AI-Driven Appendicular Skeletal Muscle Mass Index (ASMI) Prediction and Low Muscle Mass Detection from Routine Hip X-rays: A Novel Opportunistic Screening Tool

AI-Driven Appendicular Skeletal Muscle Mass Index (ASMI) Prediction and Low Muscle Mass Detection from Routine Hip X-rays: A Novel Opportunistic Screening Tool | 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 Research Article AI-Driven Appendicular Skeletal Muscle Mass Index (ASMI) Prediction and Low Muscle Mass Detection from Routine Hip X-rays: A Novel Opportunistic Screening Tool Ling Lee, Shu-Han Chuang, Yi-Jie Kuo, Lien-Chen Wu, Yu-Pin Chen This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8603410/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Purpose Sarcopenia diagnosis requires identifying low muscle mass (LMM), typically via dual-energy X-ray absorptiometry (DXA). However, DXA's limited accessibility restricts large-scale screening. This retrospective study aimed to develop and validate a deep learning model to predict DXA-derived ASMI from routine hip radiographs for opportunistic sarcopenia screening. Methods We included 1,267 patients with both hip radiography and DXA scans, split into development (n = 1,140) and external validation (n = 127) sets. A multimodal model integrating radiographic images (ResNet-34 backbone) and clinical variables (age, gender, height, weight, BMI) was trained to predict continuous ASMI and classify LMM per Asian Working Group for Sarcopenia (AWGS) 2019 criteria. Results On external validation, the model achieved strong performance with Pearson r = 0.806, R²=0.650, MAE = 0.414 kg/m², and AUC = 0.874 for LMM classification. Applying AWGS diagnostic thresholds yielded sensitivity of 70.5% and specificity of 83.3%, with consistent performance across gender and age subgroups. Gradient-weighted Class Activation Mapping confirmed focus on clinically relevant gluteal and proximal thigh muscles. Conclusions This deep learning approach enables automated LMM identification from routine hip radiographs, offering a cost-effective, accessible tool for opportunistic sarcopenia screening and early intervention in at-risk populations. Artificial intelligence Sarcopenia Low muscle mass Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Aim To develop and validate a deep learning model that predicts quantitative Appendicular Skeletal Muscle Mass Index (ASMI) and identifies low muscle mass using routine hip radiographs. Findings The deep learning model accurately predicted ASMI (Pearson r = 0.806) and effectively classified low muscle mass (AUC = 0.874), with interpretability confirmed by focus on gluteal and thigh muscles. Message This AI model enables opportunistic sarcopenia screening from standard hip X-rays without additional radiation or cost, potentially facilitating early intervention in at-risk populations. Introduction Sarcopenia, characterized by the progressive and generalized loss of skeletal muscle mass, strength, and function, constitutes a significant public health challenge, particularly within the context of a rapidly aging global population [ 1 ]. Beyond normal aging, sarcopenia strongly predicts numerous adverse outcomes, including an increased risk of falls, fragility fractures, profound disability, loss of physical independence, and premature mortality [ 2 ]. Sarcopenia affects 5–13% of adults aged 60–70 and increasing to as high as 50% in individuals over 80. This imposes a considerable economic burden on healthcare systems worldwide, with associated costs in the United States alone estimated in the tens of billions of dollars annually [ 3 , 4 ]. These severe clinical and economic problems underscore the urgent need for effective and scalable strategies for the early detection of sarcopenia, which would enable timely interventions to mitigate its debilitating effects. Sarcopenia is diagnosed in accordance with the criteria of the Working Group on Sarcopenia in Older People and is characterized by both low muscle mass (LMM) and low muscle strength [ 5 , 6 ]. Muscle strength is easily accessed via measurements such as handgrip strength or walking speed, which are simple and widely used in clinical practice for sarcopenia screening. However, muscle mass assessment requires specialized equipment, representing a major barrier to large-scale screening. Bioelectrical impedance analysis (BIA) and dual-energy X-ray absorptiometry (DXA) are the primary modalities for assessing muscle mass, while imaging techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) provide highly accurate but less practical alternatives. Despite their precision, CT and MRI are unsuitable for population-level screening due to high cost, limited accessibility, the need for specialized personnel and equipment, and radiation exposure [ 7 , 8 ]. Although BIA is noninvasive and free from radiation, it suffers from considerable measurement variability and limited availability, reducing its reliability for sarcopenia screening [ 9 ]. In contrast, DXA remains the reference standard for assessing muscle mass and diagnosing sarcopenia [ 10 ], but its high cost and limited accessibility continue to restrict its widespread application in routine practice, particularly for screening purposes. Opportunistic screening involves leveraging routinely acquired medical images to assess for findings beyond the primary clinical indication, such as quantifying muscle and bone metrics [ 11 ]. Concurrently, the field of deep learning, particularly convolutional neural networks (CNNs), has advanced rapidly in medical imaging. CNNs learn hierarchical image features and have demonstrated expert-level performance in tasks such as lesion detection, classification, and segmentation [ 12 ]. Recent studies have exploited these advances to quantify muscle mass from routine images. For example, Ryu et al. (2023) trained a CNN using chest X-rays to predict appendicular lean mass (ALM) and sarcopenia status, reporting strong agreement with DXA measurements (concordance ≈ 0.80) and an Area Under the Receiver Operating Characteristic Curve (AUC) of approximately 0.81 for sarcopenia detection [ 13 ]. Similarly, Hwang et al. (2022) employed a U-Net architecture to segment muscle on full-leg radiographs, achieving a high intersection over union (IoU ≈ 0.93) with manual segmentations and an AUC of 0.988 for sarcopenia screening based on the predicted muscle volume [ 14 ]. Furthermore, Gu et al. (2023) applied deep learning to abdominal CT scans for body composition analysis, achieving high segmentation accuracy (DSC > 0.90) and an AUC of 0.874 for automated sarcopenia classification [ 15 ]. These studies demonstrate the feasibility of using deep learning method analysis of routine imaging to reliably quantify muscle mass or sarcopenia classification. Anteroposterior (AP) hip radiographs are commonly performed in older adults presenting with hip pain or after a fall, given that over 95% of hip fractures in this demographic result from low-energy trauma [ 16 ]. Individuals with sarcopenia exhibit reduced appendicular muscle mass, and the hip and thigh regions contain the largest proportion of skeletal muscle in the body. Moreover, thigh muscle measurements are widely recognized as a standard reference for evaluating muscle wasting [ 17 ]. These factors highlight the potential of applying deep learning algorithms to AP hip radiographs for opportunistic sarcopenia screening. Therefore, we hypothesized that a CNN could be trained to predict DXA-derived appendicular skeletal muscle mass index (ASMI) from routine hip radiographs, providing a practical and accessible alternative for sarcopenia screening [ 18 ]. Therefore, this study addresses the critical need for an accessible sarcopenia screening tool. Our primary objective was to develop and rigorously validate a deep learning model capable of predicting continuous, quantitative DXA-derived ASMI values directly from standard AP hip radiographs and clinical demographic data. The secondary objective was to evaluate the model's diagnostic performance in classifying LMM according to established clinical guidelines, thereby establishing its feasibility as a novel tool for opportunistic sarcopenia screening. Methods Study Design and Data collection This study is reported in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) statement [ 19 ]. We identified all patients who underwent both a clinical DXA scan and a hip radiograph between January 2017 and December 2022. Each patient’s information including age, gender, height, weight, body mass index (BMI), ASMI, total lean mass, total fat, leg lean mass, and leg fat was collected. We used 5 numerical variables including age, gender, height, weight, and BMI as clinical information for boosting learning which was suggested by the surgeon. Patients were excluded from this study if they met any of the following criteria: (1) a time interval greater than 90 days between the two examinations; (2) the presence of missing image data or essential demographic and clinical records; or (3) recorded ASMI which may be the significant outlier and likely a data entry error. After exclusion, a final cohort of 1,267 patients was established. This cohort was then chronologically split, with the initial 90% of patients (n = 1140) forming the development set for model training and internal validation, and the subsequent 10% (n = 127) serving as a temporally distinct external validation set to assess generalizability. To the best of our knowledge, this is the largest dataset for automated sarcopenia diagnosis from images and tabular information to date. Figure 1 shows the flowchart for the generation of each dataset. The demographic and clinical characteristics of this dataset are shown in Table 1 . Table 1 Baseline characteristics in each dataset. Development (n = 1140) External validation (n = 127) p-value Sarcopenia With 479(42%) 61(42%) 0.225 Without 661(58%) 66(58%) 0.225 Demography Gender (male) 257(22.5%) 36(28.3%) 0.172 Age (years) Height (cm) Weight (kg) 72.2 ± 12.1 156.4 ± 8.2 56.8 ± 10.8 71.2 ± 11.6 156.8 ± 8.8 57.6 ± 11.1 0.325 0.652 0.440 BMI (kg/m 2 ) ASMI (kg/m 2 ) 23.2 ± 3.8 5.9 ± 1.0 23.5 ± 3.9 6.0 ± 0.9 0.358 0.883 Abbreviations: ASMI, appendicular skeletal muscle mass index. Study Outcomes and Definitions The ground truth for the primary regression task was the ASMI, calculated by normalizing DXA derived ALM to height squared, expressed in kg/m². ASMI is integral to clinical definitions of sarcopenia, including the criteria established by the Asian Working Group for Sarcopenia (AWGS). According to AWGS guidelines, LMM is defined by DXA-based ASMI cutoffs of < 7.0 kg/m² for men and < 5.4 kg/m² for women.[ 20 ] For the secondary classification task, the model's ability to identify patients with LMM was evaluated. Because the diagnostic thresholds for sarcopenia are gender-specific (ASMI < 7.0 kg/m² for males and < 5.4 kg/m² for females), a unified risk score was calculated for each patient to enable receiver operating characteristic (ROC) analysis across the entire cohort [ 20 ]. This score was defined as the gender-specific threshold minus the model-predicted ASMI value. A ROC curve was then generated using this risk score, and the AUC and further metrics were calculated. Image Preprocessing Radiographs in Digital Imaging and Communications in Medicine (DICOM) format were converted to grayscale images. Pixel intensity was normalized using window/level parameters from the DICOM metadata. To reduce confounding effects from metallic hardware, we implemented an automated implant detection algorithm. Regions with pixel intensities exceeding a threshold of 240 (on a 0-255 scale) were identified as implants and subsequently removed using an inpainting technique. All images were then resized to a standard 224×224-pixel resolution and normalized to a [0, 1] range. During the training phase, an extensive data augmentation strategy was employed to enhance model robustness and prevent overfitting. We implemented a comprehensive, policy-based augmentation scheme [ 21 ]. For each image in a training batch, one of 30 predefined augmentation policies was stochastically selected and applied. These policies comprised a comprehensive set of transformations, including: (1) Geometric transformations, such as random rotation (up to ± 30°), shearing, and translation. Horizontal flipping was specifically included to encourage the model to learn features invariant to bilateral differences in hip anatomy. (2) Photometric transformations, involving adjustments to contrast, brightness, sharpness, and color saturation, alongside operations like equalization, auto-contrast, solarization, and inversion. (3) Occlusion-based augmentation, which applied a cutout technique to randomly mask a square region of the image, compelling the model to learn from partial structural information. No augmentation was applied to the validation or external test sets to ensure a consistent and unbiased evaluation of the model's performance. Deep Learning Model Architecture We developed a multimodal deep learning model to predict ASMI from hip radiographs and associated clinical parameters. The model architecture was substantially informed by the work of Jin et al., with modifications to adapt the framework for a regression task and to incorporate advanced training strategies [ 22 ]. The architecture comprises three main components: (1) an image feature extraction pathway using a ResNet-34 backbone pretrained on ImageNet; (2) a clinical feature encoder (TextNet) for demographic and anthropometric data; and (3) a gated attention fusion module to integrate both modalities. The image pathway processes 224×224-pixel grayscale radiographs through the ResNet-34 backbone. A non-local self-attention block was applied to the final convolutional layer to capture long-range spatial dependencies relevant to muscle mass [ 23 ]. Clinical features (age, gender, height, weight, BMI) were encoded via a 1-dimensional CNN (TextNet) to produce a 64-dimensional feature vector. The gated attention fusion module then combined image features (512 dimensions) with spatially expanded clinical features using a learnable gating mechanism, allowing the model to dynamically weight the contribution of each data type. To enhance model’s focus initially, Class Activation Mapping (CAM) enhancement was incorporated during training, where spatial attention masks were elementwise multiplied with input images using the formula: \(\:{X}_{f}=X\times\:(1+\sigma\:(CAM\left)\right),\) where σ is the sigmoid function. Furthermore, contrastive learning was implemented using a noise-contrastive estimation (InfoNCE) loss function to improve the quality of feature representations, with a temperature parameter τ = 0.117 and a weighting factor β = 0.011. The final fused features were passed through a regression head to output a continuous ASMI prediction. Both CAM enhancement and contrastive learning strategy was also informed by the work of Jin et al [ 22 ]. Model Training Protocol Model Training Protocol The model was trained using a 5-fold cross-validation strategy on the development set. The model was optimized using the Concordance Correlation Coefficient (CCC) loss function, which is well-suited for clinical agreement studies: $$\:\:CCC=\frac{2\rho\:{\sigma\:}_{x}{\sigma\:}_{y}}{{{\sigma\:}_{x}}^{2}+{{\sigma\:}_{y}}^{2}{+({\mu\:}_{x}-{\mu\:}_{y})}^{2}}$$ where ρ is the Pearson correlation coefficient (r), σ is the standard deviation, and µ is the mean of the predicted and actual ASMI values [ 24 ]. Optimization was performed using the AdamW algorithm with an initial learning rate of 7.5×10⁻⁵, weight decay of 7.5×10⁻⁶, and gradient clipping at a norm of 1.0 [ 25 ]. The learning rate was dynamically adjusted during training; it was reduced by a factor of 0.33 whenever the validation loss plateaued for 8 consecutive epochs. Training was conducted for a maximum of 80 epochs with an early stopping criterion of 20 epochs. Batch sizes were set to 32 for training and 16 for validation and testing. The hyperparameters were identified as optimal through a systematic search conducted with the Optuna hyperparameter optimization framework [ 26 ]. Statistical Analysis For the primary regression task, model performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R²), Pearson r. Agreement between predicted and actual ASMI was assessed using Bland-Altman analysis. For the secondary classification task, performance was evaluated using the AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1-score, calculated at the predefined AWGS 2019 diagnostic thresholds. Subgroup analyses were performed based on gender and age. All statistical analyses were performed using Python (version 3.11.13) and R (version 4.2.3). A p-value < 0.05 was considered statistically significant. The deep learning model was implemented in PyTorch (version 2.8). Ethical Considerations This retrospective cohort study was approved by Institutional Review Board of the Taipei Medical University Ethics Committee (TMU-JIRB N201909036). Use of Artificial Intelligence Tools During the preparation of this work, the authors used Claude Code to assist in writing code for the deep learning model development and Google Gemini to refine language, improve clarity, and review content based on co-author instructions. The authors reviewed and edited the content generated by these services and take full responsibility for the content of the published article. Results Baseline Characteristics of the Study Population The final study cohort comprised 1,267 patients who met the inclusion criteria. The cohort was divided chronologically into a development set (n = 1,140) and an external validation set (n = 127). As detailed in Table 1 , there were no statistically significant differences in age, gender, height, weight, BMI, or baseline ASMI between the development and external validation cohorts (all p > 0.05). The prevalence of sarcopenia, as defined by AWGS 2019 criteria, was 42.0% in the development set and 48.0% in the external validation set. A comparative analysis between patients with and without sarcopenia is presented in Online Resource 1 . In both the development and external validation sets, patients with sarcopenia had lower body weight and BMI. Regression Performance for ASMI Prediction Model performance for ASMI prediction from hip radiographs is presented in Fig. 2 . In the 5-fold cross-validation, the model achieved a Pearson r of 0.774, a R² of 0.599, a MAE of 0.471 kg/m², and a RMSE of 0.657 kg/m². In the external validation set, the model yielded a Pearson r of 0.806, an R² of 0.650, an MAE of 0.414 kg/m², and an RMSE of 0.565 kg/m². Agreement with DXA Gold Standard Bland-Altman analysis was performed to assess the agreement between model-predicted and DXA-measured ASMI values as shown in Online Resource 2 . For the 5-fold cross-validation set, the analysis showed a mean bias of 0.006 kg/m² with 95% limits of agreement (LoA) of -1.281 to 1.293 kg/m². In the external validation set, the mean bias was − 0.001 kg/m², with 95% LoA of -1.113 to 1.111 kg/m². The plots did not show a systematic bias across the range of ASMI values. Diagnostic Performance for Sarcopenia Classification For the classification of low muscle mass, the AUC was 0.878 in the 5-fold cross-validation and 0.874 in the external validation set as shown in Fig. 3 . When applying the established AWGS 2019 diagnostic thresholds, the performance metrics for the cross-validation and external validation sets were, respectively: sensitivity of 76.2% and 70.5%; specificity of 80.5% and 83.3%; PPV of 73.9% and 79.6%; and NPV of 82.4% and 75.3%. The further metrics and confusion matrix are provided in Online Resource 3 and Online Resource 4 . Stratified Analysis The model's diagnostic performance across demographic subgroups of gender and age is shown in Online Resource 5 . In the external validation set, the Pearson r between predicted and actual ASMI was 0.777 for females and 0.772 for males. The corresponding AUC for sarcopenia classification was 0.854 (95% CI: 0.778–0.931) for females and 0.864 (95% CI: 0.710–1.000) for males. For age-based subgroups in the external set, the Pearson r was 0.836 for the age younger than aged 60, 0.832 for the 60–75 age group, and 0.768 for patients older than aged 75. The AUC was 0.854 (95% CI: 0.681-1.000) patient younger than aged 60, 0.829 (95% CI: 0.722–0.935) for the 60–75 age group, and 0.937 (95% CI: 0.872–1.000) for the > 75 years group. Model Interpretability and Effect of Implant Preprocessing Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations indicated that the model's areas of activation localized over soft tissue regions corresponding to the gluteal and proximal thigh muscles as shown in Fig. 4 . This anatomical focus was maintained in cases with orthopedic hardware, which were processed by an automated implant detection and inpainting pipeline as shown in Online Resource 6 . A comparison of model performance on the external validation dataset with and without this preprocessing step showed that the model trained with implant inpainting achieved higher values across multiple metrics, including a Pearson r of 0.80 (vs. 0.78 without) and an AUC of 0.87 (vs.0.86 without) as shown in Table 2 . Table 2 Performance comparison between with implant removal and without implant removal on external validation dataset. Mean Absolute Error R 2 Pearson correlation coefficient AUC Sensitivity Specificity F1score Accuracy Without 0.47 0.57 0.78 0.86 0.67 0.86 0.74 0.77 With 0.41 0.65 0.80 0.87 0.70 0.83 0.74 0.77 Abbreviations: ASMI, appendicular skeletal muscle mass index; R 2 , coefficient of determination; AUC, area under the receiver operating characteristic curve. Discussion Principal Results This study demonstrates that a deep learning model can accurately predict DXA-derived ASMI from routine AP hip radiographs, achieving external validation performance with a Pearson r of 0.806, R² of 0.650, and AUC of 0.874 for sarcopenia classification. The model maintained sensitivity of 70.5% and specificity of 83.3% when applying AWGS 2019 diagnostic thresholds, with consistent performance across gender and age subgroups [ 20 ]. Further Grad-CAM visualization confirmed model focus on clinically relevant gluteal and proximal thigh musculature, demonstrating interpretability essential for clinical acceptance. To the best of our knowledge, this study is the first study to use hip radiographs to predict ASMI value. Comparison With Prior Work Our model compares favorably with recent deep learning approaches for sarcopenia screening across various imaging modalities. Our Pearson r of 0.806 for ASMI prediction is competitive with the findings of Ryu et al., who reported concordance coefficients of 0.76 for ALM prediction from chest X-rays [ 13 ]. Furthermore, our model’s classification performance is comparable to that reported by Gu et al. (2023), who achieved an AUC of 0.874 using abdominal CT scans [ 15 ]. Our approach is distinguished by its clinical interpretability and alignment with established diagnostic standards. Unlike previous studies using the Youden index, our model was validated directly against the sex-specific ASMI cutoffs defined by the AWGS 2019 guidelines, thereby ensuring consistency with current clinical standards and improving its translational applicability [ 20 ]. In addition, hip radiographs provide several advantages for opportunistic sarcopenia screening. Unlike CT-based methods that require precise landmarking at the L3 level, hip radiographs are widely available, involve substantially lower radiation exposure, and are often obtained in older adults at elevated risk for falls and fragility fractures [ 16 , 27 , 28 ]. Collectively, these attributes position our deep learning model as a clinically practical, safe, and integrated solution for opportunistic sarcopenia screening in real-world settings. The SARC-F questionnaire has been widely used since its introduction in 2013 as a rapid screening tool for sarcopenia. However, its diagnostic performance has been consistently limited. A meta-analysis by Voelker et al. reported that the sensitivity of SARC-F ranges from 28.9% to 55.3% across different diagnostic criteria, leading the authors to conclude that it is “nonoptimal for sarcopenia screening” despite its good reliability. Even the modified SARC-CalF, which incorporates calf circumference, demonstrates only modest improvement, with sensitivity ranging from 45.9% to 57.2% [ 29 ]. In contrast, our model achieved a sensitivity of 70.5% and a specificity of 83.3%, substantially outperforming traditional questionnaire-based approaches. By integrating quantitative image-derived features through deep learning, our artificial intelligence (AI) framework effectively addresses a critical gap in sarcopenia screening—bridging the divide between low-sensitivity questionnaires and resource-intensive imaging modalities such as DXA. Our Grad-CAM revealed model focus on gluteal and proximal thigh muscles, which carry substantial clinical significance beyond model validation. These muscles play crucial roles in maintaining lateral balance, stability, and fall prevention, with meta-analyses confirming that sarcopenia increases fall risk 1.52-fold (OR 1.52, 95% CI: 1.32–1.77) and studies demonstrating that hip abductor strength predicts both mobility performance and functional decline [ 30 , 31 ]. Unlike "black box" AI systems that may learn spurious correlations with non-anatomical features, our Grad-CAM analysis confirms the model focuses on muscle tissues directly relevant to sarcopenia pathophysiology, an anatomical fidelity essential for regulatory approval and physician trust [ 32 ]. This anatomical focus also suggests early sarcopenia detection from hip radiographs could inform not only nutritional interventions but also targeted exercise prescriptions to preserve gluteal strength and prevent fall-related injuries. The clinical implications of our model are particularly significant within the geriatric hip fracture population. A high prevalence of sarcopenia is well-documented in these patients, and its presence is strongly associated with poorer postoperative functional recovery [ 33 , 34 ]. Indeed, muscle mass serves as a critical predictor of functional prognosis; baseline sarcopenia status has been shown to predict the extent of muscle mass loss at one-year follow-up, which in turn reflects impaired functional recovery [ 35 ]. This risk is further exacerbated when sarcopenia co-exists with osteoporosis, leading to worsened bone health, increased fall risk, and significant impairment in activities of daily living [ 36 ]. Therefore, our deep learning model provides a crucial tool for early assessment in this specific fragility fracture population, enabling opportunistic screening from routine radiographs. This facilitates the timely identification of high-risk individuals, allowing clinicians to implement interventions to improve outcomes. This approach is synergistic with similar deep learning efforts that use hip radiographs for opportunistic osteoporosis screening [ 37 ]. Looking forward, integrating such an AI tool into a Fracture Liaison Service (FLS) holds considerable potential. A multipronged FLS strategy that incorporates sarcopenia screening has already demonstrated efficacy in improving osteoporosis treatment rates and reducing subsequent fracture risk [ 38 ]. An AI system that assists clinicians in the dual detection of osteoporosis and sarcopenia could thus be a powerful advancement in secondary fracture prevention, and its clinical effectiveness warrants further validation. This study has several key strengths, including its novel application and the use of explainable AI. A significant methodological strength is our automated implant masking pipeline ( Online Resource 6) , which demonstrably improved model performance (Pearson r increased from 0.78 to 0.80, Table 2 ) and ensures its applicability in a real-world orthopedic population where hardware is common. Limitations Nevertheless, we acknowledge limitations. The single-center, retrospective design requires further validation in multi-center, prospective trials to confirm generalizability. Second, our model predicts only LMM, a key component but not the sole determinant of a sarcopenia diagnosis, which also requires assessing muscle strength. Therefore, it should be used as a screening tool to identify at-risk individuals, not as a standalone diagnostic test. Finally, our cohort was predominantly female, reflecting typical referral patterns for DXA; while our subgroup analysis showed stable performance across genders, validation in a more gender-balanced population is warranted to confirm these findings. Future work should focus on prospective validation and clinical implementation studies to measure the real-world impact of this tool on sarcopenia diagnosis rates and patient outcomes. Conclusion We demonstrate that deep learning can predict DXA-derived ASMI from routine hip radiographs with high accuracy (r = 0.806, AUC = 0.874), matching complex imaging modalities while surpassing questionnaire-based screening. This represents the first quantitative ASMI prediction from hip radiographs—imaging routinely obtained in at-risk elderly—aligned with AWGS clinical thresholds. Grad-CAM confirmed anatomically relevant focus on gluteal and thigh musculature, while automated implant masking ensures robust real-world application. With over 1.6 million hip radiographs performed annually in aging populations, this approach transforms existing workflows into opportunistic screening infrastructure at near-zero marginal cost. Given sarcopenia's substantial burden, including falls, fractures, disability, and mortality, early detection through scalable AI-driven screening offers a practical pathway for timely intervention, particularly when integrated into FLS for comprehensive musculoskeletal risk assessment. Declarations Competing Interests The authors have no relevant financial or non-financial interests to disclose. Ethics Approval This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Institutional Review Board of the Taipei Medical University Ethics Committee (TMU-JIRB N201909036). Consent to Participate The requirement for informed consent was waived by the Institutional Review Board of Taipei Medical University due to the retrospective nature of the study and the use of de-identified data. Consent to Publish Not applicable as no identifying individual person’s data or images are presented. Funding The authors declare that no funds, grants, or other support were received during the preparation of this manuscript. Author Contributions: Ling Lee: Conceptualization, Methodology, Software, Formal analysis, Writing - Original Draft, Writing - Review & Editing, Visualization. Shu-Han Chuang: Conceptualization, Methodology, Writing - Review & Editing. Yi-Jie Kuo: Writing - Review & Editing, Supervision. Lien-Chen Wu: Writing - Review & Editing, Supervision. Yu-Pin Chen: Conceptualization, Methodology, Resources, Writing - Review & Editing, Supervision. Acknowledgments: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data Availability The datasets generated and analyzed during this study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request. References Yuan S, Larsson SC (2023) Epidemiology of sarcopenia: Prevalence, risk factors, and consequences. Metabolism 144:155533 PMID: 36907247. 10.1016/j.metabol.2023.155533 Beaudart C, Alcazar J, Aprahamian I, Batsis JA, Yamada Y, Prado CM et al (2025) Health outcomes of sarcopenia: a consensus report by the outcome working group of the Global Leadership Initiative in Sarcopenia (GLIS). Aging Clin Exp Res. ;37(1):100. PMID: 40120052. 10.1007/s40520-025-02995-9 Landi F, Calvani R, Cesari M, Tosato M, Martone AM, Ortolani E et al (2018) Sarcopenia: An Overview on Current Definitions, Diagnosis and Treatment. Curr Protein Pept Sci 19(7):633–638 PMID: 28595526. 10.2174/1389203718666170607113459 Goates S, Du K, Arensberg MB, Gaillard T, Guralnik J, Pereira SL (2019) Economic Impact of Hospitalizations in US Adults with Sarcopenia. J Frailty Aging 8(2):93–99 PMID: 30997923. 10.14283/jfa.2019.10 Chen LK, Liu LK, Woo J, Assantachai P, Auyeung TW, Bahyah KS et al (2014) Sarcopenia in Asia: consensus report of the Asian Working Group for Sarcopenia. J Am Med Dir Assoc 15(2):95–101 PMID: 24461239. 10.1016/j.jamda.2013.11.025 Cruz-Jentoft AJ, Bahat G, Bauer J, Boirie Y, Bruyère O, Cederholm T et al (2019) Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing 48(1):16–31 PMID: 30312372. 10.1093/ageing/afy169 Huang WF, Chen YC, Chang CS, Wu CH (2025) Definition and diagnosis of sarcopenia: Asia-Pacific perspectives. Osteoporos Sarcopenia 11(2 Suppl):2–10 PMID: 40718357. 10.1016/j.afos.2025.05.004 Tagliafico AS, Bignotti B, Torri L, Rossi F (2022) Sarcopenia: how to measure, when and why. Radiol Med 127(3):228–237 PMID: 35041137. 10.1007/s11547-022-01450-3 Achamrah N, Colange G, Delay J, Rimbert A, Folope V, Petit A et al (2018) Comparison of body composition assessment by DXA and BIA according to the body mass index: A retrospective study on 3655 measures. PLoS ONE 13(7):e0200465 PMID: 30001381. 10.1371/journal.pone.0200465 Cheng KY, Chow SK, Hung VW, Wong CH, Wong RM, Tsang CS et al (2021) Diagnosis of sarcopenia by evaluating skeletal muscle mass by adjusted bioimpedance analysis validated with dual-energy X-ray absorptiometry. J Cachexia Sarcopenia Muscle 12(6):2163–2173 PMID: 34609065. 10.1002/jcsm.12825 Hata A, Muraguchi Y, Nakatsugawa M, Wang X, Song J, Wada N et al (2024) Automated chest CT three-dimensional quantification of body composition: adipose tissue and paravertebral muscle. Sci Rep. ;14(1):32117. PMID: 39738489. 10.1038/s41598-024-83897-0 Yamashita R, Nishio M, Do RKG, Togashi K (2018) Convolutional neural networks: an overview and application in radiology. Insights Imaging 9(4):611–629 PMID: 29934920. 10.1007/s13244-018-0639-9 Ryu J, Eom S, Kim HC, Kim CO, Rhee Y, You SC et al (2023) Chest X-ray-based opportunistic screening of sarcopenia using deep learning. J Cachexia Sarcopenia Muscle 14(1):418–428 PMID: 36457204. 10.1002/jcsm.13144 Hwang D, Ahn S, Park YB, Kim SH, Han HS, Lee MC et al (2022) Deep Learning-Based Muscle Segmentation and Quantification of Full-Leg Plain Radiograph for Sarcopenia Screening in Patients Undergoing Total Knee Arthroplasty. J Clin Med. ;11(13). PMID: 35806895. 10.3390/jcm11133612 Gu S, Wang L, Han R, Liu X, Wang Y, Chen T et al (2023) Detection of sarcopenia using deep learning-based artificial intelligence body part measure system (AIBMS). Front Physiol. ;14:1092352. PMID: 36776966. 10.3389/fphys.2023.1092352 Yang Y, Komisar V, Shishov N, Lo B, Korall AM, Feldman F et al (2020) The Effect of Fall Biomechanics on Risk for Hip Fracture in Older Adults: A Cohort Study of Video-Captured Falls in Long-Term Care. J Bone Min Res 35(10):1914–1922 PMID: 32402136. 10.1002/jbmr.4048 Kara M, Kaymak B, Ata AM, Ozkal O, Kara O, Baki A et al (2020) STAR-Sonographic Thigh Adjustment Ratio: A Golden Formula for the Diagnosis of Sarcopenia. Am J Phys Med Rehabil 99(10):902–908 PMID: 32941253. 10.1097/PHM.0000000000001439 Park H, Kim H, Yoo J (2025) AI-Enhanced Lower Extremity X-Ray Segmentation: A Promising Tool for Sarcopenia Diagnosis. Healthcare. ;13(19):2488. PMID: 10.3390/healthcare13192488 Moons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg EW et al (2015) Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 162(1):W1–73 PMID: 25560730. 10.7326/M14-0698 Chen LK, Woo J, Assantachai P, Auyeung TW, Chou MY, Iijima K et al (2020) Asian Working Group for Sarcopenia: 2019 Consensus Update on Sarcopenia Diagnosis and Treatment. J Am Med Dir Assoc. ;21(3):300–7 e2. PMID: 32033882. 10.1016/j.jamda.2019.12.012 Cubuk ED, Zoph B, Mane D, Vasudevan V, Le QV (eds) (2019) Autoaugment: Learning augmentation strategies from data. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition Jin Q, Zou C, Cui H, Sun C, Huang S-W, Kuo Y-J et al (eds) (2023) Multi-modality Contrastive Learning for Sarcopenia Screening from Hip X-rays and Clinical Information. Springer Nature Switzerland, Cham Wang X, Girshick R, Gupta A, He K (eds) (2018) Non-local neural networks. Proceedings of the IEEE conference on computer vision and pattern recognition Lawrence IKL (1989) A Concordance Correlation Coefficient to Evaluate Reproducibility. Biometrics 45(1):255–268. 10.2307/2532051 Loshchilov I, Hutter F (eds) (2017) Decoupled Weight Decay Regularization. International Conference on Learning Representations Akiba T, Sano S, Yanase T, Ohta T, Koyama M, Optuna (2019) A Next-generation Hyperparameter Optimization Framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining; Anchorage, AK, USA: Association for Computing Machinery; pp. 2623–31 Messina C, Maffi G, Vitale JA, Ulivieri FM, Guglielmi G, Sconfienza LM (2018) Diagnostic imaging of osteoporosis and sarcopenia: a narrative review. Quant Imaging Med Surg 8(1):86–99 PMID: 29541625. 10.21037/qims.2018.01.01 Stuursma A, Stroot IAS, Vermeulen KM, Slart R, Greuter MJW, Mourits MJE et al (2024) Reliability, costs, and radiation dose of dual-energy X-ray absorptiometry in diagnosis of radiologic sarcopenia in surgically menopausal women. Insights Imaging. ;15(1):104. PMID: 38589691. 10.1186/s13244-024-01677-w Voelker SN, Michalopoulos N, Maier AB, Reijnierse EM (2021) Reliability and Concurrent Validity of the SARC-F and Its Modified Versions: A Systematic Review and Meta-Analysis. J Am Med Dir Assoc. ;22(9):1864–76 e16. PMID: 34144049. 10.1016/j.jamda.2021.05.011 Woods JL, Iuliano-Burns S, King SJ, Strauss BJ, Walker KZ (2011) Poor physical function in elderly women in low-level aged care is related to muscle strength rather than to measures of sarcopenia. Clin Interv Aging 6:67–76 PMID: 21472094. 10.2147/CIA.S16979 Zhang X, Huang P, Dou Q, Wang C, Zhang W, Yang Y et al (2020) Falls among older adults with sarcopenia dwelling in nursing home or community: A meta-analysis. Clin Nutr 39(1):33–39 PMID: 30665817. 10.1016/j.clnu.2019.01.002 McNamara SL, Yi PH, Lotter W (2024) The clinician-AI interface: intended use and explainability in FDA-cleared AI devices for medical image interpretation. NPJ Digit Med. ;7(1):80. PMID: 38531952. 10.1038/s41746-024-01080-1 Chen YP, Wong PK, Tsai MJ, Chang WC, Hsieh TS, Leu TH et al (2020) The high prevalence of sarcopenia and its associated outcomes following hip surgery in Taiwanese geriatric patients with a hip fracture. J Formos Med Assoc 119(12):1807–1816 PMID: 32107098. 10.1016/j.jfma.2020.02.004 Chiang MH, Kuo YJ, Chen YP (2021) The Association Between Sarcopenia and Postoperative Outcomes Among Older Adults With Hip Fracture: A Systematic Review. J Appl Gerontol 40(12):1903–1913 PMID: 33870747. 10.1177/07334648211006519 Chen YP, Kuo YJ, Hung SW, Wen TW, Chien PC, Chiang MH et al (2021) Loss of skeletal muscle mass can be predicted by sarcopenia and reflects poor functional recovery at one year after surgery for geriatric hip fractures. Injury 52(11):3446–3452 PMID: 34404509. 10.1016/j.injury.2021.08.007 Nguyen BTT, Lin AP, Yang WW, Cheng SJ, Kuo YJ, Nguyen TT et al (2024) Impacts of osteosarcopenia on musculoskeletal health, risks of falls and fractures, and activities of daily living among population aged 50 and above: an age- and sex-matched cross-sectional analysis. Aging Clin Exp Res 37(1):8 PMID: 39725822. 10.1007/s40520-024-02902-8 Chen YP, Chan WP, Zhang HW, Tsai ZR, Peng HC, Huang SW et al (2024) Automated osteoporosis classification and T-score prediction using hip radiographs via deep learning algorithm. Ther Adv Musculoskelet Dis 16:1759720X241237872 PMID: 38665415. 10.1177/1759720X241237872 Chen YP, Chang WC, Wen TW, Chien PC, Huang SW, Kuo YJ (2022) Multipronged Programmatic Strategy for Preventing Secondary Fracture and Facilitating Functional Recovery in Older Patients after Hip Fractures: Our Experience in Taipei Municipal Wanfang Hospital. Medicina (Kaunas). ;58(7). PMID: 35888594. 10.3390/medicina58070875 Supplementary Files OnlineResources.pdf Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Major revisions 03 Feb, 2026 Reviewers agreed at journal 21 Jan, 2026 Reviewers invited by journal 21 Jan, 2026 Editor invited by journal 21 Jan, 2026 Editor assigned by journal 19 Jan, 2026 First submitted to journal 17 Jan, 2026 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-8603410","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":578161265,"identity":"afcf8516-0e15-4b86-9395-5bc58e136639","order_by":0,"name":"Ling Lee","email":"","orcid":"","institution":"Taichung Veterans General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ling","middleName":"","lastName":"Lee","suffix":""},{"id":578161266,"identity":"892a1980-4197-47df-af92-85ac9fbe5957","order_by":1,"name":"Shu-Han Chuang","email":"","orcid":"","institution":"Changhua Christian Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shu-Han","middleName":"","lastName":"Chuang","suffix":""},{"id":578161267,"identity":"4270c8e1-6a96-4a56-ab98-20c3734524b8","order_by":2,"name":"Yi-Jie Kuo","email":"","orcid":"","institution":"Taipei Municipal Wan-Fang Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yi-Jie","middleName":"","lastName":"Kuo","suffix":""},{"id":578161268,"identity":"a6acf0a9-c38b-4f3c-8166-ba0af8f992ab","order_by":3,"name":"Lien-Chen Wu","email":"","orcid":"","institution":"Taipei Municipal Wan-Fang Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lien-Chen","middleName":"","lastName":"Wu","suffix":""},{"id":578161269,"identity":"d591e395-5098-45bf-9183-7aa02cf54ec1","order_by":4,"name":"Yu-Pin Chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAy0lEQVRIiWNgGAWjYBACfvbmww8SKmBcNiK0SPYcSzN4cIaBgQemhYeQFoMbOQaSD9tI05JgYJA47zCDvdgZA4YPZUCGRAIBh515kPAgcdthBh7pHAPGGeeADEJa+I4nHDBI3HYbrIWZtw2kl4AWhgOJDRKJc6Ba/hKjReBEMoNEYgNUCyMxWoCBzGaQcOw/D8/ttIKDPefSeXjuP8CvhZ+9//PDHzVpcuyzkzc++FFmLcfec4CAX6AAHBsHGIiIllEwCkbBKBgFRAAAKNlCmM7XEUUAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-9729-6375","institution":"Taipei Municipal Wan-Fang Hospital","correspondingAuthor":true,"prefix":"","firstName":"Yu-Pin","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2026-01-14 15:42:36","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8603410/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8603410/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101074571,"identity":"c69cdf99-2fdf-47f1-9ba7-c76b3b20af55","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"xml","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8639,"visible":true,"origin":"","legend":"","description":"","filename":"egemEGEMD2600066.xml","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/58629010a789101b10be8f45.xml"},{"id":101205285,"identity":"004e24a0-49a9-4f41-ba6b-e2028d9b96bc","added_by":"auto","created_at":"2026-01-27 09:48:51","extension":"xml","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1138,"visible":true,"origin":"","legend":"","description":"","filename":"EGEMD260006621129.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/5f0b2d5711b78f517880ed12.xml"},{"id":101205428,"identity":"48b65023-4edb-424d-a551-724f908d7b64","added_by":"auto","created_at":"2026-01-27 09:49:22","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":948,"visible":true,"origin":"","legend":"","description":"","filename":"EGEMD2600066Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/407623836aa813b14f1ac038.xml"},{"id":101074576,"identity":"20d81d55-df07-41fb-8cec-72dd66ce384a","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"xml","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":113585,"visible":true,"origin":"","legend":"","description":"","filename":"EGEMD26000660enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/229fe62c3bdc5baeb910d047.xml"},{"id":101074581,"identity":"d657c300-b688-47e0-b4ab-46799458006d","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":207309,"visible":true,"origin":"","legend":"","description":"","filename":"Fig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/0db3c09080b104dadf9e9b43.png"},{"id":101206027,"identity":"2b0bff0d-693b-4c94-a29d-ddd32b7dadd5","added_by":"auto","created_at":"2026-01-27 09:51:23","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":249459,"visible":true,"origin":"","legend":"","description":"","filename":"Fig.2.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/75a63a32d993540b7f77c437.png"},{"id":101074574,"identity":"88e099e1-fce3-40d3-9a7d-de012b377e4a","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":147814,"visible":true,"origin":"","legend":"","description":"","filename":"Fig.3.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/dabce17932f147e6b5eabd8f.png"},{"id":101205892,"identity":"ddff01c5-6897-4636-a670-29cab9a1d039","added_by":"auto","created_at":"2026-01-27 09:50:29","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":408772,"visible":true,"origin":"","legend":"","description":"","filename":"Fig.4.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/0849ee9029edd81121c1336f.png"},{"id":101074586,"identity":"fb0ad925-6d26-4fe4-af7f-bf213e021169","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38215,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/dad88e146231cccd53491429.png"},{"id":101205354,"identity":"9da73a14-e010-4176-ad2b-742f3ad12241","added_by":"auto","created_at":"2026-01-27 09:49:07","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38289,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig.2.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/c1a5e2507b77fed77fb6b6dd.png"},{"id":101074585,"identity":"e11c9aa2-b126-4f68-8db3-1b942d9d1074","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32066,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig.3.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/f9d27f186cbd42cc63b3d693.png"},{"id":101205423,"identity":"e2bde4ac-4eaf-441b-a1a4-7d801501c331","added_by":"auto","created_at":"2026-01-27 09:49:21","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60953,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig.4.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/4afa1d26886d29168e3d4c73.png"},{"id":101074587,"identity":"f865b56f-ecfc-4245-bb67-ee0618e4d36b","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":110821,"visible":true,"origin":"","legend":"","description":"","filename":"EGEMD26000660structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/254e8bc468718c82473efde7.xml"},{"id":101205054,"identity":"986ba0a6-70e1-4011-b787-b35dbbd804e9","added_by":"auto","created_at":"2026-01-27 09:45:56","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125702,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/94cf72ee62fa035e66168358.html"},{"id":101205274,"identity":"b5301328-34ce-4421-a409-44279b33ce66","added_by":"auto","created_at":"2026-01-27 09:48:49","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":242957,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic of datasets generation.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/d49035024e18ae7d241a4ddd.png"},{"id":101074582,"identity":"887860ff-72b5-4865-9e70-23cde800c2a5","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":249459,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe scatter plot analysis of ASMI predicted by the deep learning model versus actual values measured by DXA in the development and external validation datasets. \u003c/strong\u003eThe dashed gray line represents the line of identity (y=x). The solid line is the linear regression fit, with the shaded area representing the 95% confidence interval.\u003c/p\u003e\n\u003cp\u003eAbbreviations: ASMI, appendicular skeletal muscle mass index; DXA, dual-energy X-ray absorptiometry; MAE, mean absolute error; R², coefficient of determination.\u003c/p\u003e","description":"","filename":"Fig.2.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/57de61f4f8c8aad88f0b20fb.png"},{"id":101074573,"identity":"24cb4239-21ce-4417-b13a-d965d9aed4b1","added_by":"auto","created_at":"2026-01-25 10:21:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":147814,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe ROC curve of deep learning model predictions based on predicted ASMI value. \u003c/strong\u003eThe cut-off point was selected based on the diagnosis cut point of Asian Working Group for Sarcopenia and presented using a circle mark.\u003c/p\u003e\n\u003cp\u003eAbbreviations: AUC, Area Under the Curve; Sens., sensitivity; Spec., specificity; PPV, positive predictive value; NPV, negative predictive value.\u003c/p\u003e","description":"","filename":"Fig.3.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/f1dabe09ad854bbeede0501f.png"},{"id":101205278,"identity":"ccd7f03f-aad9-48bb-a8e7-d867e002b63f","added_by":"auto","created_at":"2026-01-27 09:48:49","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":408772,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClass activation map (CAM) of selected patients with predicted and actual ASMI value comparison in external validation set.\u003c/strong\u003e Images demonstrate how artificial intelligence (AI) makes decisions by focusing on gluteal and proximal thigh muscles.\u003c/p\u003e\n\u003cp\u003eAbbreviations: ASMI, appendicular skeletal muscle mass index.\u003c/p\u003e","description":"","filename":"Fig.4.png","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/0d38f4a651afd0e6ab3da024.png"},{"id":101208045,"identity":"97234cc0-e7be-491b-90cc-9b0e23bdf504","added_by":"auto","created_at":"2026-01-27 10:08:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2116117,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/07932313-cc2a-4e7a-a140-0b4b357c652f.pdf"},{"id":101205332,"identity":"9f3483b8-31c3-4c4c-a249-6eda12f03ba4","added_by":"auto","created_at":"2026-01-27 09:49:03","extension":"pdf","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":1188547,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineResources.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8603410/v1/9300166e0e2e116cb22fbca7.pdf"}],"financialInterests":"","formattedTitle":"AI-Driven Appendicular Skeletal Muscle Mass Index (ASMI) Prediction and Low Muscle Mass Detection from Routine Hip X-rays: A Novel Opportunistic Screening Tool","fulltext":[{"header":"Aim","content":"\u003cp\u003eTo develop and validate a deep learning model that predicts quantitative Appendicular Skeletal Muscle Mass Index (ASMI) and identifies low muscle mass using routine hip radiographs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFindings\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe deep learning model accurately predicted ASMI (Pearson r = 0.806) and effectively classified low muscle mass (AUC = 0.874), with interpretability confirmed by focus on gluteal and thigh muscles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMessage\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis AI model enables opportunistic sarcopenia screening from standard hip X-rays without additional radiation or cost, potentially facilitating early intervention in at-risk populations.\u003c/p\u003e"},{"header":"Introduction","content":"\u003cp\u003eSarcopenia, characterized by the progressive and generalized loss of skeletal muscle mass, strength, and function, constitutes a significant public health challenge, particularly within the context of a rapidly aging global population [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Beyond normal aging, sarcopenia strongly predicts numerous adverse outcomes, including an increased risk of falls, fragility fractures, profound disability, loss of physical independence, and premature mortality [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Sarcopenia affects 5\u0026ndash;13% of adults aged 60\u0026ndash;70 and increasing to as high as 50% in individuals over 80. This imposes a considerable economic burden on healthcare systems worldwide, with associated costs in the United States alone estimated in the tens of billions of dollars annually [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. These severe clinical and economic problems underscore the urgent need for effective and scalable strategies for the early detection of sarcopenia, which would enable timely interventions to mitigate its debilitating effects.\u003c/p\u003e \u003cp\u003eSarcopenia is diagnosed in accordance with the criteria of the Working Group on Sarcopenia in Older People and is characterized by both low muscle mass (LMM) and low muscle strength [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Muscle strength is easily accessed via measurements such as handgrip strength or walking speed, which are simple and widely used in clinical practice for sarcopenia screening. However, muscle mass assessment requires specialized equipment, representing a major barrier to large-scale screening. Bioelectrical impedance analysis (BIA) and dual-energy X-ray absorptiometry (DXA) are the primary modalities for assessing muscle mass, while imaging techniques such as computed tomography (CT) and magnetic resonance imaging (MRI) provide highly accurate but less practical alternatives. Despite their precision, CT and MRI are unsuitable for population-level screening due to high cost, limited accessibility, the need for specialized personnel and equipment, and radiation exposure [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Although BIA is noninvasive and free from radiation, it suffers from considerable measurement variability and limited availability, reducing its reliability for sarcopenia screening [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. In contrast, DXA remains the reference standard for assessing muscle mass and diagnosing sarcopenia [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], but its high cost and limited accessibility continue to restrict its widespread application in routine practice, particularly for screening purposes.\u003c/p\u003e \u003cp\u003eOpportunistic screening involves leveraging routinely acquired medical images to assess for findings beyond the primary clinical indication, such as quantifying muscle and bone metrics [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Concurrently, the field of deep learning, particularly convolutional neural networks (CNNs), has advanced rapidly in medical imaging. CNNs learn hierarchical image features and have demonstrated expert-level performance in tasks such as lesion detection, classification, and segmentation [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Recent studies have exploited these advances to quantify muscle mass from routine images. For example, Ryu et al. (2023) trained a CNN using chest X-rays to predict appendicular lean mass (ALM) and sarcopenia status, reporting strong agreement with DXA measurements (concordance\u0026thinsp;\u0026asymp;\u0026thinsp;0.80) and an Area Under the Receiver Operating Characteristic Curve (AUC) of approximately 0.81 for sarcopenia detection [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Similarly, Hwang et al. (2022) employed a U-Net architecture to segment muscle on full-leg radiographs, achieving a high intersection over union (IoU\u0026thinsp;\u0026asymp;\u0026thinsp;0.93) with manual segmentations and an AUC of 0.988 for sarcopenia screening based on the predicted muscle volume [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Furthermore, Gu et al. (2023) applied deep learning to abdominal CT scans for body composition analysis, achieving high segmentation accuracy (DSC\u0026thinsp;\u0026gt;\u0026thinsp;0.90) and an AUC of 0.874 for automated sarcopenia classification [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. These studies demonstrate the feasibility of using deep learning method analysis of routine imaging to reliably quantify muscle mass or sarcopenia classification.\u003c/p\u003e \u003cp\u003eAnteroposterior (AP) hip radiographs are commonly performed in older adults presenting with hip pain or after a fall, given that over 95% of hip fractures in this demographic result from low-energy trauma [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Individuals with sarcopenia exhibit reduced appendicular muscle mass, and the hip and thigh regions contain the largest proportion of skeletal muscle in the body. Moreover, thigh muscle measurements are widely recognized as a standard reference for evaluating muscle wasting [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. These factors highlight the potential of applying deep learning algorithms to AP hip radiographs for opportunistic sarcopenia screening. Therefore, we hypothesized that a CNN could be trained to predict DXA-derived appendicular skeletal muscle mass index (ASMI) from routine hip radiographs, providing a practical and accessible alternative for sarcopenia screening [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTherefore, this study addresses the critical need for an accessible sarcopenia screening tool. Our primary objective was to develop and rigorously validate a deep learning model capable of predicting continuous, quantitative DXA-derived ASMI values directly from standard AP hip radiographs and clinical demographic data. The secondary objective was to evaluate the model's diagnostic performance in classifying LMM according to established clinical guidelines, thereby establishing its feasibility as a novel tool for opportunistic sarcopenia screening.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy Design and Data collection\u003c/h2\u003e \u003cp\u003eThis study is reported in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) statement [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. We identified all patients who underwent both a clinical DXA scan and a hip radiograph between January 2017 and December 2022. Each patient\u0026rsquo;s information including age, gender, height, weight, body mass index (BMI), ASMI, total lean mass, total fat, leg lean mass, and leg fat was collected. We used 5 numerical variables including age, gender, height, weight, and BMI as clinical information for boosting learning which was suggested by the surgeon. Patients were excluded from this study if they met any of the following criteria: (1) a time interval greater than 90 days between the two examinations; (2) the presence of missing image data or essential demographic and clinical records; or (3) recorded ASMI which may be the significant outlier and likely a data entry error. After exclusion, a final cohort of 1,267 patients was established. This cohort was then chronologically split, with the initial 90% of patients (n\u0026thinsp;=\u0026thinsp;1140) forming the development set for model training and internal validation, and the subsequent 10% (n\u0026thinsp;=\u0026thinsp;127) serving as a temporally distinct external validation set to assess generalizability. To the best of our knowledge, this is the largest dataset for automated sarcopenia diagnosis from images and tabular information to date. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the flowchart for the generation of each dataset. The demographic and clinical characteristics of this dataset are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \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\u003eBaseline characteristics in each dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\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\u003eDevelopment\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;1140)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExternal validation\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;127)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSarcopenia\u003c/p\u003e \u003cp\u003eWith\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e479(42%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e61(42%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.225\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWithout\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e661(58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66(58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.225\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDemography\u003c/b\u003e\u003c/p\u003e \u003cp\u003eGender (male)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e257(22.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e36(28.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.172\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge (years)\u003c/p\u003e \u003cp\u003eHeight (cm)\u003c/p\u003e \u003cp\u003eWeight (kg)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72.2\u0026thinsp;\u0026plusmn;\u0026thinsp;12.1\u003c/p\u003e \u003cp\u003e156.4\u0026thinsp;\u0026plusmn;\u0026thinsp;8.2\u003c/p\u003e \u003cp\u003e56.8\u0026thinsp;\u0026plusmn;\u0026thinsp;10.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e71.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.6\u003c/p\u003e \u003cp\u003e156.8\u0026thinsp;\u0026plusmn;\u0026thinsp;8.8\u003c/p\u003e \u003cp\u003e57.6\u0026thinsp;\u0026plusmn;\u0026thinsp;11.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.325\u003c/p\u003e \u003cp\u003e0.652\u003c/p\u003e \u003cp\u003e0.440\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBMI (kg/m\u003csup\u003e2\u003c/sup\u003e) ASMI (kg/m\u003csup\u003e2\u003c/sup\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e23.2\u0026thinsp;\u0026plusmn;\u0026thinsp;3.8\u003c/p\u003e \u003cp\u003e5.9\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23.5\u0026thinsp;\u0026plusmn;\u0026thinsp;3.9\u003c/p\u003e \u003cp\u003e6.0\u0026thinsp;\u0026plusmn;\u0026thinsp;0.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.358\u003c/p\u003e \u003cp\u003e0.883\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003eAbbreviations: ASMI, appendicular skeletal muscle mass index.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy Outcomes and Definitions\u003c/h3\u003e\n\u003cp\u003eThe ground truth for the primary regression task was the ASMI, calculated by normalizing DXA derived ALM to height squared, expressed in kg/m\u0026sup2;. ASMI is integral to clinical definitions of sarcopenia, including the criteria established by the Asian Working Group for Sarcopenia (AWGS). According to AWGS guidelines, LMM is defined by DXA-based ASMI cutoffs of \u0026lt;\u0026thinsp;7.0 kg/m\u0026sup2; for men and \u0026lt;\u0026thinsp;5.4 kg/m\u0026sup2; for women.[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] For the secondary classification task, the model's ability to identify patients with LMM was evaluated. Because the diagnostic thresholds for sarcopenia are gender-specific (ASMI\u0026thinsp;\u0026lt;\u0026thinsp;7.0 kg/m\u0026sup2; for males and \u0026lt;\u0026thinsp;5.4 kg/m\u0026sup2; for females), a unified risk score was calculated for each patient to enable receiver operating characteristic (ROC) analysis across the entire cohort [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. This score was defined as the gender-specific threshold minus the model-predicted ASMI value. A ROC curve was then generated using this risk score, and the AUC and further metrics were calculated.\u003c/p\u003e\n\u003ch3\u003eImage Preprocessing\u003c/h3\u003e\n\u003cp\u003eRadiographs in Digital Imaging and Communications in Medicine (DICOM) format were converted to grayscale images. Pixel intensity was normalized using window/level parameters from the DICOM metadata. To reduce confounding effects from metallic hardware, we implemented an automated implant detection algorithm. Regions with pixel intensities exceeding a threshold of 240 (on a 0-255 scale) were identified as implants and subsequently removed using an inpainting technique. All images were then resized to a standard 224\u0026times;224-pixel resolution and normalized to a [0, 1] range.\u003c/p\u003e \u003cp\u003eDuring the training phase, an extensive data augmentation strategy was employed to enhance model robustness and prevent overfitting. We implemented a comprehensive, policy-based augmentation scheme [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. For each image in a training batch, one of 30 predefined augmentation policies was stochastically selected and applied. These policies comprised a comprehensive set of transformations, including: (1) Geometric transformations, such as random rotation (up to \u0026plusmn;\u0026thinsp;30\u0026deg;), shearing, and translation. Horizontal flipping was specifically included to encourage the model to learn features invariant to bilateral differences in hip anatomy. (2) Photometric transformations, involving adjustments to contrast, brightness, sharpness, and color saturation, alongside operations like equalization, auto-contrast, solarization, and inversion. (3) Occlusion-based augmentation, which applied a cutout technique to randomly mask a square region of the image, compelling the model to learn from partial structural information. No augmentation was applied to the validation or external test sets to ensure a consistent and unbiased evaluation of the model's performance.\u003c/p\u003e\n\u003ch3\u003eDeep Learning Model Architecture\u003c/h3\u003e\n\u003cp\u003eWe developed a multimodal deep learning model to predict ASMI from hip radiographs and associated clinical parameters. The model architecture was substantially informed by the work of Jin et al., with modifications to adapt the framework for a regression task and to incorporate advanced training strategies [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. The architecture comprises three main components: (1) an image feature extraction pathway using a ResNet-34 backbone pretrained on ImageNet; (2) a clinical feature encoder (TextNet) for demographic and anthropometric data; and (3) a gated attention fusion module to integrate both modalities.\u003c/p\u003e \u003cp\u003eThe image pathway processes 224\u0026times;224-pixel grayscale radiographs through the ResNet-34 backbone. A non-local self-attention block was applied to the final convolutional layer to capture long-range spatial dependencies relevant to muscle mass [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Clinical features (age, gender, height, weight, BMI) were encoded via a 1-dimensional CNN (TextNet) to produce a 64-dimensional feature vector. The gated attention fusion module then combined image features (512 dimensions) with spatially expanded clinical features using a learnable gating mechanism, allowing the model to dynamically weight the contribution of each data type.\u003c/p\u003e \u003cp\u003eTo enhance model\u0026rsquo;s focus initially, Class Activation Mapping (CAM) enhancement was incorporated during training, where spatial attention masks were elementwise multiplied with input images using the formula: \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{X}_{f}=X\\times\\:(1+\\sigma\\:(CAM\\left)\\right),\\)\u003c/span\u003e\u003c/span\u003e where σ is the sigmoid function. Furthermore, contrastive learning was implemented using a noise-contrastive estimation (InfoNCE) loss function to improve the quality of feature representations, with a temperature parameter τ\u0026thinsp;=\u0026thinsp;0.117 and a weighting factor β\u0026thinsp;=\u0026thinsp;0.011. The final fused features were passed through a regression head to output a continuous ASMI prediction. Both CAM enhancement and contrastive learning strategy was also informed by the work of Jin et al [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003eModel Training Protocol\u003c/h3\u003e\n\u003cdiv class=\"Heading\"\u003eModel Training Protocol\u003c/div\u003e \u003cp\u003eThe model was trained using a 5-fold cross-validation strategy on the development set. The model was optimized using the Concordance Correlation Coefficient (CCC) loss function, which is well-suited for clinical agreement studies:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\:CCC=\\frac{2\\rho\\:{\\sigma\\:}_{x}{\\sigma\\:}_{y}}{{{\\sigma\\:}_{x}}^{2}+{{\\sigma\\:}_{y}}^{2}{+({\\mu\\:}_{x}-{\\mu\\:}_{y})}^{2}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere ρ is the Pearson correlation coefficient (r), σ is the standard deviation, and \u0026micro; is the mean of the predicted and actual ASMI values [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eOptimization was performed using the AdamW algorithm with an initial learning rate of 7.5\u0026times;10⁻⁵, weight decay of 7.5\u0026times;10⁻⁶, and gradient clipping at a norm of 1.0 [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The learning rate was dynamically adjusted during training; it was reduced by a factor of 0.33 whenever the validation loss plateaued for 8 consecutive epochs. Training was conducted for a maximum of 80 epochs with an early stopping criterion of 20 epochs. Batch sizes were set to 32 for training and 16 for validation and testing. The hyperparameters were identified as optimal through a systematic search conducted with the Optuna hyperparameter optimization framework [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e].\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eFor the primary regression task, model performance was evaluated using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R\u0026sup2;), Pearson r. Agreement between predicted and actual ASMI was assessed using Bland-Altman analysis. For the secondary classification task, performance was evaluated using the AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1-score, calculated at the predefined AWGS 2019 diagnostic thresholds. Subgroup analyses were performed based on gender and age. All statistical analyses were performed using Python (version 3.11.13) and R (version 4.2.3). A p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant. The deep learning model was implemented in PyTorch (version 2.8).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eEthical Considerations\u003c/h3\u003e\n\u003cp\u003e This retrospective cohort study was approved by Institutional Review Board of the Taipei Medical University Ethics Committee (TMU-JIRB N201909036).\u003c/p\u003e\n\u003ch3\u003eUse of Artificial Intelligence Tools\u003c/h3\u003e\n\u003cp\u003eDuring the preparation of this work, the authors used Claude Code to assist in writing code for the deep learning model development and Google Gemini to refine language, improve clarity, and review content based on co-author instructions. The authors reviewed and edited the content generated by these services and take full responsibility for the content of the published article.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eBaseline Characteristics of the Study Population\u003c/h2\u003e \u003cp\u003eThe final study cohort comprised 1,267 patients who met the inclusion criteria. The cohort was divided chronologically into a development set (n\u0026thinsp;=\u0026thinsp;1,140) and an external validation set (n\u0026thinsp;=\u0026thinsp;127). As detailed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, there were no statistically significant differences in age, gender, height, weight, BMI, or baseline ASMI between the development and external validation cohorts (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.05). The prevalence of sarcopenia, as defined by AWGS 2019 criteria, was 42.0% in the development set and 48.0% in the external validation set. A comparative analysis between patients with and without sarcopenia is presented in \u003cb\u003eOnline Resource 1\u003c/b\u003e. In both the development and external validation sets, patients with sarcopenia had lower body weight and BMI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eRegression Performance for ASMI Prediction\u003c/h2\u003e \u003cp\u003eModel performance for ASMI prediction from hip radiographs is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. In the 5-fold cross-validation, the model achieved a Pearson r of 0.774, a R\u0026sup2; of 0.599, a MAE of 0.471 kg/m\u0026sup2;, and a RMSE of 0.657 kg/m\u0026sup2;. In the external validation set, the model yielded a Pearson r of 0.806, an R\u0026sup2; of 0.650, an MAE of 0.414 kg/m\u0026sup2;, and an RMSE of 0.565 kg/m\u0026sup2;.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eAgreement with DXA Gold Standard\u003c/h2\u003e \u003cp\u003eBland-Altman analysis was performed to assess the agreement between model-predicted and DXA-measured ASMI values as shown in \u003cb\u003eOnline Resource 2\u003c/b\u003e. For the 5-fold cross-validation set, the analysis showed a mean bias of 0.006 kg/m\u0026sup2; with 95% limits of agreement (LoA) of -1.281 to 1.293 kg/m\u0026sup2;. In the external validation set, the mean bias was \u0026minus;\u0026thinsp;0.001 kg/m\u0026sup2;, with 95% LoA of -1.113 to 1.111 kg/m\u0026sup2;. The plots did not show a systematic bias across the range of ASMI values.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eDiagnostic Performance for Sarcopenia Classification\u003c/h2\u003e \u003cp\u003eFor the classification of low muscle mass, the AUC was 0.878 in the 5-fold cross-validation and 0.874 in the external validation set as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. When applying the established AWGS 2019 diagnostic thresholds, the performance metrics for the cross-validation and external validation sets were, respectively: sensitivity of 76.2% and 70.5%; specificity of 80.5% and 83.3%; PPV of 73.9% and 79.6%; and NPV of 82.4% and 75.3%. The further metrics and confusion matrix are provided in \u003cb\u003eOnline Resource 3\u003c/b\u003e and \u003cb\u003eOnline Resource 4\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eStratified Analysis\u003c/h2\u003e \u003cp\u003eThe model's diagnostic performance across demographic subgroups of gender and age is shown in \u003cb\u003eOnline Resource 5\u003c/b\u003e. In the external validation set, the Pearson r between predicted and actual ASMI was 0.777 for females and 0.772 for males. The corresponding AUC for sarcopenia classification was 0.854 (95% CI: 0.778\u0026ndash;0.931) for females and 0.864 (95% CI: 0.710\u0026ndash;1.000) for males. For age-based subgroups in the external set, the Pearson r was 0.836 for the age younger than aged 60, 0.832 for the 60\u0026ndash;75 age group, and 0.768 for patients older than aged 75. The AUC was 0.854 (95% CI: 0.681-1.000) patient younger than aged 60, 0.829 (95% CI: 0.722\u0026ndash;0.935) for the 60\u0026ndash;75 age group, and 0.937 (95% CI: 0.872\u0026ndash;1.000) for the \u0026gt;\u0026thinsp;75 years group.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eModel Interpretability and Effect of Implant Preprocessing\u003c/h2\u003e \u003cp\u003eGradient-weighted Class Activation Mapping (Grad-CAM) visualizations indicated that the model's areas of activation localized over soft tissue regions corresponding to the gluteal and proximal thigh muscles as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. This anatomical focus was maintained in cases with orthopedic hardware, which were processed by an automated implant detection and inpainting pipeline as shown in \u003cb\u003eOnline Resource 6\u003c/b\u003e. A comparison of model performance on the external validation dataset with and without this preprocessing step showed that the model trained with implant inpainting achieved higher values across multiple metrics, including a Pearson r of 0.80 (vs. 0.78 without) and an AUC of 0.87 (vs.0.86 without) as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance comparison between with implant removal and without implant removal on external validation dataset.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\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 \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\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\u003eMean\u003c/p\u003e \u003cp\u003eAbsolute\u003c/p\u003e \u003cp\u003eError\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eR\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePearson correlation coefficient\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eF1score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWithout\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWith\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"9\"\u003eAbbreviations: ASMI, appendicular skeletal muscle mass index; R\u003csup\u003e2\u003c/sup\u003e, coefficient of determination; AUC, area under the receiver operating characteristic curve.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003ePrincipal Results\u003c/h2\u003e \u003cp\u003eThis study demonstrates that a deep learning model can accurately predict DXA-derived ASMI from routine AP hip radiographs, achieving external validation performance with a Pearson r of 0.806, R\u0026sup2; of 0.650, and AUC of 0.874 for sarcopenia classification. The model maintained sensitivity of 70.5% and specificity of 83.3% when applying AWGS 2019 diagnostic thresholds, with consistent performance across gender and age subgroups [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Further Grad-CAM visualization confirmed model focus on clinically relevant gluteal and proximal thigh musculature, demonstrating interpretability essential for clinical acceptance. To the best of our knowledge, this study is the first study to use hip radiographs to predict ASMI value.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eComparison With Prior Work\u003c/h2\u003e \u003cp\u003eOur model compares favorably with recent deep learning approaches for sarcopenia screening across various imaging modalities. Our Pearson r of 0.806 for ASMI prediction is competitive with the findings of Ryu et al., who reported concordance coefficients of 0.76 for ALM prediction from chest X-rays [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Furthermore, our model\u0026rsquo;s classification performance is comparable to that reported by Gu et al. (2023), who achieved an AUC of 0.874 using abdominal CT scans [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Our approach is distinguished by its clinical interpretability and alignment with established diagnostic standards. Unlike previous studies using the Youden index, our model was validated directly against the sex-specific ASMI cutoffs defined by the AWGS 2019 guidelines, thereby ensuring consistency with current clinical standards and improving its translational applicability [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. In addition, hip radiographs provide several advantages for opportunistic sarcopenia screening. Unlike CT-based methods that require precise landmarking at the L3 level, hip radiographs are widely available, involve substantially lower radiation exposure, and are often obtained in older adults at elevated risk for falls and fragility fractures [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Collectively, these attributes position our deep learning model as a clinically practical, safe, and integrated solution for opportunistic sarcopenia screening in real-world settings.\u003c/p\u003e \u003cp\u003eThe SARC-F questionnaire has been widely used since its introduction in 2013 as a rapid screening tool for sarcopenia. However, its diagnostic performance has been consistently limited. A meta-analysis by Voelker et al. reported that the sensitivity of SARC-F ranges from 28.9% to 55.3% across different diagnostic criteria, leading the authors to conclude that it is \u0026ldquo;nonoptimal for sarcopenia screening\u0026rdquo; despite its good reliability. Even the modified SARC-CalF, which incorporates calf circumference, demonstrates only modest improvement, with sensitivity ranging from 45.9% to 57.2% [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. In contrast, our model achieved a sensitivity of 70.5% and a specificity of 83.3%, substantially outperforming traditional questionnaire-based approaches. By integrating quantitative image-derived features through deep learning, our artificial intelligence (AI) framework effectively addresses a critical gap in sarcopenia screening\u0026mdash;bridging the divide between low-sensitivity questionnaires and resource-intensive imaging modalities such as DXA.\u003c/p\u003e \u003cp\u003eOur Grad-CAM revealed model focus on gluteal and proximal thigh muscles, which carry substantial clinical significance beyond model validation. These muscles play crucial roles in maintaining lateral balance, stability, and fall prevention, with meta-analyses confirming that sarcopenia increases fall risk 1.52-fold (OR 1.52, 95% CI: 1.32\u0026ndash;1.77) and studies demonstrating that hip abductor strength predicts both mobility performance and functional decline [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. Unlike \"black box\" AI systems that may learn spurious correlations with non-anatomical features, our Grad-CAM analysis confirms the model focuses on muscle tissues directly relevant to sarcopenia pathophysiology, an anatomical fidelity essential for regulatory approval and physician trust [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]. This anatomical focus also suggests early sarcopenia detection from hip radiographs could inform not only nutritional interventions but also targeted exercise prescriptions to preserve gluteal strength and prevent fall-related injuries.\u003c/p\u003e \u003cp\u003eThe clinical implications of our model are particularly significant within the geriatric hip fracture population. A high prevalence of sarcopenia is well-documented in these patients, and its presence is strongly associated with poorer postoperative functional recovery [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. Indeed, muscle mass serves as a critical predictor of functional prognosis; baseline sarcopenia status has been shown to predict the extent of muscle mass loss at one-year follow-up, which in turn reflects impaired functional recovery [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. This risk is further exacerbated when sarcopenia co-exists with osteoporosis, leading to worsened bone health, increased fall risk, and significant impairment in activities of daily living [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Therefore, our deep learning model provides a crucial tool for early assessment in this specific fragility fracture population, enabling opportunistic screening from routine radiographs. This facilitates the timely identification of high-risk individuals, allowing clinicians to implement interventions to improve outcomes. This approach is synergistic with similar deep learning efforts that use hip radiographs for opportunistic osteoporosis screening [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Looking forward, integrating such an AI tool into a Fracture Liaison Service (FLS) holds considerable potential. A multipronged FLS strategy that incorporates sarcopenia screening has already demonstrated efficacy in improving osteoporosis treatment rates and reducing subsequent fracture risk [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. An AI system that assists clinicians in the dual detection of osteoporosis and sarcopenia could thus be a powerful advancement in secondary fracture prevention, and its clinical effectiveness warrants further validation.\u003c/p\u003e \u003cp\u003eThis study has several key strengths, including its novel application and the use of explainable AI. A significant methodological strength is our automated implant masking pipeline (\u003cb\u003eOnline Resource 6)\u003c/b\u003e, which demonstrably improved model performance (Pearson r increased from 0.78 to 0.80, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) and ensures its applicability in a real-world orthopedic population where hardware is common.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eNevertheless, we acknowledge limitations. The single-center, retrospective design requires further validation in multi-center, prospective trials to confirm generalizability. Second, our model predicts only LMM, a key component but not the sole determinant of a sarcopenia diagnosis, which also requires assessing muscle strength. Therefore, it should be used as a screening tool to identify at-risk individuals, not as a standalone diagnostic test. Finally, our cohort was predominantly female, reflecting typical referral patterns for DXA; while our subgroup analysis showed stable performance across genders, validation in a more gender-balanced population is warranted to confirm these findings. Future work should focus on prospective validation and clinical implementation studies to measure the real-world impact of this tool on sarcopenia diagnosis rates and patient outcomes.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe demonstrate that deep learning can predict DXA-derived ASMI from routine hip radiographs with high accuracy (r\u0026thinsp;=\u0026thinsp;0.806, AUC\u0026thinsp;=\u0026thinsp;0.874), matching complex imaging modalities while surpassing questionnaire-based screening. This represents the first quantitative ASMI prediction from hip radiographs\u0026mdash;imaging routinely obtained in at-risk elderly\u0026mdash;aligned with AWGS clinical thresholds. Grad-CAM confirmed anatomically relevant focus on gluteal and thigh musculature, while automated implant masking ensures robust real-world application. With over 1.6\u0026nbsp;million hip radiographs performed annually in aging populations, this approach transforms existing workflows into opportunistic screening infrastructure at near-zero marginal cost. Given sarcopenia's substantial burden, including falls, fractures, disability, and mortality, early detection through scalable AI-driven screening offers a practical pathway for timely intervention, particularly when integrated into FLS for comprehensive musculoskeletal risk assessment.\u003c/p\u003e"},{"header":"Declarations","content":" \u003ch2\u003eCompeting Interests\u003c/h2\u003e \u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eEthics Approval\u003c/strong\u003e \u003cp\u003e This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Institutional Review Board of the Taipei Medical University Ethics Committee (TMU-JIRB N201909036).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent to Participate\u003c/strong\u003e \u003cp\u003e The requirement for informed consent was waived by the Institutional Review Board of Taipei Medical University due to the retrospective nature of the study and the use of de-identified data.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent to Publish\u003c/strong\u003e \u003cp\u003eNot applicable as no identifying individual person\u0026rsquo;s data or images are presented.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThe authors declare that no funds, grants, or other support were received during the preparation of this manuscript.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLing Lee: Conceptualization, Methodology, Software, Formal analysis, Writing - Original Draft, Writing - Review \u0026amp; Editing, Visualization.\u003c/p\u003e\n\u003cp\u003eShu-Han Chuang: Conceptualization, Methodology, Writing - Review \u0026amp; Editing.\u003c/p\u003e\n\u003cp\u003eYi-Jie Kuo: Writing - Review \u0026amp; Editing, Supervision.\u003c/p\u003e\n\u003cp\u003eLien-Chen Wu: Writing - Review \u0026amp; Editing, Supervision.\u003c/p\u003e\n\u003cp\u003eYu-Pin Chen: Conceptualization, Methodology, Resources, Writing - Review \u0026amp; Editing, Supervision.\u003c/p\u003e\u003ch2\u003eAcknowledgments:\u003c/h2\u003e \u003cp\u003eThis research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e \u003cp\u003eThe datasets generated and analyzed during this study are not publicly available due to patient privacy and institutional data protection policies but are available from the corresponding author on reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eYuan S, Larsson SC (2023) Epidemiology of sarcopenia: Prevalence, risk factors, and consequences. Metabolism 144:155533 PMID: 36907247. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.metabol.2023.155533\u003c/span\u003e\u003cspan address=\"10.1016/j.metabol.2023.155533\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaudart C, Alcazar J, Aprahamian I, Batsis JA, Yamada Y, Prado CM et al (2025) Health outcomes of sarcopenia: a consensus report by the outcome working group of the Global Leadership Initiative in Sarcopenia (GLIS). Aging Clin Exp Res. ;37(1):100. PMID: 40120052. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40520-025-02995-9\u003c/span\u003e\u003cspan address=\"10.1007/s40520-025-02995-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLandi F, Calvani R, Cesari M, Tosato M, Martone AM, Ortolani E et al (2018) Sarcopenia: An Overview on Current Definitions, Diagnosis and Treatment. Curr Protein Pept Sci 19(7):633\u0026ndash;638 PMID: 28595526. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2174/1389203718666170607113459\u003c/span\u003e\u003cspan address=\"10.2174/1389203718666170607113459\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoates S, Du K, Arensberg MB, Gaillard T, Guralnik J, Pereira SL (2019) Economic Impact of Hospitalizations in US Adults with Sarcopenia. J Frailty Aging 8(2):93\u0026ndash;99 PMID: 30997923. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.14283/jfa.2019.10\u003c/span\u003e\u003cspan address=\"10.14283/jfa.2019.10\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen LK, Liu LK, Woo J, Assantachai P, Auyeung TW, Bahyah KS et al (2014) Sarcopenia in Asia: consensus report of the Asian Working Group for Sarcopenia. J Am Med Dir Assoc 15(2):95\u0026ndash;101 PMID: 24461239. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jamda.2013.11.025\u003c/span\u003e\u003cspan address=\"10.1016/j.jamda.2013.11.025\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCruz-Jentoft AJ, Bahat G, Bauer J, Boirie Y, Bruy\u0026egrave;re O, Cederholm T et al (2019) Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing 48(1):16\u0026ndash;31 PMID: 30312372. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/ageing/afy169\u003c/span\u003e\u003cspan address=\"10.1093/ageing/afy169\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang WF, Chen YC, Chang CS, Wu CH (2025) Definition and diagnosis of sarcopenia: Asia-Pacific perspectives. Osteoporos Sarcopenia 11(2 Suppl):2\u0026ndash;10 PMID: 40718357. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.afos.2025.05.004\u003c/span\u003e\u003cspan address=\"10.1016/j.afos.2025.05.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTagliafico AS, Bignotti B, Torri L, Rossi F (2022) Sarcopenia: how to measure, when and why. Radiol Med 127(3):228\u0026ndash;237 PMID: 35041137. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11547-022-01450-3\u003c/span\u003e\u003cspan address=\"10.1007/s11547-022-01450-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAchamrah N, Colange G, Delay J, Rimbert A, Folope V, Petit A et al (2018) Comparison of body composition assessment by DXA and BIA according to the body mass index: A retrospective study on 3655 measures. PLoS ONE 13(7):e0200465 PMID: 30001381. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1371/journal.pone.0200465\u003c/span\u003e\u003cspan address=\"10.1371/journal.pone.0200465\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCheng KY, Chow SK, Hung VW, Wong CH, Wong RM, Tsang CS et al (2021) Diagnosis of sarcopenia by evaluating skeletal muscle mass by adjusted bioimpedance analysis validated with dual-energy X-ray absorptiometry. J Cachexia Sarcopenia Muscle 12(6):2163\u0026ndash;2173 PMID: 34609065. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/jcsm.12825\u003c/span\u003e\u003cspan address=\"10.1002/jcsm.12825\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHata A, Muraguchi Y, Nakatsugawa M, Wang X, Song J, Wada N et al (2024) Automated chest CT three-dimensional quantification of body composition: adipose tissue and paravertebral muscle. Sci Rep. ;14(1):32117. PMID: 39738489. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-024-83897-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-83897-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYamashita R, Nishio M, Do RKG, Togashi K (2018) Convolutional neural networks: an overview and application in radiology. Insights Imaging 9(4):611\u0026ndash;629 PMID: 29934920. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s13244-018-0639-9\u003c/span\u003e\u003cspan address=\"10.1007/s13244-018-0639-9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRyu J, Eom S, Kim HC, Kim CO, Rhee Y, You SC et al (2023) Chest X-ray-based opportunistic screening of sarcopenia using deep learning. J Cachexia Sarcopenia Muscle 14(1):418\u0026ndash;428 PMID: 36457204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/jcsm.13144\u003c/span\u003e\u003cspan address=\"10.1002/jcsm.13144\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHwang D, Ahn S, Park YB, Kim SH, Han HS, Lee MC et al (2022) Deep Learning-Based Muscle Segmentation and Quantification of Full-Leg Plain Radiograph for Sarcopenia Screening in Patients Undergoing Total Knee Arthroplasty. J Clin Med. ;11(13). PMID: 35806895. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/jcm11133612\u003c/span\u003e\u003cspan address=\"10.3390/jcm11133612\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGu S, Wang L, Han R, Liu X, Wang Y, Chen T et al (2023) Detection of sarcopenia using deep learning-based artificial intelligence body part measure system (AIBMS). Front Physiol. ;14:1092352. PMID: 36776966. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3389/fphys.2023.1092352\u003c/span\u003e\u003cspan address=\"10.3389/fphys.2023.1092352\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Y, Komisar V, Shishov N, Lo B, Korall AM, Feldman F et al (2020) The Effect of Fall Biomechanics on Risk for Hip Fracture in Older Adults: A Cohort Study of Video-Captured Falls in Long-Term Care. J Bone Min Res 35(10):1914\u0026ndash;1922 PMID: 32402136. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1002/jbmr.4048\u003c/span\u003e\u003cspan address=\"10.1002/jbmr.4048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKara M, Kaymak B, Ata AM, Ozkal O, Kara O, Baki A et al (2020) STAR-Sonographic Thigh Adjustment Ratio: A Golden Formula for the Diagnosis of Sarcopenia. Am J Phys Med Rehabil 99(10):902\u0026ndash;908 PMID: 32941253. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/PHM.0000000000001439\u003c/span\u003e\u003cspan address=\"10.1097/PHM.0000000000001439\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark H, Kim H, Yoo J (2025) AI-Enhanced Lower Extremity X-Ray Segmentation: A Promising Tool for Sarcopenia Diagnosis. Healthcare. ;13(19):2488. PMID: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/healthcare13192488\u003c/span\u003e\u003cspan address=\"10.3390/healthcare13192488\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoons KG, Altman DG, Reitsma JB, Ioannidis JP, Macaskill P, Steyerberg EW et al (2015) Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med 162(1):W1\u0026ndash;73 PMID: 25560730. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7326/M14-0698\u003c/span\u003e\u003cspan address=\"10.7326/M14-0698\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen LK, Woo J, Assantachai P, Auyeung TW, Chou MY, Iijima K et al (2020) Asian Working Group for Sarcopenia: 2019 Consensus Update on Sarcopenia Diagnosis and Treatment. J Am Med Dir Assoc. ;21(3):300\u0026ndash;7 e2. PMID: 32033882. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jamda.2019.12.012\u003c/span\u003e\u003cspan address=\"10.1016/j.jamda.2019.12.012\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCubuk ED, Zoph B, Mane D, Vasudevan V, Le QV (eds) (2019) Autoaugment: Learning augmentation strategies from data. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJin Q, Zou C, Cui H, Sun C, Huang S-W, Kuo Y-J et al (eds) (2023) Multi-modality Contrastive Learning for Sarcopenia Screening from Hip X-rays and Clinical Information. Springer Nature Switzerland, Cham\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang X, Girshick R, Gupta A, He K (eds) (2018) Non-local neural networks. Proceedings of the IEEE conference on computer vision and pattern recognition\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLawrence IKL (1989) A Concordance Correlation Coefficient to Evaluate Reproducibility. Biometrics 45(1):255\u0026ndash;268. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2307/2532051\u003c/span\u003e\u003cspan address=\"10.2307/2532051\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLoshchilov I, Hutter F (eds) (2017) Decoupled Weight Decay Regularization. International Conference on Learning Representations\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAkiba T, Sano S, Yanase T, Ohta T, Koyama M, Optuna (2019) A Next-generation Hyperparameter Optimization Framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \u0026amp; Data Mining; Anchorage, AK, USA: Association for Computing Machinery; pp. 2623\u0026ndash;31\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMessina C, Maffi G, Vitale JA, Ulivieri FM, Guglielmi G, Sconfienza LM (2018) Diagnostic imaging of osteoporosis and sarcopenia: a narrative review. Quant Imaging Med Surg 8(1):86\u0026ndash;99 PMID: 29541625. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.21037/qims.2018.01.01\u003c/span\u003e\u003cspan address=\"10.21037/qims.2018.01.01\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStuursma A, Stroot IAS, Vermeulen KM, Slart R, Greuter MJW, Mourits MJE et al (2024) Reliability, costs, and radiation dose of dual-energy X-ray absorptiometry in diagnosis of radiologic sarcopenia in surgically menopausal women. Insights Imaging. ;15(1):104. PMID: 38589691. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13244-024-01677-w\u003c/span\u003e\u003cspan address=\"10.1186/s13244-024-01677-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVoelker SN, Michalopoulos N, Maier AB, Reijnierse EM (2021) Reliability and Concurrent Validity of the SARC-F and Its Modified Versions: A Systematic Review and Meta-Analysis. J Am Med Dir Assoc. ;22(9):1864\u0026ndash;76 e16. PMID: 34144049. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jamda.2021.05.011\u003c/span\u003e\u003cspan address=\"10.1016/j.jamda.2021.05.011\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWoods JL, Iuliano-Burns S, King SJ, Strauss BJ, Walker KZ (2011) Poor physical function in elderly women in low-level aged care is related to muscle strength rather than to measures of sarcopenia. Clin Interv Aging 6:67\u0026ndash;76 PMID: 21472094. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.2147/CIA.S16979\u003c/span\u003e\u003cspan address=\"10.2147/CIA.S16979\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Huang P, Dou Q, Wang C, Zhang W, Yang Y et al (2020) Falls among older adults with sarcopenia dwelling in nursing home or community: A meta-analysis. Clin Nutr 39(1):33\u0026ndash;39 PMID: 30665817. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.clnu.2019.01.002\u003c/span\u003e\u003cspan address=\"10.1016/j.clnu.2019.01.002\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcNamara SL, Yi PH, Lotter W (2024) The clinician-AI interface: intended use and explainability in FDA-cleared AI devices for medical image interpretation. NPJ Digit Med. ;7(1):80. PMID: 38531952. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41746-024-01080-1\u003c/span\u003e\u003cspan address=\"10.1038/s41746-024-01080-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YP, Wong PK, Tsai MJ, Chang WC, Hsieh TS, Leu TH et al (2020) The high prevalence of sarcopenia and its associated outcomes following hip surgery in Taiwanese geriatric patients with a hip fracture. J Formos Med Assoc 119(12):1807\u0026ndash;1816 PMID: 32107098. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.jfma.2020.02.004\u003c/span\u003e\u003cspan address=\"10.1016/j.jfma.2020.02.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChiang MH, Kuo YJ, Chen YP (2021) The Association Between Sarcopenia and Postoperative Outcomes Among Older Adults With Hip Fracture: A Systematic Review. J Appl Gerontol 40(12):1903\u0026ndash;1913 PMID: 33870747. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/07334648211006519\u003c/span\u003e\u003cspan address=\"10.1177/07334648211006519\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YP, Kuo YJ, Hung SW, Wen TW, Chien PC, Chiang MH et al (2021) Loss of skeletal muscle mass can be predicted by sarcopenia and reflects poor functional recovery at one year after surgery for geriatric hip fractures. Injury 52(11):3446\u0026ndash;3452 PMID: 34404509. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.injury.2021.08.007\u003c/span\u003e\u003cspan address=\"10.1016/j.injury.2021.08.007\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen BTT, Lin AP, Yang WW, Cheng SJ, Kuo YJ, Nguyen TT et al (2024) Impacts of osteosarcopenia on musculoskeletal health, risks of falls and fractures, and activities of daily living among population aged 50 and above: an age- and sex-matched cross-sectional analysis. Aging Clin Exp Res 37(1):8 PMID: 39725822. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s40520-024-02902-8\u003c/span\u003e\u003cspan address=\"10.1007/s40520-024-02902-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YP, Chan WP, Zhang HW, Tsai ZR, Peng HC, Huang SW et al (2024) Automated osteoporosis classification and T-score prediction using hip radiographs via deep learning algorithm. Ther Adv Musculoskelet Dis 16:1759720X241237872 PMID: 38665415. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1177/1759720X241237872\u003c/span\u003e\u003cspan address=\"10.1177/1759720X241237872\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen YP, Chang WC, Wen TW, Chien PC, Huang SW, Kuo YJ (2022) Multipronged Programmatic Strategy for Preventing Secondary Fracture and Facilitating Functional Recovery in Older Patients after Hip Fractures: Our Experience in Taipei Municipal Wanfang Hospital. Medicina (Kaunas). ;58(7). PMID: 35888594. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/medicina58070875\u003c/span\u003e\u003cspan address=\"10.3390/medicina58070875\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\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":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-geriatric-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"EGEM","sideBox":"Learn more about [European Geriatric Medicine](https://www.springer.com/journal/41999)","snPcode":"41999","submissionUrl":"https://www.editorialmanager.com/egem/default2.aspx","title":"European Geriatric Medicine","twitterHandle":"","acdcEnabled":false,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Artificial intelligence, Sarcopenia, Low muscle mass, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-8603410/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8603410/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003ePurpose\u003c/h2\u003e \u003cp\u003eSarcopenia diagnosis requires identifying low muscle mass (LMM), typically via dual-energy X-ray absorptiometry (DXA). However, DXA's limited accessibility restricts large-scale screening. This retrospective study aimed to develop and validate a deep learning model to predict DXA-derived ASMI from routine hip radiographs for opportunistic sarcopenia screening.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe included 1,267 patients with both hip radiography and DXA scans, split into development (n\u0026thinsp;=\u0026thinsp;1,140) and external validation (n\u0026thinsp;=\u0026thinsp;127) sets. A multimodal model integrating radiographic images (ResNet-34 backbone) and clinical variables (age, gender, height, weight, BMI) was trained to predict continuous ASMI and classify LMM per Asian Working Group for Sarcopenia (AWGS) 2019 criteria.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eOn external validation, the model achieved strong performance with Pearson r\u0026thinsp;=\u0026thinsp;0.806, R\u0026sup2;=0.650, MAE\u0026thinsp;=\u0026thinsp;0.414 kg/m\u0026sup2;, and AUC\u0026thinsp;=\u0026thinsp;0.874 for LMM classification. Applying AWGS diagnostic thresholds yielded sensitivity of 70.5% and specificity of 83.3%, with consistent performance across gender and age subgroups. Gradient-weighted Class Activation Mapping confirmed focus on clinically relevant gluteal and proximal thigh muscles.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eThis deep learning approach enables automated LMM identification from routine hip radiographs, offering a cost-effective, accessible tool for opportunistic sarcopenia screening and early intervention in at-risk populations.\u003c/p\u003e","manuscriptTitle":"AI-Driven Appendicular Skeletal Muscle Mass Index (ASMI) Prediction and Low Muscle Mass Detection from Routine Hip X-rays: A Novel Opportunistic Screening Tool","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-25 10:21:10","doi":"10.21203/rs.3.rs-8603410/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Major revisions","date":"2026-02-03T22:30:52+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"","date":"2026-01-22T00:00:34+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-21T13:38:31+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"European Geriatric Medicine","date":"2026-01-21T09:53:19+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-19T07:41:04+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Geriatric Medicine","date":"2026-01-17T06:56:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-geriatric-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"EGEM","sideBox":"Learn more about [European Geriatric Medicine](https://www.springer.com/journal/41999)","snPcode":"41999","submissionUrl":"https://www.editorialmanager.com/egem/default2.aspx","title":"European Geriatric Medicine","twitterHandle":"","acdcEnabled":false,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"d6e39c0e-8051-4a1a-84b6-905e5f39f9bd","owner":[],"postedDate":"January 25th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-03-16T07:29:34+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-25 10:21:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8603410","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8603410","identity":"rs-8603410","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}