. [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e][\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] employed convolutional neural networks and a modified algorithm-based qualitative approach, respectively, to classify fractures, focusing on binary outcomes\u0026mdash;either 'fracture' or 'no fracture'.\u003c/p\u003e \u003cp\u003eThe findings are summarised in Supplemental Digital Content 1: Supplementary Tables S4.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003ePerformance Breakdown of Vertebral Fracture Models\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e3\u003c/span\u003ea summarises the performance of 162 AI models into a decisive visualization of efficacy. With 136 models focused on diagnosis and 26 on prediction, the diagnostic models are further categorized by fracture type: 11 for osteoporotic fractures (OVFs), 39 for vertebral compression fractures (VCFs), and 30 for non-pathological fractures. Performance-wise, 49 models are at the forefront with outstanding AUROC scores above 0.9. Meanwhile, 20 models show excellent performance, 25 have acceptable levels, and 9 fall under suboptimal, reflecting a high-precision stratification in the field. The Sankey diagram underscores the concentration of superior AI models within the diagnostic realm, particularly in the detection of OVFs and VCFs, despite a notable 59 models lacking AUROC data.\u003c/p\u003e \u003cp\u003eIn the evaluation of AI models for predicting vertebral fractures (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003ea), traditional machine learning models show the highest median AUROC scores, indicating a stronger predictive performance compared to specialised ensemble and traditional machine learning models. For the diagnosis of non-pathological vertebral fractures deep learning models exhibit the highest median AUROC scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). In the context of diagnosing osteoporotic vertebral fractures (OVFs) as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003ec, specialized ensemble deep learning models showed very similar performance simple deep learning models. Lastly, for the diagnosis of vertebral compression fractures (VCFs) deep learning models again lead with higher median AUROC scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e4\u003c/span\u003ed).\u003c/p\u003e \u003cp\u003eMeta-Analysis\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003ePrediction Vertebral Fractures\u003c/h2\u003e \u003cp\u003eThe meta-analysis [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e][\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e][\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e][\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e][\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e][\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e][\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e5\u003c/span\u003e) compares different machine learning models and their effectiveness in predicting a certain outcome. With AUROCs ranging from 0.72 to 0.94, it is evident that some models perform significantly better than others. Models by Ma et al. (2023) utilizing logistic regression, gradient boosting machine, and neural networks, and Yoon et al. (2022) with CNN, achieved high predictive accuracy, with AUROCs at or above 0.90. In contrast, several models, particularly those by Cho et al. (2023) and Kong et al. (2020), show relatively lower accuracy, with AUROCs closer to 0.72. The overall predictive performance across all models, indicated by the RE Model's AUROC of 0.82, suggests excellent predictive accuracy by the models, though there is substantial heterogeneity (I\u0026sup2; \u0026gt; 99%, p\u0026thinsp;\u0026lt;\u0026thinsp;0.01).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eDiagnosis/Classification of Non-Pathological Vertebral Fractures\u003c/h2\u003e \u003cp\u003eThe forest plot [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e][\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e][\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e][\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e][\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e][\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e][\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e][\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e][\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e][\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e][\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e][\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e][\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e][\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e6\u003c/span\u003e) in question provides a comprehensive overview of the predictive accuracy of various machine learning models, as measured by AUROC. There is a notable range in performance, with AUROC values spanning from roughly 0.68 to a near perfect score of 0.99. Models by Li et al. applying ensemble deep learning techniques to different grades of fractures in 2021, demonstrated near-perfect predictive capabilities. Meanwhile, the study by Wu-Gen Li et al. explored a variety of methods including Support Vector Machine (SVM), Bayesian analysis, and logistic regression, only to display a wide array of outcomes with moderate to high accuracy. On the contrary, the models by E\u0026szlig;er-Vainicher et al. which utilised CNNs on patients where SDI\u0026thinsp;\u0026ge;\u0026thinsp;1, show lower AUROCs. The aggregate predictive accuracy across all models is indicated by the Random Effects (RE) Model's AUROC of 0.85, suggesting excellent performance. Nonetheless there is high heterogeneity (I\u0026sup2; \u0026gt; 99%, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eDiagnosis/Classification of Osteoporotic Vertebral Fractures\u003c/h2\u003e \u003cp\u003eThe forest plot [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e][\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e][\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e][\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e][\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e][\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e7\u003c/span\u003e) presents a comparative analysis of machine learning models based on their AUROC values for predicting specific outcomes. The models investigated show a considerable spread in performance, with AUROC values ranging from 0.77 to near perfection at 0.99. The models devised by Hong et al. in 2023 exhibit varying results, with internal assessments resulting in AUROCs of 0.93 and 0.85 for PVF and osteo scores respectively, indicating a solid predictive capability, whereas their external assessments reveal a slightly reduced accuracy. Yabu et al. and Yoda et al. through their incorporation of multiple CNN architectures demonstrate superior predictive performance, particularly Yoda et al. with an AUROC close to 1.00, showing an excellent fit for the predictive task. Ono et al. created a model that utilised a combination of Resnet-50, DenseNet-161, and NexResNet-50, however this resulted in a lower AUROC of 0.77, which could imply limitations in their data, or the combination of AI models used. Yilmaz et al. across three studies in 2020 and 2021 employing U-Net, CNN, and Fnet, consistently showcased high prediction accuracy, with two studies achieving AUROCs of 0.99. The combined predictive accuracy, as summarized by the Random Effects (RE) Model, reported an AUROC of 0.92, showing that on average, the models are outstandingly accurate in their predictions. However there is high heterogeneity (I\u0026sup2;=99.16, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eDiagnosis/Classification of Vertebral Compression Fractures\u003c/h2\u003e \u003cp\u003eThe forest plot [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e][\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e][\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e][\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e][\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e][\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e][\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e][\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e][\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e][\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e][\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e][\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e][\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e] (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e8\u003c/span\u003e) provided details the performances of a diverse set of machine learning models, as denoted by their AUROC values. These models range from deep learning CNNs to traditional methods like logistic regression and decision trees. The variability in performance is significant, with AUROCs as high as 0.99 for some ensemble CNN methods by Moncicka et al. down to 0.54 for certain individual models. This broad performance spectrum is further reflected in models by Zhang et al. with AUROCs spanning from 0.60 to 0.73 across different algorithmic approaches like k-nearest neighbours (KNN), logistic regression (LR), decision trees (DT), and gradient boosting (GB). The models demonstrate that ensemble methods, particularly those involving CNNs, tend to yield higher predictive accuracies (such as the study by Kim et al. which achieved an AUROC of 0.99), while traditional machine learning methods like those by Thawani et al. hovered around the 0.76 mark. The plot culminates in a Random Effects (RE) Model AUROC of 0.87. However, extreme heterogeneity (I\u0026sup2;=99.95, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) was calculated in the meta-analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eInfluence analysis and linear regression\u003c/h2\u003e \u003cp\u003eThe exclusion of outlier studies based on an influence analysis did not yield a significant change in effect size. The exclusion of outlier studies based on high levels of risk of bias scoring did not yield any significant changes to effect size of any of the outcome variables. The meta-regressions scored the influence of all co-variates on the overall effect size of the predictive AI models meta-analysis, non-pathological VF diagnostic AI models meta-analysis, OVF diagnostic AI models meta-analysis, VCF diagnostic AI meta-analysis (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). No significant (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) co-variates were found.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThe table presents the outcomes of the meta-regression analysis assessing the influence of various covariates on the performance of AI models in predicting and diagnosing different types of vertebral fractures. The covariates analysed include sample size, study type, study design, model type, validation method, imaging modality, image preprocessing, feature engineering, and year of publication. Regression coefficients with their corresponding standard errors (in round brackets) are provided for each covariate across four distinct model performance meta-analyses: Prediction, Non-Pathologic Vertebral Fracture Diagnosis, Osteoporotic Vertebral Fracture Diagnosis, and Vertebral Compression Fracture Diagnosis. P-values are shown next to the regression coefficients and standard errors, with the understanding that values greater than 0.05 indicate non-significance. The different explanatory variables were calculated singularly as sole covariates in separate meta-regression.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrediction\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNon-Pathologic VF Diagnosis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eOVF Diagnosis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eVCF Diagnosis\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSample size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.0003 [0.0002], 0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0001 [0.0001], 0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.0002 [0.00015], 0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.00025 [0.00018], 0.11\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.01 [0.02], 0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.02 [0.03], 0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.015 [0.025], 0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.018 [0.028], 0.65\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy design\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-0.005 [0.01], 0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.003 [0.01], 0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e-0.004 [0.008], 0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.006 [0.009], 0.80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.02 [0.015], 0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.018 [0.012], 0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.021 [0.016], 0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.017 [0.013], 0.09\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation method\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.01 [0.02], 0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.009 [0.019], 0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.008 [0.018], 0.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.007 [0.017], 0.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImaging modality\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.03 [0.025], 0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.027 [0.022], 0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.032 [0.03], 0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.026 [0.02], 0.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImage preprocessing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-0.015 [0.012], 0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.01 [0.008], 0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e-0.013 [0.01], 0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.011 [0.009], 0.20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature engineering\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.005 [0.007], 0.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.004 [0.006], 0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.006 [0.008], 0.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.003 [0.005], 0.23\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eYear of publication\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e-0.001 [0.002], 0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e-0.0008 [0.0015], 0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e-0.0011 [0.0018], 0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e-0.0009 [0.0016], 0.50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis meta-analysis is the first to formally assess and analyse the use of AI in prediction, diagnosis and classification of vertebral Fractures. It encompasses 39 studies incorporating data from 162 AI models. Our findings indicate that AI models exhibit an overall robust predictive capacity (AUROC\u0026thinsp;=\u0026thinsp;0.82 [0.78\u0026ndash;0.85]) and diagnostic accuracy (osteoporotic vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.92 [0.88\u0026ndash;0.96]; non-pathological vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.85 [0.81\u0026ndash;0.88] and vertebral compression fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.87 [0.83\u0026ndash;0.91]), all being statistically significant at p\u0026thinsp;\u0026lt;\u0026thinsp;0.001.\u003c/p\u003e \u003cp\u003eOur systematic review showed that traditional machine learning excels in predicting vertebral fractures, topping AUROC scores and proving its predictive reliability. Conversely, deep learning had the best accuracy in diagnosing all 3 types of vertebral fractures. Future AI should merge traditional machine learning's predictive precision with deep learning's diagnostic acuity for vertebral fracture assessment.\u003c/p\u003e \u003cp\u003eThe high predictive AUROC supports the narrative that AI can play a vital role in pre-empting fractures, an insight that dovetails with existing literature emphasizing early detection and intervention in osteoporotic conditions [\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e]. The potential of such technology to forecast risk and inform clinical decision-making prior to fracture occurrence is not only innovative but aligns with the preventive care model that is becoming increasingly crucial in an aging [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e]. Nevertheless, there remains a need for a nuanced understanding of the models\u0026rsquo; performance across diverse demographic and clinical settings, echoing calls for broader and more inclusive datasets in AI training [\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the realm of diagnosis, AI models showed particular promise in distinguishing between non-pathological, osteoporotic, and other types of vertebral fractures. These findings prompt a re-evaluation of traditional diagnostic methods, which may be augmented or, in some instances, surpassed by AI capabilities. However, the clinical integration of these models requires careful consideration of their performance in real-world settings. The consistency and reliability of AI model outputs against the gold standard of clinical diagnoses present an ongoing area of research that must address the full spectrum of clinical scenarios [\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e]. Notably, while AI models demonstrate considerable strengths, our analysis identified areas where performance is less than optimal, particularly in the prediction of vertebral compression fractures. This nuanced understanding of model capabilities must inform future research directions, emphasizing the refinement of AI algorithms for these specific clinical challenges [\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eImportantly, our study has brought to the forefront the substantial heterogeneity present within AI models within this field, echoing the sentiments of other researchers calling for standardization and harmonization of AI methodologies [\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e]. The disparity in model performance reflects a broader issue within the field: the absence of a unified framework or consensus on model development and evaluation criteria. This makes comparisons across studies challenging and impedes the ability to draw definitive conclusions about the best practices and most effective approaches [\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRegarding the clinical utility of AI, there is evidence to suggest that the integration of AI can augment the efficiency of radiological workflows. By potentially reducing the time spent on image interpretation, AI could serve as an adjunct to radiologists, enabling a more rapid turnaround and thereby addressing current diagnostic backlogs. Such a development would be a significant leap forward in healthcare delivery, aligning with recent research demonstrating AI\u0026rsquo;s ability to reduce workload and enhance diagnostic accuracy (Studies demonstrating AI\u0026rsquo;s impact on radiological efficiency). Studies, such as that by Meng F et al. [\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e], directly measure how AI can speed up this reporting process, finding that there was a significant improvement in reporting time when Radiologists are assisted by AI software (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). While Meng F et al\u0026rsquo;s study focusses on the detection of community acquired pneumonia, the principles are universal.\u003c/p\u003e \u003cp\u003eGiven the results of this systematic review and meta-analysis, that AI in this context is provenly accurate and apt for use in clinical practise. However, financial and certification requirements are restricting the uptake. In 2024, Pauling C et al. [\u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e] evaluated several commercially available AI models used to detect fractures, and found variations in pricing strategies for such models from a pay-per-use framework to an annual fee. This study highlighted the scarcity of models that are externally validated for clinical use and commercially available, in the United Kingdom post-Brexit. Pauling C et al emphasized need to develop models that are ready for use and certified by the Medical Devices Directive, the United Kingdom Conformity Assessed marking or similar bodies and certifications. Given the epidemiological burden of vertebral fractures, and the increasing constraints of healthcare systems globally, a cost-efficiency analysis is warranted to assess whether funding for AI technologies in spinal neurosurgery would have a significant positive impact at large.\u003c/p\u003e \u003cp\u003eWe undertook an exhaustive search of the literature, resulting in a study with a very large and high-powered pooled analysis. However, our findings must also be viewed in the context of the limitations of this study. Less than half of the studies included in the meta-analysis provided AUC data in the requisite format for analysis, with metrics such as specificity and sensitivity being more prevalent. The assessment of articles, in line with the PROBAST framework, revealed a general lack of information concerning missing data handling and overall data analysis procedures. Moreover, substantial variance in sample sizes was observed, with some studies having as few as 15 data points available for analysis. Additionally, confidence intervals were not consistently reported across the papers, necessitating our calculation of these intervals. Each study utilized a different AI model, each with its own parameters and methodologies. We aimed to account for the intrinsic weaknesses of the existing literature using a robust analytical approach, nonetheless it necessitates cautious interpretation of the results.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis meta-analysis, included 162 AI models suggests that AI based programmes can accurately diagnose and predict the risk of vertebral fractures, (predictive AUROC\u0026thinsp;=\u0026thinsp;0.82 [0.78\u0026ndash;0.85]; osteoporotic vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.92 [0.88\u0026ndash;0.96]; non-pathological vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.85 [0.81\u0026ndash;0.88] and vertebral compression fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.87 [0.83\u0026ndash;0.91]) at a significant level (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Traditional AI models accounted for the most successful predictive tools and deep learning models contributed to the most successful diagnostic tools. As such future development should be centred around this. However, given the high risk of bias in the papers included, likely including some level of selection and sampling bias, our findings should be interpreted with caution. We recognise the potential benefit of the widespread use of AI both predictively and diagnostically and highlight the need for a well-designed large multicentric study to further explore the benefits of AI in spine surgery, and answer questions on the practicality, efficacy, and cost-efficiency of the AI models in clinical practice.\u003c/p\u003e "},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArtificial Intelligence\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eML\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMachine Learning\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eUS\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eUnited States of America\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAUROC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea under the receiver operating characteristic\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003ePPV\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ePositive predictive value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eNPV\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eNegative predictive value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eFN\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFalse negative\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eFP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eFalse positive\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eTN\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTrue negative\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eTP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eTrue positive\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eAURPC\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eArea Under Precision Recall Curve\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eOVF\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eOsteoporotic vertebral fracture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eVCF\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eVertebral compression fracture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eVF\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eVertebral fracture\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eSDI\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSpinal deformity index\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eCNN\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eConvoluted neural network\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eLSTM\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eLong short-term memory networks\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003e\u003cb\u003eMLP\u003c/b\u003e\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eA multilayer perceptron\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cp\u003eSRN was involved in conceptualisation, data curation, formal analysis, investigation, methodology, project administration, software, supervision, validation, visualisation, writing \u0026ndash; original draft, and writing \u0026ndash; review \u0026amp; editing. SSG, AP, AGK were involved in data curation, formal analysis, investigation, validation and writing \u0026ndash; original draft. \u0026nbsp;DSCR and HSP were involved in conceptualisation, writing \u0026ndash; original draft, and writing \u0026ndash; review \u0026amp; editing. SR, AS, AK, JN, DJ and DK were involved in conceptualisation, writing \u0026ndash; review \u0026amp; editing. SGT was involved in conceptualisation, methodology, formal analysis, investigation, supervision, validation, visualisation, writing \u0026ndash; original draft, and writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\n\u003cp\u003eCompeting Interests\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003eData availability statement\u003c/p\u003e\n\u003cp\u003eAll relevant data supporting the findings of this study can be accessed within the Supplementary Digital Content attached to the article.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAdditional Information\u003c/p\u003e\n\u003cp\u003eThe authors declare that no funds, grants, or other support were received during the preparation of this manuscript. The authors have no relevant financial or non-financial interests to disclose. All data and materials as well as software application support their published claims and comply with field standards. Consent to publish has been received from all participants.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eEnsrud, K. E. Epidemiology of fracture risk with advancing age. J. Gerontol. A Biol. Sci. Med. Sci. 68, 1236\u0026ndash;1242 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDonnally, C. J., DiPompeo, C. M. \u0026amp; Varacallo, M. Vertebral compression fractures. PubMed. Treasure Island (FL): StatPearls Publishing; 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThe European Prospective Osteoporosis Study (EPOS) Group. Incidence of vertebral fracture in Europe: results from the European prospective osteoporosis study (EPOS). J. Bone Miner. Res. 17, 716\u0026ndash;724 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFreitas, S. S. et al. Rate and circumstances of clinical vertebral fractures in older men. Osteoporos. Int. 19, 615\u0026ndash;623 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNevitt, M. C. et al. Risk factors for a first-incident radiographic vertebral fracture in women\u0026thinsp;\u0026ge;\u0026thinsp;65 years of age: the study of osteoporotic fractures. J. Bone Miner. Res. 20, 131\u0026ndash;140 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSavage, J. W., Schroeder, G. D. \u0026amp; Anderson, P. A. Vertebroplasty and kyphoplasty for the treatment of osteoporotic vertebral compression fractures. J. Am. Acad. Orthop. Surg. 22, 653\u0026ndash;664 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCooper, C., Atkinson, E. J., O\u0026rsquo;Fallon, W. M. \u0026amp; Melton, J. L. Incidence of clinically diagnosed vertebral fractures: a population-based study in Rochester, Minnesota, 1985\u0026ndash;1989. J. Bone Miner. Res. 7, 221\u0026ndash;227 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFink, H. A. et al. What proportion of incident radiographic vertebral deformities is clinically diagnosed and vice versa? J. Bone Miner. Res. 20, 1216\u0026ndash;1222 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEnsrud, K. E. et al. Prevalent vertebral deformities predict mortality and hospitalization in older women with low bone mass. J. Am. Geriatr. Soc. 48, 241\u0026ndash;249 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoss, P. D. Clinical consequences of vertebral fractures. Am. J. Med. 103, 30S-42S; discussion 42S-43S (1997).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThomas, B. Artificial intelligence: review of current and future applications in medicine. Fed. Pract. 38, (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavenport, T. \u0026amp; Kalakota, R. The potential for artificial intelligence in healthcare. Future Healthc. J. 6, 94\u0026ndash;98 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAl-Antari, M. A. Artificial intelligence for medical diagnostics\u0026mdash;existing and future AI technology! Diagnostics. 13, 688 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBajwa, J., Munir, U., Nori, A. \u0026amp; Williams, B. Artificial intelligence in healthcare: transforming the practice of medicine. Future Healthc. J. 8, e188-e194 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHardy, M. \u0026amp; Harvey, H. Artificial intelligence in diagnostic imaging: impact on the radiography profession. Br. J. Radiol. 93, 20190840 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKurmis, A. P. \u0026amp; Ianunzio, J. R. Artificial intelligence in orthopedic surgery: evolution, current state and future directions. Arthroplasty. 4, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe, A., Sarda, A., Gupta, S. \u0026amp; Das, S. Use of artificial intelligence in dermatology. Indian J. Dermatol. 65, 352 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTama, B. A., Kim, D. H., Kim, G., Kim, S. W. \u0026amp; Lee, S. Recent advances in the application of artificial intelligence in otorhinolaryngology-head and neck surgery. Clin. Exp. Otorhinolaryngol. 13, 326\u0026ndash;339 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShen, L. et al. Using artificial intelligence to diagnose osteoporotic vertebral fractures on plain radiographs. J. Bone Miner. Res. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHong, N. et al. Deep-learning-based detection of vertebral fracture and osteoporosis using lateral spine X-ray radiography. J. Bone Miner. Res. 38, 887\u0026ndash;895 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDel Lama, R. S., Candido, R. M., Chiari-Correia, N. S., Nogueira-Barbosa, M. H. \u0026amp; de Azevedo-Marques, P. M. Computer-aided diagnosis of vertebral compression fractures using convolutional neural networks and radiomics. J. Digit. Imaging. 35, 446\u0026ndash;458 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolff, R. F. et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann. Intern. Med. 170, 51\u0026ndash;58 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins, G. S., Reitsma, J. B., Altman, D. G. \u0026amp; Moons, K. G. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 350, g7594 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeus, P. et al. Uniformity in measuring adherence to reporting guidelines: the example of TRIPOD for assessing completeness of reporting of prediction model studies. BMJ Open. 9, e025611 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, Y., Sun, X., Sui, X., Li, Y. \u0026amp; Wang, Z. Application of bone alkaline phosphatase and 25-oxhydryl-vitamin D in diagnosis and prediction of osteoporotic vertebral compression fractures. J. Orthop. Surg. Res. 18, 739 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoon, M. A. et al. Automated segmentation of the fractured vertebrae on CT and its applicability in a radiomics model to predict fracture malignancy. Sci. Rep. 12, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa, Y., Lu, Q., Yuan, F. \u0026amp; Chen, H. Comparison of the effectiveness of different machine learning algorithms in predicting new fractures after PKP for osteoporotic vertebral compression fractures. J. Orthop. Surg. Res. 18, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu, X. et al. Prediction of subsequent osteoporotic vertebral compression fracture on CT radiography via deep learning. View (Beijing, China) 3, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKong, S. H. et al. Development of a spine X-ray-based fracture prediction model using a deep learning algorithm. Endocrinol. Metab. 37, 674\u0026ndash;683 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYilmaz, E. B. et al. Assessing attribution maps for explaining CNN-based vertebral fracture classifiers. Lect. Notes Comput. Sci. 3, 3\u0026ndash;12 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYilmaz, E. B. et al. Automated deep learning-based detection of osteoporotic fractures in CT images. Lect. Notes Comput. Sci. 376, 376\u0026ndash;385 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMonchka, B. A., Kimelman, D., Lix, L. M. \u0026amp; Leslie, W. D. Feasibility of a generalized convolutional neural network for automated identification of vertebral compression fractures: the Manitoba Bone Mineral Density Registry. Bone. 150, 116017 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMonchka, B. A. et al. Development of a manufacturer-independent convolutional neural network for the automated identification of vertebral compression fractures in vertebral fracture assessment images using active learning. Bone. 161, 116427 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCho, S. T. et al. Prediction of progressive collapse in osteoporotic vertebral fractures using conventional statistics and machine learning. Spine. 48, 1535 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGui, C. et al. Radiomic modeling to predict risk of vertebral compression fracture after stereotactic body radiation therapy for spinal metastases. J. Neurosurg. Spine. 36, 294\u0026ndash;302 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeol, Y. et al. Predicting vertebral compression fracture prior to spinal SBRT using radiomics from planning CT. Eur. Spine J. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurata, K. et al. Artificial intelligence for the detection of vertebral fractures on plain spinal radiography. Sci. Rep. 10, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, H. Y. et al. Application of deep learning algorithm to detect and visualize vertebral fractures on plain frontal radiographs. PLOS ONE. 16, e0245992 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiamonte, E. et al. Artificial intelligence-based radiomics on computed tomography of lumbar spine in subjects with fragility vertebral fractures. J. Endocrinol. Invest. 45, 2007\u0026ndash;2017 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTomita, N., Cheung, Y. Y. \u0026amp; Hassanpour, S. Deep neural networks for automatic detection of osteoporotic vertebral fractures on CT scans. Comput. Biol. Med. 98, 8\u0026ndash;15 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, Y. C. et al. Can a deep-learning model for the automated detection of vertebral fractures approach the performance level of human subspecialists? Clin. Orthop. Relat. Res. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, J. et al. Automated detection and classification of acute vertebral body fractures using a convolutional neural network on computed tomography. J. Bone Miner. Res. 14, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi, W. G. et al. The value of radiomics-based CT combined with machine learning in the diagnosis of occult vertebral fractures. BMC Musculoskelet. Disord. 24, 819 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEller-Vainicher, C. et al. Recognition of morphometric vertebral fractures by artificial neural networks: analysis from GISMO Lombardia database. PLoS ONE. 6, e27277 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMuehlematter, U. J. et al. Vertebral body insufficiency fractures: detection of vertebrae at risk on standard CT images using texture analysis and machine learning. Eur. Radiol. 29, 2207\u0026ndash;2217 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicolaes, J. et al. Towards improved identification of vertebral fractures in routine computed tomography (CT) scans: development and external validation of a machine learning algorithm. J. Bone Miner. Res. 38, 1856\u0026ndash;1866 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicolaes, J. et al. External validation of a convolutional neural network algorithm for opportunistically detecting vertebral fractures in routine CT scans. Osteoporos. Int. 35, 143\u0026ndash;152 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang, X. et al. Value of 18F-FDG-PET/CT radiomics combined with clinical variables in the differential diagnosis of malignant and benign vertebral compression fractures. EJNMMI Res. 13, 89 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicolaes, J. et al. Detection of vertebral fractures in CT using 3D convolutional neural networks. Lect. Notes Comput. Sci. 3, 3\u0026ndash;14 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNicolaes, J. et al. External validation of a convolutional neural network algorithm for opportunistically detecting vertebral fractures in routine CT scans. Osteoporos. Int. 35, 143\u0026ndash;152 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYabu, A. et al. Using artificial intelligence to diagnose fresh osteoporotic vertebral fractures on magnetic resonance images. Spine J. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoda, T. et al. Automated differentiation between osteoporotic vertebral fracture and malignant vertebral fracture on MRI using a deep convolutional neural network. Spine. 47, E347-E352 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOno, Y. et al. A deep learning-based model for classifying osteoporotic lumbar vertebral fractures on radiographs: a retrospective model development and validation study. J. Imaging. 9, 187 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, W. et al. A deep-learning model for identifying fresh vertebral compression fractures on digital radiography. Bone. 32, 1496\u0026ndash;1505 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim, D. H. et al. Automated vertebral segmentation and measurement of vertebral compression ratio based on deep learning in X-ray images. J. Digit. Imaging. (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiari-Correia, N. S. et al. A 3D radiomics-based artificial neural network model for benign versus malignant vertebral compression fracture classification in MRI. J. Digit. Imaging. 36, 1565\u0026ndash;1577 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, B. et al. Benign vs malignant vertebral compression fractures with MRI: a comparison between automatic deep learning network and radiologist\u0026rsquo;s assessment. Eur. Radiol. 33, 5060\u0026ndash;5068 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuan, S. et al. Differential diagnosis of benign and malignant vertebral compression fractures: comparison and correlation of radiomics and deep learning frameworks based on spinal CT and clinical characteristics. Eur. J. Radiol. 165, 110899 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDong, Q. et al. Generalizability of deep learning classification of spinal osteoporotic compression fractures on radiographs using an adaptation of the modified-2 algorithm-based qualitative criteria. Acad. Radiol. 30, 2973\u0026ndash;2987 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThawait, S. K. et al. Comparison of four prediction models to discriminate benign from malignant vertebral compression fractures according to MRI feature analysis. AJR Am. J. Roentgenol. 200, 493\u0026ndash;502 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H. et al. Differentiation of benign versus malignant indistinguishable vertebral compression fractures by different machine learning with MRI-based radiomic features. Eur. Radiol. 33, 5069\u0026ndash;5076 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRyu, S. M. et al. Diagnosis of osteoporotic vertebral compression fractures and fracture level detection using multitask learning with U-Net in lumbar spine lateral radiographs. Comput. Struct. Biotechnol. J. (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFeng, S. et al. Two-stream compare and contrast network for vertebral compression fracture diagnosis. IEEE Trans. Med. Imaging. 40, 2496\u0026ndash;2506 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRinaldi, C. et al. The early detection of osteoporosis in a cohort of healthcare workers: is there room for a screening program? J. Clin. Endocrinol. Metab. 106, e485-e495 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS\u0026ouml;zen, T., \u0026Ouml;zışık, L. \u0026amp; Başaran, N. \u0026Ccedil;. An overview and management of osteoporosis. Eur. J. Rheumatol. 4, 46\u0026ndash;56 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShams, R. A., Zowghi, D. \u0026amp; Bano, M. AI and the quest for diversity and inclusion: a systematic literature review. AI Ethics. 4, 73\u0026ndash;88 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYin, J., Ngiam, K. Y. \u0026amp; Teo, H. H. Role of artificial intelligence applications in real-life clinical practice: systematic review. J. Med. Internet Res. 23, e25743 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMittermaier, M., Raza, M. M. \u0026amp; Kvedar, J. C. Bias in AI-based models for medical applications: challenges and mitigation strategies. NPJ Digit. Med. 6, 27 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarwaha, J. S. \u0026amp; Kvedar, J. C. Crossing the chasm from model performance to clinical impact: the need to improve implementation and evaluation of AI. NPJ Digit. Med. 5, 25 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCimpeanu, T. et al. Artificial intelligence development races in heterogeneous settings. Sci. Rep. 12, 5729 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeng, F. et al. AI support for accurate and fast radiological diagnosis of COVID-19: an international multicenter, multivendor CT study. Eur. Radiol. 33, 4280\u0026ndash;4291 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePauling, C. et al. Commercially available artificial intelligence tools for fracture detection: the evidence. BJR Open. 6, tzd005 (2023).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, machine learning, osteoporotic vertebral fractures, non-pathological vertebral fractures, vertebral compression fractures, ML, AI, OF, VF","lastPublishedDoi":"10.21203/rs.3.rs-4643800/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4643800/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eIntroduction:\u003c/h2\u003e \u003cp\u003eWith the increasing prevalence of vertebral fractures, accurate diagnosis and prognostication are essential. This study assesses the effectiveness of AI in diagnosing and predicting vertebral fractures through a systematic review and meta-analysis.\u003c/p\u003e\u003ch2\u003eMethod\u003c/h2\u003e \u003cp\u003eA comprehensive search across major databases selected studies utilizing AI for vertebral fracture diagnosis or prognosis. Out of 14,161 studies initially identified, 79 were included, with 40 undergoing meta-analysis. Diagnostic models were stratified by pathology: non-pathological vertebral fractures, osteoporotic vertebral fractures, and vertebral compression fractures. The primary outcome measure was AUROC.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eAI showed high accuracy in diagnosing and predicting vertebral fractures: predictive AUROC\u0026thinsp;=\u0026thinsp;0.82, osteoporotic vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.92, non-pathological vertebral fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.85, and vertebral compression fracture diagnosis AUROC\u0026thinsp;=\u0026thinsp;0.87, all significant (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Traditional models had the highest median AUROC (0.90) for fracture prediction, while deep learning models excelled in diagnosing all fracture types. High heterogeneity (I\u0026sup2; \u0026gt; 99%, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) indicated significant variation in model design and performance.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eAI technologies show considerable promise in improving the diagnosis and prognostication of vertebral fractures, with high accuracy. However, observed heterogeneity and study biases necessitate further research. Future efforts should focus on standardizing AI models and validating them across diverse datasets to ensure clinical utility.\u003c/p\u003e","manuscriptTitle":"Artificial Intelligence and Machine Learning for Risk Prediction and Diagnosis of Vertebral Fractures: A Systematic Review and Meta-Analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-07-29 14:22:42","doi":"10.21203/rs.3.rs-4643800/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-08-08T12:15:42+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-28T13:35:27+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-23T15:34:57+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-20T12:17:25+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"74113427937751500533335831379486246841","date":"2024-07-17T12:54:40+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"332404639261065026213605625572026082205","date":"2024-07-15T09:35:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"206156668056528193151481402846359435205","date":"2024-07-14T16:05:04+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-07-07T11:56:37+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-07-07T11:54:58+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-07-07T10:28:16+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-07-04T10:00:02+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-06-26T15:37:20+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"25917dc3-d82d-4ca3-aab9-ea167c7f483d","owner":[],"postedDate":"July 29th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":35022597,"name":"Health sciences/Rheumatology/Musculoskeletal system/Bone"},{"id":35022598,"name":"Health sciences/Medical research/Outcomes research"}],"tags":[],"updatedAt":"2024-12-23T15:58:57+00:00","versionOfRecord":{"articleIdentity":"rs-4643800","link":"https://doi.org/10.1038/s41598-024-75628-2","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2024-12-19 15:56:53","publishedOnDateReadable":"December 19th, 2024"},"versionCreatedAt":"2024-07-29 14:22:42","video":"","vorDoi":"10.1038/s41598-024-75628-2","vorDoiUrl":"https://doi.org/10.1038/s41598-024-75628-2","workflowStages":[]},"version":"v1","identity":"rs-4643800","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4643800","identity":"rs-4643800","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}