Consensus statement on the application of artificial intelligence in osteoporosis screening and management: perspectives from the Asia-Pacific region Authors: Chun-Feng Huang, Wen-Hui Fang, Kun-Hui Chen, Sung-Yen Lin, Cheng-Jung Ho, Jawl-Shan Hwang, Ta-Wei Tai, Yuan-Fu Liu, Chien-An Shih, Jung-Fu Chen, Shih-Te Tu, Ding-Cheng Chan, Rong-Sen Yang, Shau-Huai Fu, Hsuan-Yu Chen, Keh-Sung Tsai, Tien-Tsai Cheng, Fang-Ping Chen, Wei-Chieh Hung, Yin-Fan Chang, Der-Sheng Han, Manju Chandran, Ang Seng Bin, Joon Kiong Lee, Swan Sim Yeap, Yoon-Sok Chung, Kwang-Kyoun Kim, Peter R. Ebeling, Unnop Jaisamrarn, Dipendra Pandey, Serge Ferrari, Tsung-Han Yang, Natthinee Charatcharoenwitthaya, Akira Taguchi, Sarath Lekamwasam, Tuan Van Nguyen, E. Michael Lewiecki, Kenneth G. Saag, Ching-Chou Tsai, Fernando Marín Published: 2026 ABSTRACT SummaryOsteoporosis is a major and growing health concern in the Asia-Pacific region, y et it remains widely underdiagnosed and undertreated due to limited access to dual-energy X-ray absorptiometry (DXA) in many areas. Artificial intelligence (AI) offers new opportunities to improve osteoporosis screening and management, but unvalidated tools pose risks of inconsistent care. This consensus was developed to provide regionally harmonized guidance on the safe, effective, and equitable use of AI in osteoporosis care.PurposeThe aim of this work was to establish expert consensus recommendations on the role of AI in osteoporosis screening and management in the Asia-Pacific region. Key objectives were to define appropriate applications of AI (e.g., imaging-based bone assessment and fracture risk prediction) and specify minimum standards for validation and reporting, addressing region-specific implementation challenges and ensuring that AI use aligns with clinical guidelines and ethical principles.MethodsThis consensus was developed through multidisciplinary collaboration among experts across the Asia-Pacific region. Each participant reviewed draft statements, contributed feedback during virtual meetings, and provided insights based on clinical experience and current evidence. Consensus was reached iteratively until full agreement was achieved for all statements. The process integrated global best practices and regional adaptations, drawing from peer-reviewed studies, international AI guidelines, and local fracture registry data. The final recommendations emphasize the validation, transparency, and ethical implementation of AI within regional healthcare systems, ensuring compatibility with local regulations. Ultimately, twelve consensus statements were established to guide the responsible use of AI for osteoporosis screening and management in the Asia-Pacific region.ResultsThe panel produced 12 consensus statements covering the role of AI as an adjunct for opportunistic osteoporosis screening rather than a diagnostic tool, requirements for imaging quality and AI model transparency, standards for validation and performance reporting, integration of AI with clinical risk stratification, demonstration of clinical utility in real-world settings, adherence to data protection laws and ethical AI principles, training of clinicians in AI use, strategies for implementation and monitoring (including post-market surveillance and feedback loops), and recognition of technical, clinical, and equity limitations of AI. All 12 statements give extensive recommendations for using AI to improve osteoporosis management while ensuring patient safety, accuracy, and equity.ConclusionThis first Asia-Pacific consensus on AI in osteoporosis concludes that AI, when appropriately validated and implemented, can help bridge the osteoporosis care gap by identifying high-risk patients who would otherwise remain undiagnosed, thus facilitating earlier intervention. It emphasizes that AI should complement—not replace—standard diagnostic methods and clinical judgment. The guidance emphasizes validation, transparency, and ethical oversight to facilitate early intervention while minimizing risks associated with unvalidated or premature AI adoption. Introduction Osteoporosis is a “silent” disease characterized by low bone mass and structural deterioration, leading to increased fracture risk [1–3]. The Asia-Pacific region faces a particularly urgent challenge: as populations age, more than 50% of the world’s osteoporotic fractures are projected to occur in Asia by 2050 [4, 5]. Despite this growing burden, osteoporosis in the Asia-Pacific region remains greatly underdiagnosed and undertreated [6]. A key reason is the limited availability of dual-energy X-ray absorptiometry (DXA)—the gold standard for osteoporosis diagnosis—especially in developing countries and rural areas [7]. Many populous Asian countries have fewer than one DXA unit per million inhabitants, far below European levels, leaving a substantial proportion of individuals with low bone mineral density (BMD) undiagnosed until a fragility fracture occurs [7–9]. Given these gaps in care, there is intense interest in leveraging artificial intelligence (AI) to improve osteoporosis screening and fracture prevention [10]. AI—particularly machine learning and deep learning algorithms—can analyze medical images and patient data to detect patterns of bone loss or fracture risk that may be missed in routine practice [11]. Notably, AI models have shown the ability to estimate BMD or osteoporosis risk using existing clinical records and radiological images, including routinely acquired X-ray–based modalities such as chest radiographs and dental panoramic radiographs, as well as computed tomography (CT) scans [11–14]. This “opportunistic screening” approach repurposes scans obtained for other reasons (e.g., chest imaging for lung disease) to also screen for low bone density, without additional radiation or cost [14–17]. Studies demonstrate a wide range of accuracy depending on the imaging modality and target population: AI models applied to foot and ankle radiographs have reported AUC values of 0.85–0.94 with sensitivities of 75–90% and specificities of 80–93% [16]; models based on chest radiographs have reported AUC 0.81–0.90 [11]; CT-based models show AUC 0.87–0.96 for BMD estimation [14]. An AI system evaluating chest radiographs was found to increase early osteoporosis identification in women over 50 and was cost-effective, falling below standard willingness-to-pay thresholds [17]. Such results highlight AI’s potential to expand osteoporosis screening beyond the limited DXA facilities, reaching patients in primary care or underserved settings who would otherwise not be assessed. However, the proliferation of AI tools also brings significant concerns. Many algorithms have been developed in isolation, using narrow datasets that may not represent the diverse populations of the Asia-Pacific region. Without validation across multiple centers and ethnic groups, AI models risk yielding inaccurate or biased results when deployed broadly [18]. Indeed, a recent review found that of dozens of published AI studies for osteoporosis, only 2 had been externally validated on new patient cohorts [19]. Algorithm performance often degrades when faced with different imaging devices or patient demographics [20]. In the absence of clear guidelines, clinicians may be unsure how to integrate AI into practice—for example, how to act on an AI-generated osteoporosis risk score, or how to communicate AI findings to patients. Unregulated use of unvalidated AI could lead to misdiagnosis, inappropriate treatment, or false reassurance that leaves high-risk patients unmanaged [21]. Recognizing both the opportunities and risks of AI in osteoporosis, a panel of experts from the Asia-Pacific region convened to develop a consensus on best practices. The goal of this consensus is to provide harmonized guidance that can be adopted across countries in the Asia-Pacific region, ensuring that AI tools are used in a safe, effective, and ethically responsible manner. Importantly, the scope of this consensus is focused on applications of AI in osteoporosis screening and initial management—that is, identifying individuals at risk of osteoporosis or fracture and aiding clinical decision-making in their care. This includes AI for (1) imaging analysis (e.g., BMD estimation from X-rays or CT, vertebral fracture detection); (2) risk assessment that combines AI outputs with clinical risk factors; (3) DXA scan workflow support (such as automated analysis or quality control); (4) treatment decision support (for example, AI-guided recommendations based on risk thresholds); and (5) novel modalities such as wearable technologies for monitoring bone health. Applications beyond osteoporosis screening (such as AI in surgical treatment or rare bone disorders) are excluded from our scope. Within osteoporosis care, the emphasis is on augmenting current practice—not replacing established diagnostic criteria or human clinical judgment. For example, an AI algorithm might flag a patient as “high risk” based on a chest X-ray, but the definitive diagnosis should still follow international guidelines (DXA T-score, presence of fragility fractures, etc.) and a clinician’s evaluation [18]. Therefore, we seek to broaden this consensus by tackling critical challenges in AI implementation, including technical constraints, data privacy issues, the necessity for clinician education, and the demand for stringent regulatory oversight, while delineating vital future directions. This consensus offers a relevant, region-specific framework to assist policymakers, healthcare professionals, and technology developers in utilizing AI to enhance osteoporosis outcomes while minimizing potential dangers. Moreover, by following these recommendations, the Asia-Pacific region can progress toward uniform norms for AI implementation in osteoporosis management, fostering responsible innovation that provides significant advantages for patients in various healthcare environments. Methods This consensus was developed through extensive collaboration among multidisciplinary experts from across the Asia-Pacific region, including representatives from East Asia, Southeast Asia, South Asia, and Oceania (Taiwan, Japan, Korea, Mainland China, Hong Kong, Singapore, Malaysia, Thailand, Vietnam, the Philippines, Nepal, Sri Lanka, and Australia). In addition, internationally recognized opinion leaders from Europe and North America were invited as external experts to provide global perspectives and methodological rigor. Panel members were identified through a two-stage process. First, co-chairs (C.H.C. and C.H.W.) compiled an initial candidate list based on established expertise in osteoporosis, bone imaging, AI/machine learning, or healthcare policy, as evidenced by authorship in key journals (Osteoporosis International, Journal of Bone and Mineral Research) and participation in prior regional osteoporosis guidelines. Second, this list was reviewed by the Taiwanese Osteoporosis Association (TOA) and the Asian Federation of Osteoporosis Societies (AFOS), which nominated additional representatives to ensure geographic and professional balance. Geographic balance was intentionally sought, with at least one representative invited from each of the following subregions: East Asia, Southeast Asia, South Asia, and Oceania. All panellists were required to complete a standard conflict of interest (COI) disclosure form consistent with ICMJE guidelines prior to participation; no panellist with direct financial ties to an AI osteoporosis product company participated in statement formulation. Expert feedback was synthesized to produce comprehensive recommendations tailored to the region’s specific needs, challenges, and healthcare structures. Each expert contributed by reviewing preliminary draft statements, attending structured consensus meetings, and providing comments grounded in clinical experience, regulatory familiarity, and the latest evidence in AI and osteoporosis research. The primary author (C.F.H.) conducted a structured literature search in PubMed and Embase using predefined search terms covering AI, osteoporosis, bone mineral density, fracture risk, opportunistic screening, and Asia-Pacific, covering publications from January 2015 to September 2025. TRIPOD-AI, CONSORT-AI, and relevant regulatory guidance documents (FDA, EMA, PMDA, and WHO) were included as methodological reference frameworks. A summary evidence document was circulated to all panellists four weeks prior to the in-person meeting. The development process integrated both the best global practices and regional contextual adaptations. Evidence was drawn from peer-reviewed publications, clinical guidelines, technical standards (TRIPOD-AI: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis–Artificial Intelligence extension, CONSORT-AI: CONsolidated Standards Of Reporting Trials–Artificial Intelligence extension), and regional fracture and DXA registry data. The scope of this consensus is explicitly limited to AI applications in osteoporosis screening and management involving ionizing-radiation–based imaging modalities (radiographs, CT, and DXA). Nonionizing modalities, including quantitative ultrasound (QUS) and radiofrequency echographic multispectrometry (REMS), fall outside the current scope; the published AI validation literature for these modalities is insufficient at present to support evidence-based consensus statements. Future consensus work addressing AI-REMS and AI-QUS applications is encouraged once a sufficient evidence base accumulates. To establish the consensus recommendations, multidisciplinary experts in osteoporosis from across the Asia-Pacific region were formally invited to participate in a structured consensus process. An initial one-month consultation period was conducted, during which all panelists reviewed the background evidence and contributed to the formulation of provisional draft statements through iterative written feedback. Consensus was formally defined as ≥ 80% agreement among all panellists. During the pre-meeting consultation phase, anonymous electronic voting was conducted via a standardized form distributed by email; statements not reaching this threshold were flagged for discussion. All co-authors constituted the expert panel. The consensus methodology was designed to align with the ACCORD (Advocating for Consensus and Cooperative Research for Orderly Development) reporting framework. This process was followed by an in-person consensus meeting held in Taipei, Taiwan, on October 18, 2025. The meeting facilitated in-depth discussion, refinement, and final agreement on each statement. Through iterative review and collective deliberation, the panel ensured that all recommendations reflected both scientific rigor and practical feasibility across diverse healthcare systems in the Asia-Pacific region. The final consensus recommendations were designed to promote responsible innovation and equitable implementation of AI in osteoporosis screening and management. Emphasis was placed on transparency, validation, and ethical deployment of AI tools in real-world settings, while ensuring compatibility with local regulations such as the Personal Data Protection Act (PDPA). This structured, consultative process resulted in twelve final consensus statements that collectively provide harmonized guidance for the safe and effective integration of AI technologies into osteoporosis care throughout the region. Results Twelve consensus statements were formulated, addressing key aspects of AI use in osteoporosis care for the Asia-Pacific region. Each statement is presented below in exact wording as finalized by the consensus panel. Table 1 provides a summary of all 12 statements for quick reference. Table 1Asia-Pacific consensus statements on AI in osteoporosisNo.Consensus statement1. Rationale and scopeOsteoporosis remains underdiagnosed across the Asia-Pacific region due to limited DXA availability. The proliferation of unvalidated AI tools creates risks of inconsistent standards. This consensus establishes regionally harmonized guidance for safe, effective, and equitable deployment of AI-based osteoporosis screening tools. This consensus covers AI for: (1) imaging analysis (BMD estimation, fracture detection from radiographs/CT), (2) risk assessment integrating clinical factors, (3) DXA workflow support, (4) treatment decision support, and (5) wearable technologies. Applications beyond osteoporosis screening are excluded2. Role of AIBased on current evidence, AI is most appropriately positioned as an opportunistic screening tool for osteoporosis risk assessment, not a standalone diagnostic instrument. Diagnosis must adhere to guidelines based on DXA T-scores, fragility fractures, or validated bone health indicators. AI applications span imaging analysis, risk stratification, DXA assistance, treatment support, and wearables—but all require clinical oversight and regional adaptation3. Image sources and qualityAI-based bone health assessment should primarily repurpose preexisting images (chest, abdomen, spine, hip, and jaw radiographs) to avoid additional radiation. However, in specific contexts like emergency fracture detection, dedicated imaging may be appropriate. Accuracy depends on standardized protocols and adequate image quality4. Model transparency and explainabilityAI reports must specify imaging device brands/models used for training and reference datasets (e.g., NHANES III). Systems should incorporate Explainable AI (XAI) principles with visual explanations, confidence scores, and plain-language outputs. Clear definitions must distinguish accuracy (systematic error), precision (random error), and reliability (reproducibility)5. Validation and reportingIndependent multicenter external validation with representative regional datasets is mandatory. Reports must include AUC, AUPRC, sensitivity, specificity, and calibration metrics—not accuracy alone. Studies should follow TRIPOD-AI, CONSORT-AI standards and align with FDA/EMA/PMDA and WHO/IOF frameworks. Performance varies with population prevalence; temporal and geographic generalizability must be assessed6. Risk stratificationAI should support risk stratification (high/moderate/low) by providing evidence-based categories that, combined with clinical factors (age, fractures, and comorbidities), guide referral pathways. Risk depends on multiple factors beyond AI outputs; AI provides contributory information, not standalone determinations. Thresholds must be evidence-based and linked to clear actions like DXA confirmation7. Integration with risk calculatorsAI-estimated BMD may augment fracture risk calculators (e.g., FRAX®) when DXA is unavailable, only if: validated against DXA in target populations, calibrated to local fracture data, and using compatible reference databases. AI-BMD is a surrogate with larger error than DXA and should not replace DXA when available or delay high-risk referrals8. Clinical utility and real-world evidenceAI utility must be confirmed by prospective studies and real-world evidence assessing fracture incidence, diagnosis rates, treatment initiation, and cost-effectiveness. Regional fracture epidemiology (varying incidence and refracture rates) and healthcare disparities must guide local optimization. Health economic evaluations should use local data and account for diverse funding models9. Data protection and Ethics AI must comply with local laws (PDPA) and international standards (GDPR and HIPAA). Systems should align with WHO AI Ethics Guidance and OECD AI Principles emphasizing transparency, accountability, and equity. Address Asia-Pacific data sovereignty and cross-border restrictions. Establish feedback mechanisms for clinicians to report discordant AI-DXA cases to vendors for retraining and continuous improvement10. Training and competencyStructured training should focus on practical application: interpreting outputs, recognizing limitations, integrating results clinically, and maintaining professional judgment. Deep algorithmic understanding is unnecessary—“bus drivers don’t need to know how gearboxes work, only how to shift gears.” Emphasize AI as an adjunct, not replacement, and train clinicians to communicate results as risk alerts, not diagnoses11. Implementation, monitoring, and communicationPrioritize deployment in primary care and underserved regions with seamless PACS/HIS integration and multilingual accessibility. Implement mandatory postmarket surveillance to detect performance drift. Establish clinician feedback loops to report discordant cases for vendor model refinement. Evaluate cost-effectiveness locally and review guidelines every 3–5 years. Multistakeholder collaboration (clinicians, regulators, industry, and patients) is essential12. Limitations of AIAI has technical (generalizability limits, image quality dependency, and blackbox interpretation), clinical (screening not diagnostic, variable subgroup performance, and cannot replace comprehensive assessment), and equity limitations (digital divide, algorithmic bias, and uncertain cost-effectiveness). Deployment must consider resource availability and equity; transparent acknowledgment of limitations and ongoing validation are vital for safe adoptionAI, artificial intelligence; BMD, bone mineral density; CONSORT-AI, Consolidated Standards of Reporting Trials–Artificial Intelligence extension; CT, computed tomography; DXA, dual-energy X-ray absorptiometry; FDA, US Food and Drug Administration; EMA, European Medicines Agency; FRAX, Fracture Risk Assessment Tool; GDPR, General Data Protection Regulation; HIPAA, Health Insurance Portability and Accountability Act; IOF, International Osteoporosis Foundation; NHANES III, The Third National Health and Nutrition Examination Survey; OECD, Organisation for Economic Co-operation and Development; PACS, Picture Archiving and Communication System; HIS, Hospital Information System; PDPA, Personal Data Protection Act; TRIPOD-AI, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis–Artificial Intelligence extension; PMDA, Pharmaceuticals and Medical Devices Agency; WHO, World Health Organization Table 1Asia-Pacific consensus statements on AI in osteoporosisNo.Consensus statement1. Rationale and scopeOsteoporosis remains underdiagnosed across the Asia-Pacific region due to limited DXA availability. The proliferation of unvalidated AI tools creates risks of inconsistent standards. This consensus establishes regionally harmonized guidance for safe, effective, and equitable deployment of AI-based osteoporosis screening tools. This consensus covers AI for: (1) imaging analysis (BMD estimation, fracture detection from radiographs/CT), (2) risk assessment integrating clinical factors, (3) DXA workflow support, (4) treatment decision support, and (5) wearable technologies. Applications beyond osteoporosis screening are excluded2. Role of AIBased on current evidence, AI is most appropriately positioned as an opportunistic screening tool for osteoporosis risk assessment, not a standalone diagnostic instrument. Diagnosis must adhere to guidelines based on DXA T-scores, fragility fractures, or validated bone health indicators. AI applications span imaging analysis, risk stratification, DXA assistance, treatment support, and wearables—but all require clinical oversight and regional adaptation3. Image sources and qualityAI-based bone health assessment should primarily repurpose preexisting images (chest, abdomen, spine, hip, and jaw radiographs) to avoid additional radiation. However, in specific contexts like emergency fracture detection, dedicated imaging may be appropriate. Accuracy depends on standardized protocols and adequate image quality4. Model transparency and explainabilityAI reports must specify imaging device brands/models used for training and reference datasets (e.g., NHANES III). Systems should incorporate Explainable AI (XAI) principles with visual explanations, confidence scores, and plain-language outputs. Clear definitions must distinguish accuracy (systematic error), precision (random error), and reliability (reproducibility)5. Validation and reportingIndependent multicenter external validation with representative regional datasets is mandatory. Reports must include AUC, AUPRC, sensitivity, specificity, and calibration metrics—not accuracy alone. Studies should follow TRIPOD-AI, CONSORT-AI standards and align with FDA/EMA/PMDA and WHO/IOF frameworks. Performance varies with population prevalence; temporal and geographic generalizability must be assessed6. Risk stratificationAI should support risk stratification (high/moderate/low) by providing evidence-based categories that, combined with clinical factors (age, fractures, and comorbidities), guide referral pathways. Risk depends on multiple factors beyond AI outputs; AI provides contributory information, not standalone determinations. Thresholds must be evidence-based and linked to clear actions like DXA confirmation7. Integration with risk calculatorsAI-estimated BMD may augment fracture risk calculators (e.g., FRAX®) when DXA is unavailable, only if: validated against DXA in target populations, calibrated to local fracture data, and using compatible reference databases. AI-BMD is a surrogate with larger error than DXA and should not replace DXA when available or delay high-risk referrals8. Clinical utility and real-world evidenceAI utility must be confirmed by prospective studies and real-world evidence assessing fracture incidence, diagnosis rates, treatment initiation, and cost-effectiveness. Regional fracture epidemiology (varying incidence and refracture rates) and healthcare disparities must guide local optimization. Health economic evaluations should use local data and account for diverse funding models9. Data protection and Ethics AI must comply with local laws (PDPA) and international standards (GDPR and HIPAA). Systems should align with WHO AI Ethics Guidance and OECD AI Principles emphasizing transparency, accountability, and equity. Address Asia-Pacific data sovereignty and cross-border restrictions. Establish feedback mechanisms for clinicians to report discordant AI-DXA cases to vendors for retraining and continuous improvement10. Training and competencyStructured training should focus on practical application: interpreting outputs, recognizing limitations, integrating results clinically, and maintaining professional judgment. Deep algorithmic understanding is unnecessary—“bus drivers don’t need to know how gearboxes work, only how to shift gears.” Emphasize AI as an adjunct, not replacement, and train clinicians to communicate results as risk alerts, not diagnoses11. Implementation, monitoring, and communicationPrioritize deployment in primary care and underserved regions with seamless PACS/HIS integration and multilingual accessibility. Implement mandatory postmarket surveillance to detect performance drift. Establish clinician feedback loops to report discordant cases for vendor model refinement. Evaluate cost-effectiveness locally and review guidelines every 3–5 years. Multistakeholder collaboration (clinicians, regulators, industry, and patients) is essential12. Limitations of AIAI has technical (generalizability limits, image quality dependency, and blackbox interpretation), clinical (screening not diagnostic, variable subgroup performance, and cannot replace comprehensive assessment), and equity limitations (digital divide, algorithmic bias, and uncertain cost-effectiveness). Deployment must consider resource availability and equity; transparent acknowledgment of limitations and ongoing validation are vital for safe adoptionAI, artificial intelligence; BMD, bone mineral density; CONSORT-AI, Consolidated Standards of Reporting Trials–Artificial Intelligence extension; CT, computed tomography; DXA, dual-energy X-ray absorptiometry; FDA, US Food and Drug Administration; EMA, European Medicines Agency; FRAX, Fracture Risk Assessment Tool; GDPR, General Data Protection Regulation; HIPAA, Health Insurance Portability and Accountability Act; IOF, International Osteoporosis Foundation; NHANES III, The Third National Health and Nutrition Examination Survey; OECD, Organisation for Economic Co-operation and Development; PACS, Picture Archiving and Communication System; HIS, Hospital Information System; PDPA, Personal Data Protection Act; TRIPOD-AI, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis–Artificial Intelligence extension; PMDA, Pharmaceuticals and Medical Devices Agency; WHO, World Health Organization Asia-Pacific consensus statements on AI in osteoporosis AI, artificial intelligence; BMD, bone mineral density; CONSORT-AI, Consolidated Standards of Reporting Trials–Artificial Intelligence extension; CT, computed tomography; DXA, dual-energy X-ray absorptiometry; FDA, US Food and Drug Administration; EMA, European Medicines Agency; FRAX, Fracture Risk Assessment Tool; GDPR, General Data Protection Regulation; HIPAA, Health Insurance Portability and Accountability Act; IOF, International Osteoporosis Foundation; NHANES III, The Third National Health and Nutrition Examination Survey; OECD, Organisation for Economic Co-operation and Development; PACS, Picture Archiving and Communication System; HIS, Hospital Information System; PDPA, Personal Data Protection Act; TRIPOD-AI, Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis–Artificial Intelligence extension; PMDA, Pharmaceuticals and Medical Devices Agency; WHO, World Health Organization The consensus statements presented above reflect the final agreement of the expert panel. Collectively, they underscore that AI should be applied to enhance osteoporosis screening and care within a framework of careful clinical and regulatory oversight. Each statement is addressed in the subsequent sections, where it is discussed in context and integrated into the broader narrative rather than presented as a point-by-point enumeration. To facilitate practical usage, the aforementioned 12 statements outline a comprehensive framework—from the overarching role of AI in screening to technical requirements, integration into clinical pathways, and acknowledging limitations graphically (Fig. 1).Fig. 1Conceptual framework and consensus domains for the safe and effective use of artificial intelligence (AI) in opportunistic osteoporosis screening in the Asia-Pacific region. Regulatory science considerations (e.g., FDA qualification of DXA-derived total hip BMD as a surrogate endpoint) may further inform future AI validation targets and trial designs Conceptual framework and consensus domains for the safe and effective use of artificial intelligence (AI) in opportunistic osteoporosis screening in the Asia-Pacific region. Regulatory science considerations (e.g., FDA qualification of DXA-derived total hip BMD as a surrogate endpoint) may further inform future AI validation targets and trial designs Discussion Implications and potential of AI in osteoporosis care The consensus statements underscore that AI has a valuable but supporting role in osteoporosis care, particularly as a tool for opportunistic screening in the Asia-Pacific context [22]. Given the underdiagnosis of osteoporosis in the region due to various reasons such as limited DXA and human resources, AI can help identify at-risk individuals from data that are already available in the healthcare system in a time-effective manner. Statement 1 highlights the rationale: many Asia-Pacific patients who could benefit from osteoporosis assessment are missed under current practice, and unregulated AI use could lead to inconsistent standards [23, 24]. By establishing a unified guidance, the consensus aims to channel AI development and deployment toward filling the diagnostic gap safely. It is important to note that the incremental value of AI opportunistic screening varies substantially across the Asia-Pacific region depending on DXA availability. In settings with low DXA density (much of South and Southeast Asia), AI screening offers clear incremental value by identifying high-risk individuals who would otherwise remain undiagnosed. In contrast, in high-DXA-density settings such as Japan and South Korea—where DXA accessibility and cost are comparable to opportunistic imaging—the marginal benefit is less straightforward. Routine deployment of AI opportunistic screening tools in such contexts may carry a meaningful risk of increasing false-positive referrals without a commensurate improvement in diagnostic yield, and health technology assessments should carefully evaluate whether AI screening is additive or merely duplicative of existing DXA pathways. One of the most significant implications is the use of AI to analyze existing medical images for signs of low bone density or fractures (Statement 3). This approach can dramatically expand the screening reach. For example, an individual receiving a chest X-ray for a cough could simultaneously be screened for osteoporosis risk by an AI algorithm analyzing the thoracic spine or rib bone patterns on that X-ray [11]. Such opportunistic screening is particularly relevant in the Asia-Pacific region, where dedicated osteoporosis screenings are often not performed [22]. AI enables “missed” patients to be evaluated using scans performed for other purposes. Studies from within and outside the region support this strategy: a deep learning model was able to identify patients with osteopenia or osteoporosis with high accuracy [25]. Likewise, algorithms processing CT scans have shown a strong correlation with DXA measurements of BMD [26–28]. These findings demonstrate that AI can estimate or correlate with DXA-derived BMD values using more ubiquitous imaging modalities, providing a surrogate assessment in settings where DXA is unavailable. The consensus embraces this potential, suggesting that preexisting images (chest, abdomen, hip X-rays, etc.) be the primary data source for AI-based bone health assessments (Statement 3) to avoid unnecessary radiation exposure and leverage whatever imaging is available in resource-limited settings. Moreover, AI can assist in fracture risk stratification by integrating imaging findings with clinical risk factors (Statements 6 and 7). This aligns with current clinical tools such as FRAX® (Fracture Risk Assessment Tool), which calculates 10-year fracture probability based on risk factors and BMD [29, 30]. The panel recommends that AI-derived BMD or other outputs may augment such risk calculators only under certain conditions; notably, the AI should be well-validated against DXA in the target population and calibrated to local fracture epidemiology (Statement 7). This caution is warranted because while AI models might estimate BMD, these estimates have an error margin larger than that of DXA and could misclassify some patients [30]. Nonetheless, in settings where DXA is unavailable, a validated AI tool could provide a provisional risk assessment to guide management. For instance, an AI that analyzes a hip X-ray to predict BMD could be used to feed a FRAX calculation, as long as it is shown to be accurate for that ethnic population and fracture data. This approach can potentially improve risk stratification in primary care: high-risk patients (e.g., an elderly woman flagged by AI as likely osteoporotic) can be referred for confirmatory DXA or started on preventive therapy sooner, while low-risk patients are reassured or advised on lifestyle changes [30]. One study demonstrated that a deep learning–based model (DeepSurv) achieved fracture-risk prediction performance comparable to that of the conventional FRAX tool in diabetic patients [31]. These findings indicate that AI has the potential to improve fracture-risk estimation in clinical subgroups where conventional tools have known limitations. Regulatory science is also evolving in a direction that may broaden the translational pathway for BMD-centric innovations, including AI-enabled bone health assessment. On December 19, 2025, the US Food and Drug Administration (FDA) qualified total hip BMD (assessed by DXA) as a validated surrogate endpoint for phase 3 clinical trials of investigational therapies for postmenopausal women with osteoporosis at risk for fracture, defined as the percentage change from baseline at 24 months in total hip BMD [32]. This milestone signals increasing regulatory acceptance of BMD as an efficacy outcome (under specific conditions), which may create clearer opportunities to evaluate AI tools not only against DXA correlation metrics but also within BMD-driven clinical pathways. Nevertheless, it is critical to distinguish FDA-qualified DXA-derived total hip BMD from AI-estimated BMD: AI outputs should be treated as screening surrogates unless and until robust bridging evidence demonstrates equivalence, calibration, and clinical utility in the intended population (including device- and workflow-specific performance). By following Statement 2’s guidance that based on current evidence, AI is most appropriately positioned as an aid for risk assessment rather than a standalone diagnostic instrument, healthcare providers can incorporate AI outputs into their decision-making without over-relying on them [33]. The consensus clearly positions DXA and clinical evaluation as the gold standards for diagnosis—consistent with international definitions of osteoporosis and national guidelines. AI’s role is to alert and inform. This delineation is crucial for maintaining clinical prudence: while AI might flag a patient as high risk, the physician still must confirm osteoporosis via DXA or recognize that AI’s suggestion of “osteoporosis” is not definitive. This approach is akin to using AI as a highly sensitive screening test that needs a specific confirmatory test [34]. It also helps manage patient expectations—if patients are made aware that an AI screening indicated they might have low bone density, they can be counseled that a formal DXA is needed to diagnose and that AI is not a “virtual DXA” [18]. In short, the consensus integrates AI into the existing care pathway, rather than replacing elements of it. This integration ensures that improvements in early detection (through AI) do not come at the cost of diagnostic accuracy or adherence to evidence-based treatment thresholds [35–37]. The consensus statements also highlight some forward-looking applications that can transform osteoporosis care. For example, wearable technologies (mentioned in the scope of Statement 1) could continuously monitor risk factors such as physical activity, gait stability, or even bone relevant biomarkers, feeding data to AI algorithms. Although still emerging, such approaches could enable “continuous fracture risk monitoring” in the community [38, 39]. It should be noted that the application of wearable technologies to bone health monitoring remains at an early investigational stage (Technology Readiness Level 3–4); current evidence is largely proof-of-concept, and wearables are included in the consensus scope as a forward-looking direction requiring dedicated validation before clinical deployment. One can envision AI-driven assessments from smartwatches or sensor insoles that gauge frailty or fall risk, prompting timely interventions [40]. The panel’s inclusion of wearables acknowledges this potential. AI may analyze mobility patterns from a wearable device to predict if a patient’s fracture risk is increasing (due to declining gait speed or balance), triggering a clinic visit [41]. Another novel application area is AI-driven treatment decision support (Statement 1 scope). This may involve algorithms that support clinicians in determining the most appropriate therapeutic strategies (e.g., calcium/vitamin D supplementation and exercise) through the synthesis of patient-specific risk factors, individual preferences, and evidence-based guidelines [42]. While few systems currently do this, the consensus anticipates a future where AI might suggest, for instance, that a high-risk patient with certain comorbidities might best benefit from a specific osteoporosis medication, or that a patient with moderate risk could be monitored rather than medicated—all under physician oversight [43]. Importantly, the consensus calls for these innovations to be adapted to regional needs. “Regional adaptation” (Statement 2) implies that AI models and thresholds might need adjustment for different Asia-Pacific populations [44]. The epidemiology of osteoporosis varies widely across Asia. For instance, hip fracture rates are very high in parts of East Asia but lower in some other areas [4]. An AI predicting absolute fracture risk must account for those baseline differences—perhaps via calibration with local fracture incidence data (as Statement 7 suggests). Likewise, risk factors such as body size, diet, or genetics differ among Asian subpopulations, meaning that an AI trained predominantly on Western data might underperform if applied without modification. The consensus therefore encourages the development of localized reference standards (e.g., using local BMD reference databases and incorporating regional risk factors such as high prevalence of vitamin D deficiency and population-specific lifestyle factors). By doing so, AI tools will yield more accurate predictions for the populations they serve than using an AI tool without including population-specific risk factors and/or calibration. It should also be noted that the current consensus statements were developed primarily with reference to postmenopausal women, who represent the highest burden group and for whom the majority of AI validation data exist. AI model performance in men, individuals with glucocorticoid-induced osteoporosis (GIO), and those with secondary causes of bone loss is less well established, and validation data in these groups remain sparse. Direct application of current AI tools to these populations warrants caution pending dedicated validation studies. Future prospective studies should explicitly include these groups. The immediate benefit of implementing this consensus will likely be improved early identification of osteoporosis. If health systems in the Asia-Pacific region adopt AI screening (for example, automatically running a validated AI on chest X-rays or CT scans of older patients), many more high-risk individuals can be funneled into appropriate care pathways. Early detection is critical—as noted, osteoporotic fractures (especially hip fractures) carry significant morbidity and mortality, with a one-year mortality rate of approximately 15% in some studies [45]. Treating osteoporosis with pharmacotherapy can substantially reduce fracture risk, so catching patients before the first fracture is a public health priority. The consensus-driven approach ensures that such AI-based screening would be performed in a standardized and clinically responsible manner across the region. AI-driven opportunistic screening via chest X-rays improved quality-adjusted life years and was cost-effective in an older population [17]. While real-world confirmation is still needed (addressed in Statement 8), AI-driven screening has the potential to improve early detection and address unmet diagnostic needs in the Asia-Pacific region. Challenges and considerations in implementation While AI presents exciting opportunities, the consensus panel is emphatic that its implementation must overcome several challenges and adhere to strict standards to truly benefit patients [46]. One major theme in the statements is the need for rigorous validation and transparent reporting of AI performance (Statement 5). This arises from the recognition that many AI models perform impressively in initial studies but may not generalize well. The literature shows numerous examples where AI algorithms achieved high accuracy in a development dataset but failed to replicate those results externally. For instance, a systematic review by Yu et al. (2022) found that in 27 of 31 radiology AI studies (87%), external validation performance was significantly lower than internal performance, with median AUC decreasing by 0.06–0.12 across modalities [47]. In the osteoporosis AI literature specifically, Gatineau et al. (2024) documented that of dozens of published AI studies for osteoporosis, only 2 had undergone formal external validation [18]. These examples underscore the risk of over-optimistic internal performance metrics. In osteoporosis, this risk is heightened by heterogeneity in imaging techniques and patient demographics across the Asia-Pacific region. The consensus therefore mandates independent multicenter external validation before any AI tool is widely deployed (Statement 5). This means that if an AI is trained on scans from Thailand, it should be evaluated on data from other countries (such as India, Japan, and Malaysia) to guarantee that it continues to perform correctly. External validation is essential because, as noted, reduced algorithm performance on external datasets is common when models encounter shifts in population or imaging settings [47]. By requiring representative regional datasets, the panel is essentially saying an AI must prove itself on the diversity of real-world cases in Asia-Pacific—different ethnicities, ages, and clinical profiles—and not just in a controlled or homogeneous sample. The panel also stresses comprehensive performance reporting: metrics such as area under the ROC curve (AUC), area under the precision–recall curve (AUPRC), sensitivity, specificity, and calibration should all be reported (Statement 5). This addresses a prevalent issue where studies might tout a single metric (often “accuracy”) which can be misleading. For example, in a class-imbalanced scenario, accuracy alone can be high (for example, if few people have osteoporosis, a fundamental model stating “no osteoporosis” for everyone may be 90% correct because 90% do not have it, but it fails to discover any true cases). Reporting AUC and calibration gives a fuller picture of how well the model distinguishes risk and whether its probability outputs are reliable [48]. The consensus echoes emerging AI reporting guidelines such as TRIPOD-AI for prediction models and CONSORT-AI for clinical trials involving AI, urging investigators and developers to follow these standards [49, 50]. By aligning with international frameworks FDA, European Medicines Agency (EMA), Pharmaceuticals and Medical Devices Agency (PMDA) regulatory guidance and World Health Organization (WHO)/International Osteoporosis Foundation (IOF) recommendations), the statements push for a high quality of evidence. In practice, this means that any AI tool should ideally undergo the kind of evaluation we expect of medical devices or new diagnostic tests: preferably a prospective trial (if used in patient care decisions) or at least a thorough validation on unseen data according to a predefined protocol [51]. Another challenge is ensuring model transparency and explainability, as elaborated in Statement 4. “Black-box” algorithms can undermine clinician and patient trust, and make it difficult to identify errors or biases. The consensus advises that AI systems incorporate explainable AI (XAI) features such as visual heatmaps showing which areas of an X-ray contribute to a low BMD prediction, or provide a confidence score with the result [52, 53]. This is in line with broader AI ethical principles that call for transparency and accountability in AI decisions. For instance, the WHO’s six principles for AI in healthcare include transparency, explainability, and intelligibility of AI systems [54]. In practical terms, if an AI report on a patient’s radiograph says “High risk of osteoporosis,” it should also include additional info like “because cortical bone thickness appears reduced at these points (highlighted on image)” or “model confidence: 85%.” Such features help the clinician understand and trust the result, or potentially question it if it contradicts clinical exam. The consensus also calls for clear definitions of terms such as accuracy, precision, and reliability to avoid confusion (Statement 4). This may seem technical, but it matters when AI results are communicated. Clinicians should know, for example, the difference between the AI’s precision (consistency) and accuracy (closeness to true value)—an AI might be very consistent but consistently off by a margin in BMD estimation, which has different implications than random error. The panel’s desire for clarity ensures that end users of AI (doctors and radiographers) are not deceived by overly optimistic statements. Data quality is another crucial consideration. Statement 3 notes that AI accuracy depends on standardized imaging protocols and adequate quality. If images are noisy, low-resolution, or taken with varied techniques, AI predictions can be erroneous. Many healthcare facilities in low-resource settings may have older X-ray machines or less standardized acquisition, which could degrade AI performance. Therefore, implementing AI might require investments in quality control—e.g., ensuring X-ray machines are calibrated and images meet certain resolution criteria, or the AI might need a quality-check module to reject images that are not suitable (for instance, an algorithm might flag “image too low quality for analysis”) [55]. Additionally, patient position and artifacts can impact measurements such as femoral geometry on X-rays, so training radiographers on proper techniques remains important even when AI is involved [56]. The agreement basically warns that AI is not magical—“garbage in, garbage out” remains true. Efforts to deploy AI should go hand-in-hand with improving image quality and consistency across sites. A significant challenge in the Asia-Pacific deployment of AI is ensuring equity and avoiding bias (Statement 12’s equity limitations, and Statement 9’s emphasis on fairness). The “digital divide” exists: urban tertiary hospitals may have advanced information technology infrastructure to run AI tools, whereas rural clinics may lack even reliable internet or picture archiving systems (e.g., PACS (Picture Archiving and Communication System)) [57]. If AI is only implemented in wealthy centers, it could widen health disparities, benefiting patients who already have more access. The consensus insists on prioritizing deployment in primary care and underserved regions (Statement 11)—which implies making the technology accessible and user-friendly in those settings. For example, an AI tool might need to be embedded in a lightweight application that can run offline or on portable devices for rural outreach. In regions without advanced information technology infrastructure, AI tools can be run in a cloud-based server. Additionally, algorithmic bias must be addressed: if an AI was trained mostly on, say, postmenopausal women from one ethnic group, its performance on men or on other ethnicities might be poor, potentially leading to underestimation or overestimation of risk in those groups. There is evidence that using Caucasian BMD reference standards can misclassify Asian patients; similarly, an AI might inherit such biases if not retrained on local data [58]. The panel’s recommendations to use local reference databases and to continuously collect feedback (Statements 7 and 9) serve as safeguards. We suggest a feedback loop where if clinicians notice systematic errors (e.g., AI missing osteoporosis in very petite individuals or in men), these cases should be reported back to developers so the model can be improved. This is a form of active bias mitigation—acknowledging that the first version of any AI might not be perfect, but with real-world testing and iterative updates (retraining with more diverse data), it can become fairer and more robust. Data governance and privacy pose another set of challenges. Statement 9 addresses compliance with local and international data protection laws. Asia-Pacific countries enforce various regulations (for example, Singapore’s PDPA and Europe’s GDPR (General Data Protection Regulation)) which influence many global practices, such as the HIPAA (Health Insurance Portability and Accountability Act) for any collaboration with US entities) [59]. AI development often requires large datasets, and sometimes cross-border data sharing (to obtain enough cases of certain populations). However, many nations are increasingly concerned about data sovereignty—patient data may not be allowed to leave the country or even the hospital network. This complicates multicenter collaborations [60]. The consensus suggests acknowledging and addressing these restrictions explicitly. Solutions might include federated learning approaches (where AI models are trained on-site in each location and only model weights are shared, not raw data) or establishing secure regional data hubs with proper agreements [61]. Additionally, compliance means having robust patient consent processes and de-identification for any data used in AI training [60, 62]. A related point is that AI algorithms, once deployed, should also protect patient privacy—e.g., if an AI platform is cloud-based, it must ensure secure transmission and storage of medical images (encryption, etc.) [62]. The consensus aligning with WHO and OECD (Organisation for Economic Co-operation and Development) AI principles also implies the need for accountability: if an AI makes a significant error that harms a patient, there should be clarity on responsibility (the clinician? the hospital? the AI vendor?) [63]. Currently, regulatory agencies are grappling with this. The US FDA has begun treating some AI algorithms as medical devices requiring approval, and similar regulatory pathways are developing in Asia [64]. The consensus essentially prepares the stakeholders to consider these issues: only use AI that meets regulatory approval where required and maintain monitoring postimplementation. Regulatory frameworks for AI in healthcare across the Asia-Pacific region The regulatory classification of AI as a medical device (Software as Medical Device (SaMD)) varies considerably across the Asia-Pacific region, and understanding this landscape is essential for responsible AI deployment. In Japan, the Pharmaceuticals and Medical Devices Agency (PMDA) governs AI-based medical devices under the Pharmaceutical and Medical Device Act, with a dedicated SaMD pathway for AI/ML-based diagnostic tools requiring premarket review. In South Korea, the Ministry of Food and Drug Safety (MFDS) has published dedicated AI medical device guidelines since 2019, classifying AI diagnostic tools by risk level and requiring clinical performance evaluation. In Australia, the Therapeutic Goods Administration (TGA) regulates SaMD under an updated framework (2023), aligned with International Medical Device Regulators Forum (IMDRF) guidance. In Singapore, the Health Sciences Authority (HSA) has issued a regulatory framework for software-based medical devices. In Taiwan, the Taiwan Food and Drug Administration (TFDA) has issued AI medical device guidance documents requiring performance and explainability reporting. In China, the National Medical Products Administration (NMPA) has issued AI medical device classification guidelines, with a risk-tiered approval pathway. Across the region, these frameworks broadly align with IMDRF SaMD guidelines and the IMDRF AI/ML SaMD action plan, though the pace of implementation differs. A key distinction from the US FDA’s approach is that several Asian jurisdictions currently require a full pre-market submission for AI diagnostic tools rather than the FDA’s pre-determined change control plan for AI/ML devices, which may affect the speed of iterative model updates post-deployment. A significant regulatory gap exists in lower-middle-income Asia-Pacific countries where formal AI medical device classification frameworks are nascent or absent, reinforcing the tiered compliance framework recommended in Statement 9 and the need for minimum baseline standards applicable across all resource settings. The statements on training and communication (Statement 10) highlight that human factors are as important as technical factors. Many clinicians and radiologists in the region may not have had formal training in AI concepts. A survey by the American Medical Association found that while physician use of AI is increasing, many still have concerns and knowledge gaps about AI in practice [65]. The consensus recommends structured training focusing on practical competencies—how to interpret AI outputs and integrate them into patient care. This could mean developing short courses or workshops for healthcare professionals on topics such as: understanding AI confidence levels, recognizing scenarios where AI might be less reliable (e.g., unusual patient anatomy or imaging artefacts), and explaining AI findings to patients. The delightful metaphor provided (“bus drivers don’t need to know how gearboxes work, only how to shift gears”) captures an important point: clinicians do not need to be data scientists, but they do need to know when and how to use the AI “gearbox.” For example, a primary care doctor receiving an AI-based report should know that a “red flag” from the AI is a call to action (e.g., refer for DXA or specialist consult), while a “green flag” (low risk) still does not guarantee that the patient is fine if other clinical risk factors suggest otherwise. Training should also emphasize that AI is an adjunct—a second pair of eyes—not an oracle. This mindset will help clinicians maintain their own clinical judgment, thereby reducing over-reliance on AI and preventing de-skilling [66]. From an implementation standpoint, the consensus advocates for the integration of AI into existing clinical workflows and information systems (Statement 11). This is a nontrivial challenge: many healthcare IT systems in the Asia-Pacific region are not fully interoperable, and AI tools often start as standalone software. If using AI requires multiple extra steps (e.g., exporting an image to a separate application, then importing results back), busy clinicians may ignore it. Thus, seamless integration with PACS and HIS (Hospital Information System) is crucial [67, 68]. Multilingual accessibility is another important point the panel made: Asia-Pacific’s linguistic diversity means AI interfaces and reports should be available (or translatable) in local languages, not just English. Otherwise, critical information might be lost on the end-user [44]. Monitoring and feedback loops (Statement 11) are highlighted to ensure that AI remains safe over time. Unlike a static medical device, AI software can “drift”—for example, if an algorithm is deployed in 2019 and in 2025, the patient population’s characteristics have shifted (perhaps due to new treatment patterns, post-COVID health changes, etc.), the AI might start performing worse. Therefore, postmarket surveillance akin to pharmacovigilance is recommended [69]. This could involve periodic audits of AI performance metrics, user reports of any errors, and even a system for end-users to flag cases where the AI was wrong (as mentioned, clinicians reporting discordant AI vs. DXA results to the vendor; Statement 9). The consensus effectively preempts this by advising it as best practice. One outcome of monitoring could be the decision to retrain the AI model after a certain time with new data. The feedback from clinicians is invaluable in this process—it grounds AI development in reality. For example, if multiple doctors report that the AI tends to miss fracture in patients with osteoarthritis (perhaps because osteophytes confuse it), the developers know what specific issue to address in the next version. This collaborative improvement cycle is how AI can become safer and more accurate over time [70]. Cost and economic considerations are also a challenge. Statements 8 and 11 mention evaluating cost-effectiveness and performing local health economic analyses. Introducing AI has upfront costs—software licensing, hardware (perhaps GPUs for processing), integration expenses, training staff, etc. Healthcare budgets in many Asia-Pacific countries are constrained [71]. It is generally agreed that cost-effectiveness must be demonstrated through the use of local data. For instance, an AI screening program implemented in a country with a very low fracture incidence may be less cost-effective than one in a country with a high incidence, unless the cost of AI remains substantially low. Conversely, in regions where current screening practices are nearly absent, even moderately effective AI could substantially impact fracture prevention and reduce costs by preventing fracture treatments. To make Statement 8’s requirement for clinical utility demonstration more operational, meaningful benefit should be assessed against a framework of evaluable endpoints: (1) incremental change in osteoporosis diagnosis rate (new DXA-confirmed diagnoses per 1000 patients screened versus usual care); (2) reduction in time-to-first-treatment among newly diagnosed patients (months from AI screening flag to pharmacotherapy initiation); (3) proportion of fragility fractures preceded by an AI-generated risk alert (retrospective audit metric); and (4) incremental cost-effectiveness ratio (ICER) expressed as cost per quality-adjusted life year (QALY) gained relative to local willingness-to-pay thresholds, which vary across the region (approximately 1–3 × GDP per capita in most Asia-Pacific countries). These endpoints are proposed as minimum reporting requirements for real-world implementation studies. The consensus recommends reviewing the outcomes and economics every 3–5 years (Statement 11). This means that if AI technology becomes cheaper or more effective, or if epidemiology changes, the guidelines on who to screen and how might be updated. The 3–5 year review interval aligns with the typical guideline update cycle used by major clinical bodies (e.g., International Osteoporosis Foundation, NICE) and reflects the pace at which meaningful new evidence from prospective studies is likely to accumulate. In addition to scheduled reviews, event-driven reviews should be triggered by: (a) documented AI performance drift exceeding predefined thresholds; (b) significant changes in fracture epidemiology or DXA availability at the regional level; or (c) major regulatory changes in AI medical device classification across Asia-Pacific jurisdictions. Finally, multistakeholder collaboration is emphasized (Statement 11) as an overarching need to tackle these challenges. Healthcare providers can work with tech companies to co-design AI tools that fit clinical workflows and address real-world pain points. Patients and public health officials can help guide AI deployments in a way that maximizes reach to those in need (for instance, focusing on community health centers or screening programs). This multistakeholder approach is vital to overcome inertia and address concerns from all angles—from legal liability to ethical use to technical support [72]. The consensus statements naturally point toward several future directions in both research and policy to fully realize AI’s benefits for osteoporosis in the Asia-Pacific region. First, prospective clinical trials and real-world implementation studies are urgently needed (as highlighted in Statement 8). Another future direction is the use of federated learning and collaborative networks across the Asia-Pacific region to train AI models without breaching data privacy (tying into Statement 9). For example, hospitals in various nations could work together to develop a model by training on local data and communicating only model parameters (not patient data) with a central aggregator. In this way, an “Asia-Pacific osteoporosis AI” could be developed that learns from the heterogeneity of the region while respecting data sovereignty. Early examples of such federated approaches in medicine are appearing and given the emphasis on cross-border restrictions in the consensus, this could be a key strategy moving forward. It will require coordination and trust between institutions, and possibly facilitation by international bodies (perhaps IOF or the Asia Pacific Consortium on Osteoporosis (APCO) could spearhead such an initiative for bone health AI, such as how they do audits and frameworks). Furthermore, in Asia-Pacific region, where medical curricula vary greatly, there may be initiatives (possibly spearheaded by major organizations or regional associations) to build modules on AI in osteoporosis/bone health. Trainees may learn how to understand an AI-generated bone report or how to critically evaluate an AI tool’s validation paper. The consensus’ attitude on training (Statement 10) supports this trend, and as a result, healthcare firms should explore upskilling their personnel concurrently with AI adoption. Despite its promise, AI in osteoporosis care remains constrained by several important limitations (Statement 12) that must temper expectations. Technically, AI performance depends heavily on image quality and dataset representativeness; algorithms trained on narrow datasets may fail when confronted with new imaging devices, ethnicities, or disease patterns [20]. Clinically, AI is a screening adjunct, not a diagnostic replacement—false negatives could delay treatment, while false positives could increase unnecessary testing. Furthermore, algorithmic bias may exacerbate inequities if underserved populations are underrepresented in training data, reinforcing the “digital divide” across the Asia-Pacific region [73]. Economically, cost-effectiveness remains uncertain, as many AI tools have yet to prove their real-world impact on fracture reduction [31, 74]. Therefore, continuous post-market validation, transparency about model uncertainty, and equitable access must be prioritized to ensure that AI enhances rather than undermines clinical care. This awareness of AI limitations forms the ethical cornerstone of safe adoption and underlines the consensus call for ongoing evaluation and refinement. Conclusion By synthesizing current evidence and expert opinion, the 12 consensus statements provide clear guidance on how to deploy AI safely, effectively, and equitably: AI should serve as an adjunct to clinical care—identifying at-risk individuals from existing data, enhancing risk stratification, and supporting decision-making—while final diagnosis and management remain grounded in established clinical standards and physician judgment. The consensus places strong emphasis on the prerequisites for AI use, including robust validation, transparency, and oversight, thereby guarding against the potential pitfalls of unproven technology. This highlights that AI should supplement, not replace, traditional diagnostic procedures and professional judgment. This AI consensus will help Asia-Pacific healthcare professionals and systems enhance early osteoporosis detection and secondary fracture prevention in an evidence-based and patient-centered manner. The expert panel encourages stakeholders to integrate these consensus statements into their policies and procedures, as well as to engage in ongoing evaluation and reform. With a coordinated effort, the Asia-Pacific region has the potential to set a new benchmark for how artificial intelligence may be ethically integrated into clinical practice to combat the osteoporosis pandemic and improve the bone health of millions.