Transformer-based multimodal model for estimation of appendicular lean mass using incomplete chest radiographs and electronic health record

Transformer-based multimodal model for estimation of appendicular lean mass using incomplete chest radiographs and electronic health record | 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 Transformer-based multimodal model for estimation of appendicular lean mass using incomplete chest radiographs and electronic health record Kosuke Kita, Yuki Suzuki, Takashi Fujimoto, Keisuke Uemura, Yoshito Otake, and 10 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8556884/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 02 Apr, 2026 Read the published version in Journal of Translational Medicine → Version 1 posted 4 You are reading this latest preprint version Abstract Background Sarcopenia is a muscle disease that increases the risk of falls, fractures, and mortality. Appendicular lean mass (ALM) assessment is central to its diagnosis, but standard methods like dual-energy X-ray absorptiometry (DXA) have accessibility and cost issues. Previous artificial intelligence (AI) studies for sarcopenia assessment have been limited to single modalities and have not adequately addressed missing data modalities. This study aimed to develop and validate a multimodal AI model using frontal and lateral chest radiographs and electronic health record (EHR) data to estimate ALM and detect low muscle mass, and to investigate the robustness of a transformer-based algorithm to missing modalities. Methods This model development and validation study adhered to TRIPOD guidelines. The derivation cohort included 3,295 observations from 1,524 participants (mean age 61.9 years, 30% female). External validation was performed on an independent cohort of 3,771 observations from 1,976 participants (mean age 58.4 years, 59% female). Our multimodal model uses a transformer-based TabPFN to predict ALM from 75 features, which integrate 18 features extracted from each of the frontal and lateral chest radiographs by TorchXray with 39 features from EHR data (demographics and blood tests). Model performance for ALM estimation was evaluated using the root mean square error (RMSE), and mean absolute error (MAE), and Pearson correlation coefficient. Estimation of appendicular lean mass performance was assessed using the area under the receiver operating characteristic curve (AUROC). Results The prevalence of low muscle mass was 30% in the derivation set and 29% in the external validation set. The multimodal model achieved high accuracy in ALM estimation, with an RMSE of 1.23 kg, MAE of 0.96 kg, and correlation coefficient of 0.958 in the internal test set. In external validation, the model yielded an RMSE of 1.87 kg, and MAE of 1.41 kg, and correlation of 0.930. For estimation of appendicular lean mass in the external validation set, the model yielded an AUROC of 0.825 for males and 0.730 for females. Sensitivity analyses demonstrated the model's robustness to missing modalities, maintaining stable prediction errors in participants missing part of the EHR (RMSE: 2.00 kg), frontal (RMSE: 1.99 kg), or lateral (RMSE: 2.08 kg) chest radiographs. Conclusions Our transformer-based multimodal AI model accurately estimates ALM and detects low muscle mass from routinely collected clinical data, outperforming unimodal approaches. The model demonstrated robustness even with missing data modalities, supporting its potential utility as a screening tool in clinical settings where data completeness varies. This approach has the potential to serve as an accessible screening tool for low muscle mass, especially in settings where DXA is not readily available. Sarcopenia Multimodal artificial intelligence Transformer Missing modality Chest radiography Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Sarcopenia is a progressive and systemic muscle disease characterized by age-related decline in skeletal muscle mass and strength, leading to serious health consequences including increased risk of falls, fractures, dependency, and mortality 1 . In aging societies, its rising prevalence represents a significant health issue that diminishes patients' quality of life and consumes healthcare resources 2 . Early detection and diagnosis are crucial, with appendicular lean mass (ALM) assessment playing a central role in diagnostic criteria for sarcopenia as a musculoskeletal disorder 3 . International consensus definitions from EWGSOP2 and AWGS adopt "low muscle strength + low muscle mass" for sarcopenia diagnosis, recommending ALM measurement via dual-energy X-ray absorptiometry (DXA) or bioelectrical impedance analysis (BIA) 2 , 3 . However, DXA's high cost and limited availability, combined with BIA's susceptibility to hydration status and measurement variability, create barriers to widespread clinical implementation 4 . Consequently, there is a need for developing alternative methods that can assess muscle mass accurately and conveniently using existing data. Recent artificial intelligence (AI) studies utilizing existing data for sarcopenia assessment have predominantly focused on single modalities. Examples include deep learning attempts to estimate ALM from frontal chest X-rays for sarcopenia screening 5 and machine learning models detecting sarcopenia patients using structured electronic health record (EHR) data including blood test results 6 . These studies demonstrated the potential for non-invasive muscle mass estimation that was previously difficult to measure. However, there are currently no AI studies utilizing lateral chest X-rays in the sarcopenia field, nor are there any multimodal approaches that integrate different data types such as frontal chest X-rays and EHR data. Additionally, handling missing data modalities in multimodal AI remains inadequately addressed 7 . In previous multimodal analyses in the skeletal muscle field 8 – 10 , cases with missing modalities were excluded, which may introduce bias and prevent accurate evaluation of model performance. To advance precision medicine in sarcopenia and frailty, the importance of multimodal diagnostic approaches combining clinical information, imaging, and biological data has been discussed 11 . Previous skeletal muscle research has reported improved accuracy through integration of multiple modalities compared to single modalities 10 , 12 . Furthermore, transformer-based algorithms have shown potential effectiveness for handling missing modalities in multimodal medical data 13 , suggesting they may achieve sufficient accuracy in multimodal analysis of appendicular lean mass. Therefore, this study aims to develop and validate a multimodal AI model that estimates ALM using frontal and lateral chest X-rays along with EHR data. Furthermore, we will investigate whether transformer-based algorithms remain effective even when dealing with missing modalities. Methods Study design and TRIPOD adherence This study developed and validated a multimodal AI model for estimating ALM and detecting low muscle mass in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) guidelines 14 . Following transparent reporting standards for prediction models, we conducted internal validation on the derivation cohort and external validation on an independent cohort. Study population The derivation cohort included participants who visited the health screening center at Sumitomo Hospital between January 2016 and June 2023. For participants who visited multiple times, all visits meeting the inclusion criteria were included as separate observations, resulting in 3,295 observations from 1,524 unique participants. The study population was defined according to specific inclusion criteria. For the derivation cohort, inclusion criteria were: (1) aged ≥ 20 years, (2) DXA measurement of ALM available (iDXA, GE/Lunar), and (3) at least one of chest X-ray (frontal or lateral) or blood test (complete blood count and biochemistry) performed (Table S1 ). A total of 3,295 observations met these criteria. This study was performed in accordance with the Declaration of Helsinki and received approval from the Institutional Ethics Committee of the University of Osaka Graduate School of Medicine (approval number: 22099). Informed consent was waived because of the retrospective nature of the study. For external validation, data from the University of Osaka Hospital were used. Between January 2012 and April 2023, 2,738 participants underwent DXA (Horizon A, Hologic) for ALM measurement. Chest X-rays (frontal and lateral) and blood tests (complete blood count and biochemistry) performed within 3 months before or after DXA were collected. Applying the same inclusion criteria as the derivation cohort, 1,976 participants remained. For participants who underwent multiple DXA examinations during the study period, all examinations were included as separate observations, resulting in 3,771 observations for the external validation cohort. Multimodal AI model architecture Our multimodal AI model was developed using a two-step approach, as illustrated in Fig. 1 . The model architecture, including the pre-trained CNN, was prespecified and not modified based on the results. Step 1 involved extracting 18 features (Table S2) from each of the frontal and lateral chest X-rays using a pre-trained CNN, resulting in a total of 36 imaging features. We used TorchXray, which was trained on large chest X-ray datasets including PadChest, NIH, CheXpert, and MIMIC 15 . Step 2 combined the 36 features extracted from frontal and lateral chest X-rays with EHR data consisting of demographics (height, weight, sex, age) and blood test results, resulting in a total of 75 features. These combined features were then processed through TabPFN, a transformer-based model for tabular data 16 , to predict ALM. Missing data or modalities were handled by treating them as missing values without imputation, which was pre-planned in our analysis strategy. Model comparison To compare multimodal and unimodal approaches, we developed unimodal models for each modality. For EHR data, we used TabPFN. To process the chest X-ray images, we employed two models: VGG16, and an alternative model that used TabPFN to analyze features extracted by TorchXray. Additionally, we evaluated the performance of the transformer-based TabPFN against existing machine learning models by replacing TabPFN in Step 2 of our multimodal model with other machine learning models. We used PyCaret (version 3.3.2) to validate the accuracy of these alternative machine learning models. Data preprocessing For each participant’s raw Digital Imaging and Communications in Medicine (DICOM) image, pixel values were extracted using the pydicom library (version 2.4.4) in Python. Images were resized to 224 × 224 pixels and processed by the TorchXray model. For the unimodal VGG16 model, contrast-limited adaptive histogram equalization was additionally applied to increase image contrast. Low muscle mass was defined according to AWGS criteria 3 based on the skeletal muscle index (SMI), which was calculated as ALM (kg) divided by height squared (m²). Low SMI was defined as < 7.0 kg/m² in males and < 5.4 kg/m² in females. For EHR data, no imputation was performed for missing values. Model validation Model performance was evaluated using predefined metrics. For the regression task of ALM prediction, we assessed model performance using root mean square error (RMSE) in kg, mean absolute error (MAE) in kg, and Pearson correlation coefficient between predicted and actual ALM values. These metrics were used to compare the performance of TabPFN-multimodal model with unimodal models (TabPFN-EHR, VGG16-chest X-ray, and TabPFN-chest X-ray) as well as to compare TabPFN with existing machine learning models in the multimodal framework (step 2 of the model architecture). For classification tasks of low muscle mass detection based on AWGS criteria, model performance was evaluated using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), sensitivity, specificity, and F1 score. Internal validation was performed using an 80:20 split of the derivation dataset, with further 80:20 splitting for training and validation. External validation was conducted on an independent cohort from the University of Osaka Hospital to assess model transportability. Model discrimination and agreement for the TabPFN-multimodal model were further visualized using Pearson correlation plots and Bland-Altman plots with 95% limits of agreement. To ensure the TabPFN-multimodal model's robustness in the external validation set, we conducted subgroup analyses by age and sex categories to assess model performance across different populations. Additionally, sensitivity analyses were performed to evaluate model performance in participant subsets with incomplete data: participants missing at least one EHR parameter (blood test or demographic data), participants missing frontal chest X-ray, and participants missing lateral chest X-ray. Model interpretability analysis To enhance model transparency and clinical applicability, we performed interpretability analyses on our models. For the unimodal VGG16 model that processes chest X-ray images, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize and understand the regions in chest X-ray images that the model focused on for feature extraction 17 . For the TabPFN-based multimodal model, SHAP (SHapley Additive exPlanations) scores were calculated to interpret the model output and identify the contribution of each input feature to ALM prediction. These analyses provide insights into which anatomical regions in chest X-rays are important for the VGG16 model and which features from both chest X-rays and EHR data most significantly influence the multimodal model's predictions. Hyperparameters For the training of the unimodal VGG16 model, data augmentation was performed on the training images using random shift, rotation, horizontal flip, zoom, brightness, and random multiplication five times. The model was trained for 100 epochs with early stopping (patience = 15), a batch size of 32, and a cosine annealing learning rate. The Adam optimizer was used to minimize the Gaussian loss function (negative log-likelihood loss following Gaussian distributions). Statistical analysis All statistical methods were prespecified to ensure transparent reporting. Baseline characteristics were summarized for both derivation and external validation cohorts. Continuous variables were presented as mean ± standard deviation based on their distribution and compared using the independent two-sample t-test or Wilcoxon rank-sum test as appropriate. Categorical variables were presented as number (percentage) and compared using chi-square test or Fisher's exact test between groups with or without low muscle mass. To derive the 95% confidence intervals of model performance metrics in the internal test and external validation sets, the bootstrap method with 2000 repetitions was used. All statistical analyses were performed using Scikit-Learn (version 0.24.2) and SciPy (version 1.10.1). A two-sided P-value < 0.05 was considered statistically significant. Results Clinical characteristics The mean age of the derivation cohort (n = 3,295 observations from 1,524 participants, females 30%) and the external validation cohort (n = 3,771 observations from 1,976 participants, females 59%) was 61.9 and 58.4 years, respectively. The prevalence of low muscle mass was 30% in the derivation set and 29% in the external validation set (Table 1 ). In both cohorts, individuals with low muscle mass had lower weight and ALM (P < 0.001 for both). The proportion of females was higher in the low muscle mass group in the derivation set but lower in the external validation set (P < 0.001 for both cohorts). Table 1 Clinical characteristics of study subjects Derivation set External validation set Low muscle mass (n = 991, 30%) No low muscle mass (n = 2304, 70%) P value Low muscle mass (n = 1107, 29%) No low muscle mass (n = 2664, 71%) P value Age, year 64.6 ± 12.6 60.8 ± 11.8 < 0.001 57.0 ± 19.0 59.0 ± 15.8 0.087 Female, n (%) 334 (34) 634 (28) < 0.001 511 (46) 1699 (64) < 0.001 Weight, kg 56.8 ± 9.6 70.0 ± 12.7 < 0.001 49.9 ± 9.9 60.3 ± 14.0 < 0.001 Height, m 1.64 ± 0.08 1.66 ± 0.09 < 0.001 1.59 ± 0.11 1.58 ± 0.10 < 0.001 ALM, kg 16.4 ± 3.4 20.8 ± 4.4 < 0.001 14.4 ± 3.5 17.9 ± 4.7 < 0.001 ALM, appendicular lean mass Comparing the multimodal model with unimodal models: Multiple models were developed to predict ALM from EHR, frontal or lateral chest X-rays, or these three modalities integrated (Table 2 ). Among all these models, the TabPFN-multimodal model combining chest X-ray features with EHR data achieved the highest performance. In the internal test set, the multimodal model achieved prediction errors of RMSE 1.23 kg (95% CI: 1.18–1.29) and MAE 0.96 kg (0.91–1.02), with a correlation coefficient of 0.958 (0.951–0.964). In the external validation set, the model yielded an RMSE of 1.87 kg (1.83–1.91) and MAE of 1.41 kg (1.38–1.45), with a correlation coefficient of 0.930 (0.925–0.935). The correlation and agreement between predicted and measured ALM values for the multimodal model are shown in Fig. 2 . Bland-Altman analysis revealed minimal systematic bias in both datasets. In the internal test set, the mean difference was 0.07 kg, with 95% limits of agreement ranging from − 2.35 to 2.48 kg. In the external validation set, the model maintained a low mean bias of -0.22 kg, although the 95% limits of agreement widened slightly to -3.52 to 3.08 kg. The plots showed no distinct proportional bias across the range of ALM values. Table 2 Performance comparison of multimodal and unimodal models for predicting appendicular lean mass Model - Modality Internal validation External validation RMSE, kg MAE, kg Correlation Coefficient RMSE, kg MAE, kg Correlation Coefficient VGG16 - AP* 3.09 (2.96–3.22) 2.43 (2.30–2.56) 0.840 (0.819–0.859) 4.97 (4.87–5.07) 4.23 (4.13–4.34) 0.578 (0.552–0.604) VGG16 - LAT* 3.13 (3.00–3.25) 2.49 (2.36–2.62) 0.807 (0.782–0.829) 4.22 (4.08–4.37) 3.18 (3.03–3.32) 0.438 (0.395–0.479) TabPFN - AP* 3.06 (2.91–3.21) 2.38 (2.23–2.53) 0.718 (0.678–0.754) 4.08 (3.98–4.18) 3.03 (2.92–3.13) 0.493 (0.463–0.521) TabPFN - LAT* 2.90 (2.76–3.04) 2.30 (2.16–2.44) 0.745 (0.707–0.778) 3.45 (3.38–3.52) 2.71 (2.64–2.78) 0.692 (0.675–0.708) TabPFN - EHR 1.43(1.36–1.50) 1.11 (1.04–1.18) 0.946 (0.937–0.954) 2.02 (1.93–2.11) 1.57 (1.53–1.61) 0.920 (0.915–0.925) TabPFN - Multimodal 1.23 (1.18–1.29) 0.96 (0.91–1.02) 0.958 (0.951–0.964) 1.87 (1.83–1.91) 1.41 (1.38–1.45) 0.930 (0.925–0.935) * Performance was measured exclusively on cases with available chest X-ray images AP, frontal chest X-rays; LAT, lateral chest X-rays; EHR, electronic health record; RMSE, root mean square error; MAE, mean absolute error. Values in parentheses are 95% confidence intervals, calculated using the bootstrap method. The TabPFN-EHR model, using EHR data alone, achieved an RMSE of 1.43 kg (95% CI: 1.36–1.50) and MAE of 1.11 kg (1.04–1.18) in the internal test set, while maintaining robust performance in the external validation set with an RMSE of 2.02 kg (1.93–2.11) and MAE of 1.57 kg (1.53–1.61). The correlation coefficients were 0.946 (0.937–0.954) and 0.920 (0.915–0.925), respectively. In contrast, models utilizing chest X-ray features alone (VGG16-AP, VGG16-LAT, TabPFN-AP, TabPFN-LAT) exhibited higher prediction errors, with RMSE values ranging from 2.90 to 3.13 kg (MAE: 2.30–2.49 kg) in internal validation and 3.45 to 4.97 kg (MAE: 2.71–4.23 kg) in external validation, corresponding to correlation coefficients of 0.718–0.840 and 0.438–0.692, respectively. Comparing TabPFN with traditional machine learning models To evaluate a range of algorithms for the final prediction stage of our multimodal architecture (Step 2), we compared the performance of the transformer-based TabPFN against several traditional machine learning models for ALM prediction (Table S3). In the internal validation set, TabPFN demonstrated superior performance, achieving the lowest prediction errors with an RMSE of 1.23 kg (95% CI: 1.18–1.29) and MAE of 0.96 kg (0.91–1.02), alongside the highest correlation coefficient of 0.958 (0.951–0.964). The best-performing traditional machine learning model was the Gradient Boosting Regressor, which yielded slightly higher errors with an RMSE of 1.25 kg (1.18–1.32) and MAE of 0.98 kg (0.92–1.04), and a correlation coefficient of 0.957 (0.951–0.962). In the external validation dataset, TabPFN maintained its superiority, recording an RMSE of 1.87 kg (1.83–1.91) and MAE of 1.41 kg (1.38–1.45). This substantially outperformed the best conventional model (Gradient Boosting Regressor), which showed increased prediction errors with an RMSE of 2.07 kg (2.01–2.13) and MAE of 1.59 kg (1.55–1.63), corresponding to a lower correlation coefficient of 0.919 (0.912–0.925). Notably, while TabPFN kept the RMSE below 1.9 kg in external validation, all traditional models exhibited RMSE values exceeding 2.0 kg. These results demonstrate that the transformer-based TabPFN model consistently minimized prediction errors compared to established machine learning approaches across all evaluation metrics and datasets. Performance of multimodal model for low muscle mass prediction The TabPFN-multimodal model predicted low muscle mass based on AWGS criteria (Table 3 ). In the internal test set, sex-specific AUROCs were 0.832 (95% CI: 0.793–0.869) for males and 0.808 (95% CI: 0.747–0.864) for females (Figure S1 ), with sensitivity of 0.766 (95% CI: 0.700-0.821) and specificity of 0.836 (95% CI: 0.801–0.867). As shown in Table 3 , the model demonstrated robust diagnostic performance in the external validation set, with AUROCs of 0.825 (95% CI: 0.804–0.843) for males and 0.730 (95% CI: 0.708–0.754) for females. While the sensitivity was 0.558 (95% CI: 0.534–0.582), the specificity remained high at 0.910 (95% CI: 0.897–0.921). AUPRCs in the internal test set were 0.909 (95% CI: 0.879–0.936) for males and 0.895 (95% CI: 0.846–0.937) for females, while external validation AUPRCs were 0.880 (95% CI: 0.860–0.901) for males and 0.904 (95% CI: 0.891–0.916) for females. Table 3 Performance of the multimodal model for low muscle mass prediction Performance metrics Internal test set 95% CI External test set 95% CI AUROC-Male 0.832 0.793–0.869 0.825 0.804–0.843 AUROC-Female 0.808 0.747–0.864 0.730 0.708–0.754 AUPRC-Male 0.909 0.879–0.936 0.880 0.860–0.901 AUPRC-Female 0.895 0.846–0.937 0.904 0.891–0.916 Sensitivity-AWGS 0.766 0.700–0.821 0.558 0.534–0.582 Specificity-AWGS 0.836 0.801–0.867 0.910 0.897–0.921 F1 score 0.698 0.645–0.744 0.666 0.645–0.685 95% CI was calculated using bootstrapping method. Subgroup and sensitivity analysis Subgroup analysis in the external validation cohort showed RMSE values ranging from 1.62 kg (95% CI: 1.46–1.79 kg) in participants aged ≥ 80 years to 2.17 kg (95% CI: 2.05–2.30 kg) in participants < 50 years (Table 4 ). Performance differed by sex, with males exhibiting higher predictive errors (RMSE: 2.37 kg, 95% CI: 2.29–2.47 kg; MAE: 1.86 kg, 95% CI: 1.79–1.93 kg) compared to females (RMSE: 1.64 kg, 95% CI: 1.58–1.70 kg; MAE: 1.25 kg, 95% CI: 1.21–1.30 kg). Sensitivity analyses in participants with missing modalities showed RMSE of 2.00 kg (95% CI: 1.85–2.15 kg) in 261 participants missing at least one EHR parameter, 1.99 kg (95% CI: 1.92–2.06 kg) in 1,197 participants missing frontal chest X-ray, and 2.08 kg (95% CI: 2.03–2.13 kg) in 2,346 participants missing lateral chest X-ray. Table 4 Subgroup and sensitivity analyses of the multimodal model's performance in the external validation cohort. Analysis Type Subgroup / Missing modality N RMSE, kg MAE, kg Correlation Coefficient Complete case 3771 1.87 (1.83–1.91) 1.41 (1.38–1.45) 0.930 (0.925–0.935) Subgroup Analysis Age: <50 1086 2.17 (2.05–2.30) 1.64 (1.54–1.72) 0.939 (0.930–0.947) Age: 50–64 985 2.03 (1.92–2.14) 1.53 (1.44–1.61) 0.926 (0.915–0.935) Age: 65–79 1459 1.84 (1.76–1.91) 1.43 (1.37–1.50) 0.918 (0.907–0.928) Age: 80< 241 1.62 (1.46–1.79) 1.25 (1.12–1.39) 0.906 (0.876–0.928) Sex: Male 1561 2.37 (2.29–2.47) 1.86 (1.79–1.93) 0.882 (0.867–0.897) Sex: Female 2210 1.64 (1.58–1.70) 1.25 (1.21–1.30) 0.846 (0.828–0.862) Age_Sex: <50_Male 503 2.65 (2.47–2.85) 2.07 (1.92–2.21) 0.903 (0.883–0.922) Age_Sex: <50_Female 583 1.66 (1.54–1.79) 1.27 (1.18–1.36) 0.907 (0.884–0.926) Age_Sex: 50–64_Male 356 2.52 (2.32–2.68) 1.96 (1.80–2.11) 0.830 (0.788–0.863) Age_Sex: 50–64_Female 629 1.69 (1.58–1.81) 1.28 (1.20–1.37) 0.811 (0.780–0.839) Age_Sex: 65–79_Male 602 2.13 (2.01–2.24) 1.71 (1.61–1.81) 0.829 (0.796–0.858) Age_Sex: 65–79_Female 857 1.60 (1.51–1.69) 1.24 (1.18–1.31) 0.763 (0.719–0.805) Age_Sex: 80<_Male 100 1.68 (1.44–1.91) 1.35 (1.15–1.56) 0.868 (0.827–0.904) Age_Sex: 80<_Female 141 1.57 (1.32–1.79) 1.17 (0.988–1.35) 0.808 (0.725–0.871) Sensitivity Analysis EHR 261 2.00 (1.85–2.15) 1.58 (1.43–1.73) 0.951 (0.938–0.962) Frontal chest X-ray 1197 1.99 (1.92–2.06) 1.54 (1.47–1.61) 0.932 (0.924–0.939) Lateral chest X-ray 2346 2.08 (2.03–2.13) 1.60 (1.55–1.66) 0.923 (0.917–0.929) Values in parentheses are 95% confidence intervals, calculated using the bootstrap method. Model Interpretability Analysis: The Grad-CAM analysis for the VGG16 models predicting ALM is shown in Fig. 3 . Saliency maps were generated separately for frontal and lateral chest X-rays from both the internal test set and external validation set, and each map was averaged across all subjects to produce representative saliency images for each view. The averaged Grad-CAM visualizations demonstrated that the models consistently highlighted specific anatomical regions. For frontal chest X-rays, the internal validation model showed prominent activation around both upper arm roots (proximal humerus regions), while the external validation model focused on both upper arm roots and additionally highlighted the right lung field. For lateral chest X-rays, both internal and external validation models consistently focused on the thoracic spine region. These findings suggest that the VGG16 models utilize anatomically relevant features, particularly musculoskeletal landmarks and surrounding structures, for ALM prediction. SHAP values for the TabPFN in the multimodal model on the external validation set are illustrated in Fig. 4 . The feature importance analysis revealed that demographic features were the most important for ALM predictions, with weight, sex, height, and age ranking as the top four, respectively. Among blood test parameters, creatinine and LDH were the most important factors, ranking fifth and sixth overall, respectively. Following these top six features, the 'Atelectasis' feature from frontal chest X-rays contributed to the predictions. The SHAP analysis demonstrated distinct directional relationships: weight, sex (male: 1, female: 0), height, and creatinine showed a positive correlation with ALM predictions (higher values associated with higher predicted ALM), while age exhibited negative correlations (higher values associated with lower predicted ALM). Discussion Summary of key findings This study developed and validated a multimodal AI model that combines chest X-ray images (frontal and lateral views) with EHR data (blood test and demographics) to predict ALM and detect low muscle mass. The TabPFN-based multimodal model achieved superior performance compared to unimodal approaches, with an RMSE of 1.23 kg (95% CI: 1.18–1.29) and MAE of 0.96 kg (0.91–1.02) in internal validation, and an RMSE of 1.87 kg (1.83–1.91) and MAE of 1.41 kg (1.38–1.45) in external validation. For low muscle mass detection, the model demonstrated performance with AUPRCs of 0.88 for males and 0.90 for females in external validation. Comparative Performance against Prior Studies and Clinical Benchmarks The superior performance of the multimodal approach over unimodal models aligns withprior studies in medical AI research demonstrating that integration of multiple data modalities enhances predictive accuracy in complex clinical tasks 10 , 11 , 18 . The multimodal model achieved reduced prediction errors compared to the EHR-only model. Specifically, in the external validation set, the multimodal model yielded a lower RMSE of 1.87 kg compared to 2.02 kg for the EHR-only model, demonstrating the added value of integrating chest X-rays. Our unimodal chest X-ray models showed mixed results compared to previous work, with strong internal validation performance (correlation coefficient 0.840) but weaker external validation (correlation coefficient 0.578) compared to previous chest X-ray-based models for ALM prediction, which reported concordance correlation coefficients of 0.80 and 0.76 in internal and external validation, respectively 5 . This performance drop in external validation highlights the challenge of generalizability with single-modality approaches, and it appears more pronounced than that observed in the previous work 5 . A direct comparison with the prior study is challenging, as it does not specify the X-ray acquisition protocols or equipment used. The performance of an AI model can be sensitive to such variations in imaging technique, and it is possible that these unclarified technical differences contributed to the performance degradation observed in our external validation. In contrast, our EHR-only model demonstrated robust performance for direct ALM prediction in both internal and external validation settings (correlation coefficients of 0.946 and 0.920, respectively.) This extends previous machine learning research on EHR data for low muscle mass detection 6 , which showed promising results for identifying patients meeting AWGS criteria using a logistic regression model (AUROC values between 0.72–0.91), but was limited to internal validation only. Our study's external validation performance demonstrates that EHR data alone can reliably predict ALM and potentially identify low muscle mass across different clinical settings, addressing a key limitation in previous research where generalizability remained uncertain. Both the previous and our studies highlight the potential value of EHR-based screening approaches for early detection of low muscle mass in clinical practice, using routinely collected healthcare data. Our multimodal approach demonstrates accuracy comparable to BIA reports when using DXA as the reference standard. While BIA typically overestimates ALM by 1.97 kg compared to DXA measurements, our TabPFN-based multimodal model yielded a MAE of 1.41 kg (95% CI: 1.38–1.45) in external validation. These findings suggest that our model offers a viable alternative to BIA, a traditional method recommended in guidelines, while offering the practical advantage of using routinely available clinical data. Our Bland-Altman analysis (Fig. 2 ) further confirmed that the multimodal model is reliable for regression tasks, showing negligible fixed bias even in the external cohort. While the 95% limits of agreement in external validation (approximately ± 3.5 kg) were wider than in internal validation, this range is comparable to the reported error margins between BIA 19 . Although some outliers exceeded these limits—likely due to image artifacts or extreme body compositions inherent in real-world clinical data—the majority of predictions fell within a clinically acceptable range. These findings support the model's robustness and its potential utility as a screening tool, where minimizing systematic bias is crucial for identifying individuals at risk. Performance Differences by Sex In our external validation, we observed a divergence in performance metrics between sexes: the model achieved a higher AUROC for detecting low muscle mass in males (0.825) compared to females (0.730), whereas the estimation error (RMSE) for absolute ALM values was larger in males (2.37 kg) than in females (1.64 kg) (Tables 3 and 4 ). This discrepancy in RMSE is likely attributable to the physiological difference in baseline muscle mass; males typically possess greater total muscle mass, resulting in a wider range of values and consequently larger absolute errors, a phenomenon consistent with scaling effects. Conversely, the superior discriminative ability (AUROC) in males aligns with findings from previous AI studies in sarcopenia 20 . This may be because the age-related decline in muscle mass is often more pronounced and distinct in males, making the differentiation between normal and low muscle mass relatively clearer for the model. Importantly, our SHAP analysis identified sex as the second most important feature, suggesting that the multimodal model successfully learned and adjusted for these physiological sex-based differences within a single architecture. Model Interpretability and Technical Advantages Analysis of SHAP values revealed that the top predictors were six EHR-derived features—weight, sex, height, age, creatinine, and LDH, in that order—followed by the 'Atelectasis' feature from frontal chest X-rays (Fig. 4 ). SHAP analysis demonstrated a positive correlation between creatinine levels and predicted ALM (higher creatinine associated with higher predicted ALM). This finding is consistent with the fact that creatinine is a byproduct of muscle metabolism, and its serum levels are directly influenced by an individual's skeletal muscle mass 21 , 22 . The prominence of demographics (weight, sex, height, and age) as top predictors underscores their fundamental importance in determining muscle mass, while the contribution of chest X-ray features suggests that imaging provides complementary information about thoracic anatomy that relates to overall body composition. Although the Grad-CAM visualization was derived from the unimodal VGG16 model rather than the final transformer-based architecture, the consistent focus on anatomically relevant regions—specifically the proximal humerus and thoracic spine—validates that the deep learning algorithms are leveraging biologically plausible signals from the chest radiographs. This supports the clinical interpretability of the imaging features integrated into our multimodal framework. The TabPFN-based multimodal model consistently outperformed traditional machine learning approaches across all evaluation metrics (Table S3). While several conventional models showed comparable performance in the internal validation, TabPFN maintained its superior predictive capability in the external validation dataset where traditional methods exhibited substantial performance degradation. This finding aligns with research demonstrating that transformer-based models excel on tabular datasets, particularly for smaller datasets with fewer than 10,000 samples, compared to conventional approaches 16 , 23 . Notably, our sensitivity analysis revealed that the TabPFN-multimodal model maintained robust performance even with missing data modalities. The model demonstrated stable prediction errors even when handling patients with missing EHR parameters (RMSE: 2.00 kg), missing frontal chest X-rays (RMSE: 1.99 kg), or missing lateral chest X-rays (RMSE: 2.08 kg). This excellent performance with incomplete data supports recent findings that transformer-based architectures effectively manage missing healthcare data through their self-attention mechanisms 24 . Such robustness to missing modalities is particularly valuable for clinical implementation, where data completeness varies significantly across patients and institutions, potentially eliminating the need for complex imputation strategies while maintaining high predictive accuracy. Clinical Significance and Practicality The multimodal AI model developed in this study demonstrated the potential to accurately estimate ALM and detect low muscle mass from readily available clinical data, including chest X-ray images, blood test results, and basic anthropometric information without requiring special examinations. This could be a powerful solution to the current challenges where DXA, the gold standard for measuring ALM, is expensive and has limited availability. Specifically, by utilizing data from regular health check-ups or routine inpatient examinations, this model is expected to serve as a gatekeeper, efficiently screening patients at potential risk of sarcopenia and facilitating early intervention or more detailed examinations. For example, integrating this model into EHR systems to automatically present estimated ALM values and the likelihood of low muscle mass, when relevant test data are available would enable physicians to quickly recognize sarcopenia risk and promote appropriate nutritional guidance, exercise therapy, or referrals to specialists. Particularly in regions or facilities where access to DXA scans is difficult, this model could contribute to improving the quality of sarcopenia care and reducing disparities in medical access. Furthermore, the model's robust performance even with missing data is a crucial factor enhancing its practicality, as stable operation can be expected under diverse data acquisition conditions in real-world clinical practice. However, from a practical standpoint, the trade-off between maximal accuracy and implementation feasibility must be considered. While the multimodal model achieved the highest predictive accuracy, the performance gain over the EHR-only model was incremental. This suggests that the EHR-only model may serve as a sufficient and scalable screening tool in resource-limited settings. Nevertheless, the multimodal approach offers additional value in terms of robustness against missing biochemical data and provides complementary anatomical information, serving as a comprehensive gatekeeper for sarcopenia risk assessment. As a tangible implementation of this approach, the web application we have developed (available at: https://alm-application.onrender.com ) validates the value of the EHR-only model as a tool that clinicians can use immediately without requiring special infrastructure. A critical aspect for optimizing this model as a screening tool is the adjustment of its diagnostic threshold. In the external validation, the model demonstrated high specificity (0.910) but a moderate sensitivity of 0.558. This sensitivity was observed at the fixed diagnostic cut-off values defined by the AWGS criteria, which are designed for confirmatory diagnosis rather than screening. Given the model's acceptable discriminative performance (AUROC > 0.80 for males and > 0.73 for females), the sensitivity can be readily enhanced for screening purposes by adjusting the decision threshold (e.g., using the Youden Index) to prioritize sensitivity over specificity. This flexibility allows the model to be tailored to specific clinical needs, where identifying potential at-risk individuals is prioritized over minimizing false positives. Limitations This study has several limitations. First, the study population may have a selection bias. Although the derivation cohort consisted of participants from a single health screening center and the external validation cohort used data from a different university hospital, these populations may not necessarily represent the general population or patient groups with diverse disease backgrounds. However, the model's demonstrated accuracy in external validation across two potentially distinct cohorts suggests a degree of generalizability. Further multicenter collaborative research is desirable to validate the model's performance in broader medical settings and among diverse racial and ethnic groups. Second, a consideration for the future application of our model concerns its scalability. In this study, we employed TabPFN, which demonstrated excellent performance on our dataset sizes (derivation: 3,295; external validation: 3,771). This aligns with reports that TabPFN performs particularly well on tabular data with fewer than 10,000 samples 25 . However, the scalability of this specific approach could be a limitation when applying it to much larger datasets, such as nationwide registries or biobanks with tens of thousands of observations. In such scenarios, the computational efficiency and performance might not be maintained. Therefore, future work to scale our multimodal framework would necessitate exploring alternative or supplementary methods, such as ensembling TabPFN with gradient boosting models like LightGBM, or adopting other transformer-based architectures designed for large-scale data. Conclusion This study developed and validated a novel multimodal AI model using transformer-based TabPFN that accurately estimates ALM and detects low muscle mass from routinely collected clinical data (chest X-rays, blood tests, demographics), achieving superior performance over unimodal approaches. The model demonstrated robust performance even with missing data modalities, supporting its potential utility as a screening tool in clinical settings where data completeness varies. These findings underscore the potential of this multimodal AI approach as an accessible screening tool for early low muscle mass identification, particularly in settings where DXA availability is limited, facilitating timely interventions and improved patient outcomes. Abbreviations ALM: Appendicular Lean Mass; DXA: Dual-energy X-ray Absorptiometry; EHR: Electronic Health Record; AI: Artificial Intelligence; RMSE: Root Mean Square Error; MAE: Mean Absolute Error; AUROC: Area Under the Receiver Operating Characteristic Curve; SMI: Skeletal Muscle Index; CNN: Convolutional Neural Network; Grad-CAM: Gradient-weighted Class Activation Mapping; SHAP: SHapley Additive exPlanations. Declarations Ethics approval and consent to participate This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Ethics Committee of the University of Osaka Graduate School of Medicine (approval number: 22099). Informed consent was waived due to the retrospective nature of the study. Consent for publication Not applicable. Availability of data and materials The datasets generated and/or analysed during the current study are not publicly available due to patient privacy protocols but are available from the corresponding author on reasonable request. Competing interests The authors declare that they have no competing interests. Funding This work was supported by Japan Osteoporosis Foundation, and JSPS KAKENHI Grant Numbers JP21K20966, JP24KJ1559. Authors' contributions KK and TF conceived and designed the study. KK, YS, and TF performed the data analysis. KK wrote the manuscript. TF supervised the project. All authors read and approved the final manuscript. Acknowledgements Not applicable. References Cruz-Jentoft AJ, Sayer AA, Sarcopenia. Lancet. 2019;393:2636–46. Cruz-Jentoft AJ, Bahat G, Bauer J, Boirie Y, Bruyère O, Cederholm T, et al. Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing. 2019;48:16–31. Chen L-K, Woo J, Assantachai P, Auyeung T-W, Chou M-Y, Iijima K, et al. Asian Working Group for Sarcopenia: 2019 Consensus Update on Sarcopenia Diagnosis and Treatment. J Am Med Dir Assoc. 2020;21:300–e3072. Lee SY, Ahn S, Kim YJ, Ji MJ, Kim KM, Choi SH, et al. Comparison between Dual-Energy X-ray Absorptiometry and Bioelectrical Impedance Analyses for Accuracy in Measuring Whole Body Muscle Mass and Appendicular Skeletal Muscle Mass. Nutrients. 2018;10:738. Ryu J, Eom S, Kim HC, Kim CO, Rhee Y, You SC, et al. Chest X-ray-based opportunistic screening of sarcopenia using deep learning. J Cachexia Sarcopenia Muscle. 2023;14:418–28. Luo X, Ding H, Broyles A, Warden SJ, Moorthi RN, Imel EA. Using machine learning to detect sarcopenia from electronic health records. Digit HEALTH. 2023;9:20552076231197098. Ghassemi M, Naumann T, Schulam P, Beam AL, Chen IY, Ranganath R. Practical guidance on artificial intelligence for health-care data. Lancet Digit Health. 2019;1:e157–9. Maki S, Furuya T, Katsumi K, Nakajima H, Honjoh K, Watanabe S et al. Multimodal Deep Learning-based Radiomics Approach for Predicting Surgical Outcomes in Patients with Cervical Ossification of the Posterior Longitudinal Ligament. Spine. 2024;49:1561–1569. Kita K, Fujimori T, Suzuki Y, Okada S, Kido S. Bi-modal network combining convolutional neural network and TabNet, differentiating spinal tumors based on images and clinical risk factors. In: Iftekharuddin KM, Chen W, editors. Medical Imaging 2023: Computer-Aided Diagnosis. SPIE: San Diego, United States; 2023. p. 4. Kita K, Fujimori T, Suzuki Y, Kanie Y, Takenaka S, Kaito T, et al. Bimodal artificial intelligence using TabNet for differentiating spinal cord tumors-Integration of patient background information and images. iScience. 2023;26:107900. Lynch DH, Spangler HB, Franz JR, Krupenevich RL, Kim H, Nissman D, et al. Multimodal Diagnostic Approaches to Advance Precision Medicine in Sarcopenia and Frailty. Nutrients. 2022;14:1384. Kline A, Wang H, Li Y, Dennis S, Hutch M, Xu Z et al. Multimodal machine learning in precision health: A scoping review. npj Digit Med 2022;5. Wang K, Wang Y, Ren R, Zou H, Shao Z. Transformer-Based Feature Aggregation and Stitching Network for Crowd Counting. IEEE Access. 2023;11:124833–44. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024;e078378. Cohen JP, Viviano JD, Bertin P, Morrison P, Torabian P, Guarrera M et al. TorchXRayVision: A library of chest X-ray datasets and models. 2021 doi: 10.48550/arXiv.2111.00595 Hollmann N, Müller S, Eggensperger K, Hutter F. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. 2023 doi: 10.48550/arXiv.2207.01848 Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). IEEE: Venice; 2017. pp. 618–626. Yoo HJ, Kim YJ, Hong H, Hong SH, Chae HD, Choi J-Y. Deep learning–based fully automated body composition analysis of thigh CT: comparison with DXA measurement. Eur Radiol. 2022;32:7601–11. Nickerson BS, McLester CN, McLester JR, Kliszczewicz BM. Agreement Between 2 Segmental Bioimpedance Devices, BOD POD, and DXA in Obese Adults. J Clin Densitometry. 2020;23:138–48. Yin L, Zhao J. An Artificial Intelligence Approach for Test-Free Identification of Sarcopenia. J Cachexia Sarcopenia Muscle. 2024;15:2765–80. Yim J, Son N-H, Kyong T, Park Y, Kim J-H. Muscle mass has a greater impact on serum creatinine levels in older males than in females. Heliyon. 2023;9:e21866. Thongprayoon C, Cheungpasitporn W, Kashani K. Serum creatinine level, a surrogate of muscle mass, predicts mortality in critically ill patients. J Thorac Dis. 2016;8:E305–311. Tran VQ, Byeon H. Predicting dementia in Parkinson’s disease on a small tabular dataset using hybrid LightGBM–TabPFN and SHAP. Digit HEALTH 2024;10. Casella M, Milano N, Dolce P, Marocco D. Transformers deep learning models for missing data imputation: an application of the ReMasker model on a psychometric scale. Front Psychol. 2024;15:1449272. Hollmann N, Müller S, Purucker L, Krishnakumar A, Körfer M, Hoo SB, et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637:319–26. Supplementary Files Supportinginformation20260106.docx Cite Share Download PDF Status: Published Journal Publication published 02 Apr, 2026 Read the published version in Journal of Translational Medicine → Version 1 posted Reviewers agreed at journal 19 Jan, 2026 Reviewers invited by journal 19 Jan, 2026 Editor assigned by journal 16 Jan, 2026 First submitted to journal 13 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-8556884","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":576924469,"identity":"4105bc74-2e7e-4d43-9071-1d116677d69f","order_by":0,"name":"Kosuke Kita","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Kosuke","middleName":"","lastName":"Kita","suffix":""},{"id":576924470,"identity":"b52aff85-0186-4d33-9474-5c112764016c","order_by":1,"name":"Yuki Suzuki","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Yuki","middleName":"","lastName":"Suzuki","suffix":""},{"id":576924471,"identity":"4566db75-1ac7-4d6e-bdfb-e7887c9fe3b7","order_by":2,"name":"Takashi Fujimoto","email":"","orcid":"","institution":"Sumitomo Hospital","correspondingAuthor":false,"prefix":"","firstName":"Takashi","middleName":"","lastName":"Fujimoto","suffix":""},{"id":576924472,"identity":"36fc0ab9-df64-4fd9-b167-134018df91f0","order_by":3,"name":"Keisuke Uemura","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Keisuke","middleName":"","lastName":"Uemura","suffix":""},{"id":576924473,"identity":"a0673685-82d3-4c06-93d4-ad626947f5f9","order_by":4,"name":"Yoshito Otake","email":"","orcid":"","institution":"Nara Institute of Science and Technology: Nara Sentan Kagaku Gijutsu Daigakuin Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Yoshito","middleName":"","lastName":"Otake","suffix":""},{"id":576924474,"identity":"4b41b333-85df-481b-810f-3254030f34c9","order_by":5,"name":"Masayuki Furuya","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Masayuki","middleName":"","lastName":"Furuya","suffix":""},{"id":576924475,"identity":"09590a00-77e0-41e3-b998-7e82e59fa46a","order_by":6,"name":"Yuya Kanie","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Yuya","middleName":"","lastName":"Kanie","suffix":""},{"id":576924476,"identity":"f0461533-4583-48e6-b491-8eadc18edbd6","order_by":7,"name":"Tomohiro Wataya","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Tomohiro","middleName":"","lastName":"Wataya","suffix":""},{"id":576924477,"identity":"4f0006ad-1b2b-4a4e-9f6a-8ce872a321b7","order_by":8,"name":"Daiki Nishigaki","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Daiki","middleName":"","lastName":"Nishigaki","suffix":""},{"id":576924478,"identity":"a1ca7ad4-e8f6-426e-a12d-d9d16f2cda74","order_by":9,"name":"Junya Sato","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Junya","middleName":"","lastName":"Sato","suffix":""},{"id":576924479,"identity":"61bd3ffb-df45-4cd2-b5ac-dd1214399a3b","order_by":10,"name":"Miyuki Tomiyama","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Miyuki","middleName":"","lastName":"Tomiyama","suffix":""},{"id":576924480,"identity":"29030b34-d653-44d0-91f8-5be8cf48b3a5","order_by":11,"name":"Noriyuki Tomiyama","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Noriyuki","middleName":"","lastName":"Tomiyama","suffix":""},{"id":576924481,"identity":"f4ebcc93-6fb5-40bd-afc5-b3900b0f506e","order_by":12,"name":"Seiji Okada","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Seiji","middleName":"","lastName":"Okada","suffix":""},{"id":576924482,"identity":"1ee63ff3-0c2c-44ce-92c8-c8745140ccf9","order_by":13,"name":"Masatoshi Hori","email":"","orcid":"","institution":"Osaka University: Osaka Daigaku","correspondingAuthor":false,"prefix":"","firstName":"Masatoshi","middleName":"","lastName":"Hori","suffix":""},{"id":576924483,"identity":"10e6b04f-0fca-42c1-97d6-82bd827e7def","order_by":14,"name":"Takahito Fujimori","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0002-2758-7919","institution":"Osaka University School of Medicine Graduate School of Medicine: Osaka Daigaku Daigakuin Igakukei Kenkyuka Igakubu","correspondingAuthor":true,"prefix":"","firstName":"Takahito","middleName":"","lastName":"Fujimori","suffix":""}],"badges":[],"createdAt":"2026-01-09 05:07:53","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8556884/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8556884/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12967-026-08079-0","type":"published","date":"2026-04-02T15:59:30+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":100857521,"identity":"9713d808-38e5-4676-a77a-6f12010fad22","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":530524,"visible":true,"origin":"","legend":"","description":"","filename":"Figures.docx","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/1433e40e90c803bffc400e3c.docx"},{"id":101751095,"identity":"059be9ba-9b77-4a48-8d0d-9070cb3dfd9b","added_by":"auto","created_at":"2026-02-03 10:13:35","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":25670,"visible":true,"origin":"","legend":"","description":"","filename":"Tables20260109.docx","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/c4f1c6582aee1ab4a526e185.docx"},{"id":100857522,"identity":"0bc1c8d4-de72-4135-b937-5e7e808bae11","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":18662,"visible":true,"origin":"","legend":"","description":"","filename":"jtrmJTRMD2600531.xml","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/d949a5824cd33c6170f3e02c.xml"},{"id":100857515,"identity":"1ac0c3c8-6053-4156-a4dd-89f50b80bf72","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"xml","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1073,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD2600531168407.go.xml","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/b3f4b4d443b193c158ae8d1e.xml"},{"id":100857519,"identity":"830b1f0d-917d-4c63-a876-13708310512b","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"xml","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":937,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD2600531Import.xml","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/1e2c3786d6c6803c54ebe6bc.xml"},{"id":101397572,"identity":"a6530ac8-b3bc-4217-93bf-021e504f6ec3","added_by":"auto","created_at":"2026-01-29 09:29:23","extension":"xml","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":117198,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD26005310enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/55730ec63754697f3ad7036e.xml"},{"id":101397573,"identity":"8671e0b2-4494-40cc-855d-aa8854066f70","added_by":"auto","created_at":"2026-01-29 09:29:32","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":100369,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/49671467d960d43cefe3b527.png"},{"id":100857520,"identity":"fa59aaad-5683-43e4-9c11-7a965ec19d24","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":184066,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/03e7ab6c2dfc905089a44b88.png"},{"id":101202394,"identity":"e500c42b-f1f0-43c0-8366-24a35c7ebfaf","added_by":"auto","created_at":"2026-01-27 09:29:53","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130121,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/1664d89d3d0b20372affb6f6.jpeg"},{"id":100857529,"identity":"40e841ff-c9c6-4909-8eac-68f1c7296848","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98436,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/a3f5b2facda37a0ab08c2530.png"},{"id":100857526,"identity":"3f071f24-fc1e-4713-a31a-a3aa35e90ce1","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24527,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/f0d6b1e1c728ff846893ba74.png"},{"id":100949267,"identity":"1d2cb451-6ea3-454d-86e4-169b815dea94","added_by":"auto","created_at":"2026-01-23 06:56:42","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":45488,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/e032a0ca61941509ab828c38.png"},{"id":100857523,"identity":"c7b07ad3-d0d5-4bbd-b8e0-171f206542cb","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":57835,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/f4ade6c7045161744f49b44d.png"},{"id":100857528,"identity":"6949642c-325e-428d-8032-3b2050d9487c","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":22304,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/e5fade661f5cdf635dab0335.png"},{"id":100857524,"identity":"237b3de6-8aa3-4489-b97e-c9c5055be6a8","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"xml","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114035,"visible":true,"origin":"","legend":"","description":"","filename":"JTRMD26005310structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/55dda7e45dc91e5c63e44656.xml"},{"id":101202392,"identity":"97e88d09-113d-49ff-9ea9-17ca5cb14337","added_by":"auto","created_at":"2026-01-27 09:29:51","extension":"html","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123605,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/6dc1fc4edc3dac76493aa633.html"},{"id":100857511,"identity":"5c67e652-5511-4572-8f88-ed5458c2448e","added_by":"auto","created_at":"2026-01-22 07:15:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":92956,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the multimodal AI model for appendicular lean mass (ALM) estimation. The model uses a two-step approach. In Step 1, a pre-trained convolutional neural network (TorchXray) extracts 18 features each from frontal and lateral chest X-rays. In Step 2, these 36 image-derived features are concatenated with patient demographics (4 elements) and blood test results (35 elements), creating a combined set of 75 features. This tabular data is then processed by TabPFN, a transformer-based model, to predict ALM.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/f4dc212e2f1f21712ad815c9.png"},{"id":101202388,"identity":"4e42026b-7dcc-4ea8-a4ed-1f48ed070602","added_by":"auto","created_at":"2026-01-27 09:29:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":209119,"visible":true,"origin":"","legend":"\u003cp\u003eTabPFN-Multimodal model performance results. Scatterplots showing ground truth and predicted values of the TabPFN-Multimodal model for the internal test set (A) and the external validation set (B). Pearson's correlation coefficients are 0.958 (95% CI 0.951–0.964) for the internal test set and 0.930 (95% CI 0.925–0.935) for the external validation set. Bland-Altman plots showing the difference between predicted and measured ALM values versus their mean for the internal test set (C) and the external validation set (D). The plots contain three horizontal lines: the middle black dashed line representing the mean difference, and the upper and lower red dashed lines representing the 95% limits of agreement.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/137a0f1adbd941fc2bc48011.png"},{"id":100857512,"identity":"ce7a4b25-c9cf-4211-91c4-76c496f67f3e","added_by":"auto","created_at":"2026-01-22 07:15:43","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":276225,"visible":true,"origin":"","legend":"\u003cp\u003eGrad-CAM visualization of VGG16 models for ALM prediction. Average saliency maps from the VGG16 models predicting appendicular lean mass from chest X-rays. The top panels show averaged chest radiographs and the bottom panels show corresponding averaged Grad-CAM saliency maps. Panels A and B display frontal chest X-rays and their activation patterns from the internal and external test sets, respectively. Panels C and D show lateral chest X-rays and their activation patterns from the internal and external test sets, respectively. Hot areas (red-yellow) in the saliency maps indicate regions most influential for appendicular lean mass prediction.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/ab780e95d227bc3aac518e7b.png"},{"id":100857513,"identity":"9c64b941-cf9d-4fe0-8b20-e4375afa5aa5","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":98436,"visible":true,"origin":"","legend":"\u003cp\u003eSHapley Additive exPlanations (SHAP) value analysis of the TabPFN in the multimodal model. SHAP summary plot showing the distribution of SHAP values for the top 20 features in the TabPFN in the multimodal model for appendicular lean mass prediction. Each dot represents a SHAP value for a feature per participant. The x-axis represents the SHAP value (impact on model output), and the color varying from blue to red represents the feature value from low to high, respectively. For example, higher values of 'Weight' tend to have positive SHAP values, suggesting a higher predicted\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/fa6ca9c6cc04630e22066577.png"},{"id":106344287,"identity":"c36b291f-1d2c-4cbb-a9cb-3c11f20e0dee","added_by":"auto","created_at":"2026-04-07 16:13:13","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1887449,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/f4c64971-c181-4ca9-a689-3f2538347237.pdf"},{"id":100857517,"identity":"f5be1e88-835a-4eac-9c4c-81e8a7f3cd6f","added_by":"auto","created_at":"2026-01-22 07:15:44","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":180285,"visible":true,"origin":"","legend":"","description":"","filename":"Supportinginformation20260106.docx","url":"https://assets-eu.researchsquare.com/files/rs-8556884/v1/7bd02aba04875f231cbf01a4.docx"}],"financialInterests":"","formattedTitle":"Transformer-based multimodal model for estimation of appendicular lean mass using incomplete chest radiographs and electronic health record","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSarcopenia is a progressive and systemic muscle disease characterized by age-related decline in skeletal muscle mass and strength, leading to serious health consequences including increased risk of falls, fractures, dependency, and mortality\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. In aging societies, its rising prevalence represents a significant health issue that diminishes patients' quality of life and consumes healthcare resources\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Early detection and diagnosis are crucial, with appendicular lean mass (ALM) assessment playing a central role in diagnostic criteria for sarcopenia as a musculoskeletal disorder\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. International consensus definitions from EWGSOP2 and AWGS adopt \"low muscle strength\u0026thinsp;+\u0026thinsp;low muscle mass\" for sarcopenia diagnosis, recommending ALM measurement via dual-energy X-ray absorptiometry (DXA) or bioelectrical impedance analysis (BIA)\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. However, DXA's high cost and limited availability, combined with BIA's susceptibility to hydration status and measurement variability, create barriers to widespread clinical implementation\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Consequently, there is a need for developing alternative methods that can assess muscle mass accurately and conveniently using existing data.\u003c/p\u003e \u003cp\u003eRecent artificial intelligence (AI) studies utilizing existing data for sarcopenia assessment have predominantly focused on single modalities. Examples include deep learning attempts to estimate ALM from frontal chest X-rays for sarcopenia screening\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e and machine learning models detecting sarcopenia patients using structured electronic health record (EHR) data including blood test results\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. These studies demonstrated the potential for non-invasive muscle mass estimation that was previously difficult to measure. However, there are currently no AI studies utilizing lateral chest X-rays in the sarcopenia field, nor are there any multimodal approaches that integrate different data types such as frontal chest X-rays and EHR data. Additionally, handling missing data modalities in multimodal AI remains inadequately addressed\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. In previous multimodal analyses in the skeletal muscle field\u003csup\u003e\u003cspan additionalcitationids=\"CR9\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, cases with missing modalities were excluded, which may introduce bias and prevent accurate evaluation of model performance.\u003c/p\u003e \u003cp\u003eTo advance precision medicine in sarcopenia and frailty, the importance of multimodal diagnostic approaches combining clinical information, imaging, and biological data has been discussed\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Previous skeletal muscle research has reported improved accuracy through integration of multiple modalities compared to single modalities\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Furthermore, transformer-based algorithms have shown potential effectiveness for handling missing modalities in multimodal medical data\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, suggesting they may achieve sufficient accuracy in multimodal analysis of appendicular lean mass.\u003c/p\u003e \u003cp\u003eTherefore, this study aims to develop and validate a multimodal AI model that estimates ALM using frontal and lateral chest X-rays along with EHR data. Furthermore, we will investigate whether transformer-based algorithms remain effective even when dealing with missing modalities.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy design and TRIPOD adherence\u003c/h2\u003e \u003cp\u003eThis study developed and validated a multimodal AI model for estimating ALM and detecting low muscle mass in accordance with the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) guidelines\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Following transparent reporting standards for prediction models, we conducted internal validation on the derivation cohort and external validation on an independent cohort.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy population\u003c/h3\u003e\n\u003cp\u003eThe derivation cohort included participants who visited the health screening center at Sumitomo Hospital between January 2016 and June 2023. For participants who visited multiple times, all visits meeting the inclusion criteria were included as separate observations, resulting in 3,295 observations from 1,524 unique participants. The study population was defined according to specific inclusion criteria. For the derivation cohort, inclusion criteria were: (1) aged\u0026thinsp;\u0026ge;\u0026thinsp;20 years, (2) DXA measurement of ALM available (iDXA, GE/Lunar), and (3) at least one of chest X-ray (frontal or lateral) or blood test (complete blood count and biochemistry) performed (Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). A total of 3,295 observations met these criteria. This study was performed in accordance with the Declaration of Helsinki and received approval from the Institutional Ethics Committee of the University of Osaka Graduate School of Medicine (approval number: 22099). Informed consent was waived because of the retrospective nature of the study.\u003c/p\u003e \u003cp\u003eFor external validation, data from the University of Osaka Hospital were used. Between January 2012 and April 2023, 2,738 participants underwent DXA (Horizon A, Hologic) for ALM measurement. Chest X-rays (frontal and lateral) and blood tests (complete blood count and biochemistry) performed within 3 months before or after DXA were collected. Applying the same inclusion criteria as the derivation cohort, 1,976 participants remained. For participants who underwent multiple DXA examinations during the study period, all examinations were included as separate observations, resulting in 3,771 observations for the external validation cohort.\u003c/p\u003e\n\u003ch3\u003eMultimodal AI model architecture\u003c/h3\u003e\n\u003cp\u003eOur multimodal AI model was developed using a two-step approach, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The model architecture, including the pre-trained CNN, was prespecified and not modified based on the results. Step 1 involved extracting 18 features (Table S2) from each of the frontal and lateral chest X-rays using a pre-trained CNN, resulting in a total of 36 imaging features. We used TorchXray, which was trained on large chest X-ray datasets including PadChest, NIH, CheXpert, and MIMIC\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Step 2 combined the 36 features extracted from frontal and lateral chest X-rays with EHR data consisting of demographics (height, weight, sex, age) and blood test results, resulting in a total of 75 features. These combined features were then processed through TabPFN, a transformer-based model for tabular data\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, to predict ALM. Missing data or modalities were handled by treating them as missing values without imputation, which was pre-planned in our analysis strategy.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eModel comparison\u003c/h3\u003e\n\u003cp\u003eTo compare multimodal and unimodal approaches, we developed unimodal models for each modality. For EHR data, we used TabPFN. To process the chest X-ray images, we employed two models: VGG16, and an alternative model that used TabPFN to analyze features extracted by TorchXray. Additionally, we evaluated the performance of the transformer-based TabPFN against existing machine learning models by replacing TabPFN in Step 2 of our multimodal model with other machine learning models. We used PyCaret (version 3.3.2) to validate the accuracy of these alternative machine learning models.\u003c/p\u003e\n\u003ch3\u003eData preprocessing\u003c/h3\u003e\n\u003cp\u003eFor each participant\u0026rsquo;s raw Digital Imaging and Communications in Medicine (DICOM) image, pixel values were extracted using the pydicom library (version 2.4.4) in Python. Images were resized to 224 \u0026times; 224 pixels and processed by the TorchXray model. For the unimodal VGG16 model, contrast-limited adaptive histogram equalization was additionally applied to increase image contrast. Low muscle mass was defined according to AWGS criteria\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e based on the skeletal muscle index (SMI), which was calculated as ALM (kg) divided by height squared (m\u0026sup2;). Low SMI was defined as \u0026lt;\u0026thinsp;7.0 kg/m\u0026sup2; in males and \u0026lt;\u0026thinsp;5.4 kg/m\u0026sup2; in females. For EHR data, no imputation was performed for missing values.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eModel validation\u003c/h2\u003e \u003cp\u003eModel performance was evaluated using predefined metrics. For the regression task of ALM prediction, we assessed model performance using root mean square error (RMSE) in kg, mean absolute error (MAE) in kg, and Pearson correlation coefficient between predicted and actual ALM values. These metrics were used to compare the performance of TabPFN-multimodal model with unimodal models (TabPFN-EHR, VGG16-chest X-ray, and TabPFN-chest X-ray) as well as to compare TabPFN with existing machine learning models in the multimodal framework (step 2 of the model architecture). For classification tasks of low muscle mass detection based on AWGS criteria, model performance was evaluated using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), sensitivity, specificity, and F1 score.\u003c/p\u003e \u003cp\u003eInternal validation was performed using an 80:20 split of the derivation dataset, with further 80:20 splitting for training and validation. External validation was conducted on an independent cohort from the University of Osaka Hospital to assess model transportability. Model discrimination and agreement for the TabPFN-multimodal model were further visualized using Pearson correlation plots and Bland-Altman plots with 95% limits of agreement.\u003c/p\u003e \u003cp\u003eTo ensure the TabPFN-multimodal model's robustness in the external validation set, we conducted subgroup analyses by age and sex categories to assess model performance across different populations. Additionally, sensitivity analyses were performed to evaluate model performance in participant subsets with incomplete data: participants missing at least one EHR parameter (blood test or demographic data), participants missing frontal chest X-ray, and participants missing lateral chest X-ray.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eModel interpretability analysis\u003c/h3\u003e\n\u003cp\u003eTo enhance model transparency and clinical applicability, we performed interpretability analyses on our models. For the unimodal VGG16 model that processes chest X-ray images, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize and understand the regions in chest X-ray images that the model focused on for feature extraction\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. For the TabPFN-based multimodal model, SHAP (SHapley Additive exPlanations) scores were calculated to interpret the model output and identify the contribution of each input feature to ALM prediction. These analyses provide insights into which anatomical regions in chest X-rays are important for the VGG16 model and which features from both chest X-rays and EHR data most significantly influence the multimodal model's predictions.\u003c/p\u003e\n\u003ch3\u003eHyperparameters\u003c/h3\u003e\n\u003cp\u003eFor the training of the unimodal VGG16 model, data augmentation was performed on the training images using random shift, rotation, horizontal flip, zoom, brightness, and random multiplication five times. The model was trained for 100 epochs with early stopping (patience\u0026thinsp;=\u0026thinsp;15), a batch size of 32, and a cosine annealing learning rate. The Adam optimizer was used to minimize the Gaussian loss function (negative log-likelihood loss following Gaussian distributions).\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eAll statistical methods were prespecified to ensure transparent reporting. Baseline characteristics were summarized for both derivation and external validation cohorts. Continuous variables were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation based on their distribution and compared using the independent two-sample t-test or Wilcoxon rank-sum test as appropriate. Categorical variables were presented as number (percentage) and compared using chi-square test or Fisher's exact test between groups with or without low muscle mass. To derive the 95% confidence intervals of model performance metrics in the internal test and external validation sets, the bootstrap method with 2000 repetitions was used. All statistical analyses were performed using Scikit-Learn (version 0.24.2) and SciPy (version 1.10.1). A two-sided P-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eClinical characteristics\u003c/h2\u003e \u003cp\u003eThe mean age of the derivation cohort (n\u0026thinsp;=\u0026thinsp;3,295 observations from 1,524 participants, females 30%) and the external validation cohort (n\u0026thinsp;=\u0026thinsp;3,771 observations from 1,976 participants, females 59%) was 61.9 and 58.4 years, respectively. The prevalence of low muscle mass was 30% in the derivation set and 29% in the external validation set (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). In both cohorts, individuals with low muscle mass had lower weight and ALM (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001 for both). The proportion of females was higher in the low muscle mass group in the derivation set but lower in the external validation set (P\u0026thinsp;\u0026lt;\u0026thinsp;0.001 for both cohorts).\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\u003eClinical characteristics of study subjects\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\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 \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eDerivation set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eExternal validation set\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLow muscle mass (n\u0026thinsp;=\u0026thinsp;991, 30%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo low muscle mass (n\u0026thinsp;=\u0026thinsp;2304, 70%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eLow muscle mass (n\u0026thinsp;=\u0026thinsp;1107, 29%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo low muscle mass (n\u0026thinsp;=\u0026thinsp;2664, 71%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, year\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e64.6\u0026thinsp;\u0026plusmn;\u0026thinsp;12.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e60.8\u0026thinsp;\u0026plusmn;\u0026thinsp;11.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e57.0\u0026thinsp;\u0026plusmn;\u0026thinsp;19.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e59.0\u0026thinsp;\u0026plusmn;\u0026thinsp;15.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.087\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale, \u003cem\u003en\u003c/em\u003e (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e334 (34)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e634 (28)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e511 (46)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1699 (64)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWeight, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e56.8\u0026thinsp;\u0026plusmn;\u0026thinsp;9.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e70.0\u0026thinsp;\u0026plusmn;\u0026thinsp;12.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e49.9\u0026thinsp;\u0026plusmn;\u0026thinsp;9.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e60.3\u0026thinsp;\u0026plusmn;\u0026thinsp;14.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHeight, m\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.64\u0026thinsp;\u0026plusmn;\u0026thinsp;0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.66\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.59\u0026thinsp;\u0026plusmn;\u0026thinsp;0.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.58\u0026thinsp;\u0026plusmn;\u0026thinsp;0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eALM, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e16.4\u0026thinsp;\u0026plusmn;\u0026thinsp;3.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20.8\u0026thinsp;\u0026plusmn;\u0026thinsp;4.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14.4\u0026thinsp;\u0026plusmn;\u0026thinsp;3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e17.9\u0026thinsp;\u0026plusmn;\u0026thinsp;4.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003eALM, appendicular lean mass\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eComparing the multimodal model with unimodal models:\u003c/h2\u003e \u003cp\u003eMultiple models were developed to predict ALM from EHR, frontal or lateral chest X-rays, or these three modalities integrated (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Among all these models, the TabPFN-multimodal model combining chest X-ray features with EHR data achieved the highest performance. In the internal test set, the multimodal model achieved prediction errors of RMSE 1.23 kg (95% CI: 1.18\u0026ndash;1.29) and MAE 0.96 kg (0.91\u0026ndash;1.02), with a correlation coefficient of 0.958 (0.951\u0026ndash;0.964). In the external validation set, the model yielded an RMSE of 1.87 kg (1.83\u0026ndash;1.91) and MAE of 1.41 kg (1.38\u0026ndash;1.45), with a correlation coefficient of 0.930 (0.925\u0026ndash;0.935). The correlation and agreement between predicted and measured ALM values for the multimodal model are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Bland-Altman analysis revealed minimal systematic bias in both datasets. In the internal test set, the mean difference was 0.07 kg, with 95% limits of agreement ranging from \u0026minus;\u0026thinsp;2.35 to 2.48 kg. In the external validation set, the model maintained a low mean bias of -0.22 kg, although the 95% limits of agreement widened slightly to -3.52 to 3.08 kg. The plots showed no distinct proportional bias across the range of ALM values.\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 of multimodal and unimodal models for predicting appendicular lean mass\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\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=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel - Modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eInternal validation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eExternal validation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRMSE, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMAE, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCorrelation Coefficient\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRMSE, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eMAE, kg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCorrelation Coefficient\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGG16 - AP*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.09 (2.96\u0026ndash;3.22)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.43 (2.30\u0026ndash;2.56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.840 (0.819\u0026ndash;0.859)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.97 (4.87\u0026ndash;5.07)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.23 (4.13\u0026ndash;4.34)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.578 (0.552\u0026ndash;0.604)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVGG16 - LAT*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.13 (3.00\u0026ndash;3.25)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.49 (2.36\u0026ndash;2.62)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.807 (0.782\u0026ndash;0.829)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.22 (4.08\u0026ndash;4.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.18 (3.03\u0026ndash;3.32)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.438 (0.395\u0026ndash;0.479)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN - AP*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.06 (2.91\u0026ndash;3.21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.38 (2.23\u0026ndash;2.53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.718 (0.678\u0026ndash;0.754)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e4.08 (3.98\u0026ndash;4.18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3.03 (2.92\u0026ndash;3.13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.493 (0.463\u0026ndash;0.521)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN - LAT*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.90 (2.76\u0026ndash;3.04)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2.30 (2.16\u0026ndash;2.44)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.745 (0.707\u0026ndash;0.778)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.45 (3.38\u0026ndash;3.52)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2.71 (2.64\u0026ndash;2.78)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.692 (0.675\u0026ndash;0.708)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN - EHR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.43(1.36\u0026ndash;1.50)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.11 (1.04\u0026ndash;1.18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.946 (0.937\u0026ndash;0.954)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.02 (1.93\u0026ndash;2.11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.57 (1.53\u0026ndash;1.61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.920 (0.915\u0026ndash;0.925)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTabPFN - Multimodal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.23 (1.18\u0026ndash;1.29)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.96 (0.91\u0026ndash;1.02)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.958 (0.951\u0026ndash;0.964)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.87 (1.83\u0026ndash;1.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.41 (1.38\u0026ndash;1.45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.930 (0.925\u0026ndash;0.935)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003e* Performance was measured exclusively on cases with available chest X-ray images\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003eAP, frontal chest X-rays; LAT, lateral chest X-rays; EHR, electronic health record; RMSE, root mean square error; MAE, mean absolute error.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"7\"\u003eValues in parentheses are 95% confidence intervals, calculated using the bootstrap method.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe TabPFN-EHR model, using EHR data alone, achieved an RMSE of 1.43 kg (95% CI: 1.36\u0026ndash;1.50) and MAE of 1.11 kg (1.04\u0026ndash;1.18) in the internal test set, while maintaining robust performance in the external validation set with an RMSE of 2.02 kg (1.93\u0026ndash;2.11) and MAE of 1.57 kg (1.53\u0026ndash;1.61). The correlation coefficients were 0.946 (0.937\u0026ndash;0.954) and 0.920 (0.915\u0026ndash;0.925), respectively. In contrast, models utilizing chest X-ray features alone (VGG16-AP, VGG16-LAT, TabPFN-AP, TabPFN-LAT) exhibited higher prediction errors, with RMSE values ranging from 2.90 to 3.13 kg (MAE: 2.30\u0026ndash;2.49 kg) in internal validation and 3.45 to 4.97 kg (MAE: 2.71\u0026ndash;4.23 kg) in external validation, corresponding to correlation coefficients of 0.718\u0026ndash;0.840 and 0.438\u0026ndash;0.692, respectively.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eComparing TabPFN with traditional machine learning models\u003c/h2\u003e \u003cp\u003eTo evaluate a range of algorithms for the final prediction stage of our multimodal architecture (Step 2), we compared the performance of the transformer-based TabPFN against several traditional machine learning models for ALM prediction (Table S3). In the internal validation set, TabPFN demonstrated superior performance, achieving the lowest prediction errors with an RMSE of 1.23 kg (95% CI: 1.18\u0026ndash;1.29) and MAE of 0.96 kg (0.91\u0026ndash;1.02), alongside the highest correlation coefficient of 0.958 (0.951\u0026ndash;0.964). The best-performing traditional machine learning model was the Gradient Boosting Regressor, which yielded slightly higher errors with an RMSE of 1.25 kg (1.18\u0026ndash;1.32) and MAE of 0.98 kg (0.92\u0026ndash;1.04), and a correlation coefficient of 0.957 (0.951\u0026ndash;0.962). In the external validation dataset, TabPFN maintained its superiority, recording an RMSE of 1.87 kg (1.83\u0026ndash;1.91) and MAE of 1.41 kg (1.38\u0026ndash;1.45). This substantially outperformed the best conventional model (Gradient Boosting Regressor), which showed increased prediction errors with an RMSE of 2.07 kg (2.01\u0026ndash;2.13) and MAE of 1.59 kg (1.55\u0026ndash;1.63), corresponding to a lower correlation coefficient of 0.919 (0.912\u0026ndash;0.925). Notably, while TabPFN kept the RMSE below 1.9 kg in external validation, all traditional models exhibited RMSE values exceeding 2.0 kg. These results demonstrate that the transformer-based TabPFN model consistently minimized prediction errors compared to established machine learning approaches across all evaluation metrics and datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003ePerformance of multimodal model for low muscle mass prediction\u003c/h2\u003e \u003cp\u003eThe TabPFN-multimodal model predicted low muscle mass based on AWGS criteria (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). In the internal test set, sex-specific AUROCs were 0.832 (95% CI: 0.793\u0026ndash;0.869) for males and 0.808 (95% CI: 0.747\u0026ndash;0.864) for females (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e), with sensitivity of 0.766 (95% CI: 0.700-0.821) and specificity of 0.836 (95% CI: 0.801\u0026ndash;0.867). As shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the model demonstrated robust diagnostic performance in the external validation set, with AUROCs of 0.825 (95% CI: 0.804\u0026ndash;0.843) for males and 0.730 (95% CI: 0.708\u0026ndash;0.754) for females. While the sensitivity was 0.558 (95% CI: 0.534\u0026ndash;0.582), the specificity remained high at 0.910 (95% CI: 0.897\u0026ndash;0.921). AUPRCs in the internal test set were 0.909 (95% CI: 0.879\u0026ndash;0.936) for males and 0.895 (95% CI: 0.846\u0026ndash;0.937) for females, while external validation AUPRCs were 0.880 (95% CI: 0.860\u0026ndash;0.901) for males and 0.904 (95% CI: 0.891\u0026ndash;0.916) for females.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of the multimodal model for low muscle mass prediction\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePerformance metrics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInternal test set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eExternal test set\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUROC-Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.832\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.793\u0026ndash;0.869\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.825\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.804\u0026ndash;0.843\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUROC-Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.808\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.747\u0026ndash;0.864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.730\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.708\u0026ndash;0.754\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUPRC-Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.879\u0026ndash;0.936\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.880\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.860\u0026ndash;0.901\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAUPRC-Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.895\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.846\u0026ndash;0.937\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.904\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.891\u0026ndash;0.916\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity-AWGS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.766\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.700\u0026ndash;0.821\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.558\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.534\u0026ndash;0.582\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity-AWGS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.836\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.801\u0026ndash;0.867\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.910\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.897\u0026ndash;0.921\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eF1 score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.698\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.645\u0026ndash;0.744\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.666\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.645\u0026ndash;0.685\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e95% CI was calculated using bootstrapping method.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eSubgroup and sensitivity analysis\u003c/h2\u003e \u003cp\u003eSubgroup analysis in the external validation cohort showed RMSE values ranging from 1.62 kg (95% CI: 1.46\u0026ndash;1.79 kg) in participants aged\u0026thinsp;\u0026ge;\u0026thinsp;80 years to 2.17 kg (95% CI: 2.05\u0026ndash;2.30 kg) in participants\u0026thinsp;\u0026lt;\u0026thinsp;50 years (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Performance differed by sex, with males exhibiting higher predictive errors (RMSE: 2.37 kg, 95% CI: 2.29\u0026ndash;2.47 kg; MAE: 1.86 kg, 95% CI: 1.79\u0026ndash;1.93 kg) compared to females (RMSE: 1.64 kg, 95% CI: 1.58\u0026ndash;1.70 kg; MAE: 1.25 kg, 95% CI: 1.21\u0026ndash;1.30 kg). Sensitivity analyses in participants with missing modalities showed RMSE of 2.00 kg (95% CI: 1.85\u0026ndash;2.15 kg) in 261 participants missing at least one EHR parameter, 1.99 kg (95% CI: 1.92\u0026ndash;2.06 kg) in 1,197 participants missing frontal chest X-ray, and 2.08 kg (95% CI: 2.03\u0026ndash;2.13 kg) in 2,346 participants missing lateral chest X-ray.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSubgroup and sensitivity analyses of the multimodal model's performance in the external validation cohort.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\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=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnalysis Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubgroup / Missing modality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE, kg\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMAE, kg\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCorrelation Coefficient\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eComplete case\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e3771\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.87 (1.83\u0026ndash;1.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.41 (1.38\u0026ndash;1.45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.930 (0.925\u0026ndash;0.935)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSubgroup Analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge: \u0026lt;50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1086\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.17 (2.05\u0026ndash;2.30)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.64 (1.54\u0026ndash;1.72)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.939 (0.930\u0026ndash;0.947)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge: 50\u0026ndash;64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e985\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.03 (1.92\u0026ndash;2.14)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.53 (1.44\u0026ndash;1.61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.926 (0.915\u0026ndash;0.935)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge: 65\u0026ndash;79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1459\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.84 (1.76\u0026ndash;1.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.43 (1.37\u0026ndash;1.50)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.918 (0.907\u0026ndash;0.928)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge: 80\u0026lt;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e241\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.62 (1.46\u0026ndash;1.79)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.25 (1.12\u0026ndash;1.39)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.906 (0.876\u0026ndash;0.928)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSex: Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1561\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.37 (2.29\u0026ndash;2.47)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.86 (1.79\u0026ndash;1.93)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.882 (0.867\u0026ndash;0.897)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSex: Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2210\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.64 (1.58\u0026ndash;1.70)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.25 (1.21\u0026ndash;1.30)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.846 (0.828\u0026ndash;0.862)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: \u0026lt;50_Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e503\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.65 (2.47\u0026ndash;2.85)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.07 (1.92\u0026ndash;2.21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.903 (0.883\u0026ndash;0.922)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: \u0026lt;50_Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e583\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.66 (1.54\u0026ndash;1.79)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.27 (1.18\u0026ndash;1.36)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.907 (0.884\u0026ndash;0.926)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 50\u0026ndash;64_Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e356\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.52 (2.32\u0026ndash;2.68)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.96 (1.80\u0026ndash;2.11)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.830 (0.788\u0026ndash;0.863)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 50\u0026ndash;64_Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e629\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.69 (1.58\u0026ndash;1.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.28 (1.20\u0026ndash;1.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.811 (0.780\u0026ndash;0.839)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 65\u0026ndash;79_Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e602\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.13 (2.01\u0026ndash;2.24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.71 (1.61\u0026ndash;1.81)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.829 (0.796\u0026ndash;0.858)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 65\u0026ndash;79_Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e857\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.60 (1.51\u0026ndash;1.69)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.24 (1.18\u0026ndash;1.31)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.763 (0.719\u0026ndash;0.805)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 80\u0026lt;_Male\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.68 (1.44\u0026ndash;1.91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.35 (1.15\u0026ndash;1.56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.868 (0.827\u0026ndash;0.904)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAge_Sex: 80\u0026lt;_Female\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.57 (1.32\u0026ndash;1.79)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.17 (0.988\u0026ndash;1.35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.808 (0.725\u0026ndash;0.871)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity Analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEHR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e261\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.00 (1.85\u0026ndash;2.15)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.58 (1.43\u0026ndash;1.73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.951 (0.938\u0026ndash;0.962)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrontal chest X-ray\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1197\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.99 (1.92\u0026ndash;2.06)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.54 (1.47\u0026ndash;1.61)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.932 (0.924\u0026ndash;0.939)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLateral chest X-ray\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2346\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.08 (2.03\u0026ndash;2.13)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.60 (1.55\u0026ndash;1.66)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.923 (0.917\u0026ndash;0.929)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"6\" nameend=\"c6\" namest=\"c1\"\u003e \u003cp\u003eValues in parentheses are 95% confidence intervals, calculated using the bootstrap method.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eModel Interpretability Analysis:\u003c/h2\u003e \u003cp\u003eThe Grad-CAM analysis for the VGG16 models predicting ALM is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Saliency maps were generated separately for frontal and lateral chest X-rays from both the internal test set and external validation set, and each map was averaged across all subjects to produce representative saliency images for each view. The averaged Grad-CAM visualizations demonstrated that the models consistently highlighted specific anatomical regions. For frontal chest X-rays, the internal validation model showed prominent activation around both upper arm roots (proximal humerus regions), while the external validation model focused on both upper arm roots and additionally highlighted the right lung field. For lateral chest X-rays, both internal and external validation models consistently focused on the thoracic spine region. These findings suggest that the VGG16 models utilize anatomically relevant features, particularly musculoskeletal landmarks and surrounding structures, for ALM prediction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSHAP values for the TabPFN in the multimodal model on the external validation set are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The feature importance analysis revealed that demographic features were the most important for ALM predictions, with weight, sex, height, and age ranking as the top four, respectively. Among blood test parameters, creatinine and LDH were the most important factors, ranking fifth and sixth overall, respectively. Following these top six features, the 'Atelectasis' feature from frontal chest X-rays contributed to the predictions. The SHAP analysis demonstrated distinct directional relationships: weight, sex (male: 1, female: 0), height, and creatinine showed a positive correlation with ALM predictions (higher values associated with higher predicted ALM), while age exhibited negative correlations (higher values associated with lower predicted ALM).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eSummary of key findings\u003c/h2\u003e \u003cp\u003eThis study developed and validated a multimodal AI model that combines chest X-ray images (frontal and lateral views) with EHR data (blood test and demographics) to predict ALM and detect low muscle mass. The TabPFN-based multimodal model achieved superior performance compared to unimodal approaches, with an RMSE of 1.23 kg (95% CI: 1.18\u0026ndash;1.29) and MAE of 0.96 kg (0.91\u0026ndash;1.02) in internal validation, and an RMSE of 1.87 kg (1.83\u0026ndash;1.91) and MAE of 1.41 kg (1.38\u0026ndash;1.45) in external validation. For low muscle mass detection, the model demonstrated performance with AUPRCs of 0.88 for males and 0.90 for females in external validation.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eComparative Performance against Prior Studies and Clinical Benchmarks\u003c/h2\u003e \u003cp\u003eThe superior performance of the multimodal approach over unimodal models aligns withprior studies in medical AI research demonstrating that integration of multiple data modalities enhances predictive accuracy in complex clinical tasks\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. The multimodal model achieved reduced prediction errors compared to the EHR-only model. Specifically, in the external validation set, the multimodal model yielded a lower RMSE of 1.87 kg compared to 2.02 kg for the EHR-only model, demonstrating the added value of integrating chest X-rays. Our unimodal chest X-ray models showed mixed results compared to previous work, with strong internal validation performance (correlation coefficient 0.840) but weaker external validation (correlation coefficient 0.578) compared to previous chest X-ray-based models for ALM prediction, which reported concordance correlation coefficients of 0.80 and 0.76 in internal and external validation, respectively\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. This performance drop in external validation highlights the challenge of generalizability with single-modality approaches, and it appears more pronounced than that observed in the previous work\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. A direct comparison with the prior study is challenging, as it does not specify the X-ray acquisition protocols or equipment used. The performance of an AI model can be sensitive to such variations in imaging technique, and it is possible that these unclarified technical differences contributed to the performance degradation observed in our external validation.\u003c/p\u003e \u003cp\u003eIn contrast, our EHR-only model demonstrated robust performance for direct ALM prediction in both internal and external validation settings (correlation coefficients of 0.946 and 0.920, respectively.) This extends previous machine learning research on EHR data for low muscle mass detection\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e, which showed promising results for identifying patients meeting AWGS criteria using a logistic regression model (AUROC values between 0.72\u0026ndash;0.91), but was limited to internal validation only. Our study's external validation performance demonstrates that EHR data alone can reliably predict ALM and potentially identify low muscle mass across different clinical settings, addressing a key limitation in previous research where generalizability remained uncertain. Both the previous and our studies highlight the potential value of EHR-based screening approaches for early detection of low muscle mass in clinical practice, using routinely collected healthcare data. Our multimodal approach demonstrates accuracy comparable to BIA reports when using DXA as the reference standard. While BIA typically overestimates ALM by 1.97 kg compared to DXA measurements, our TabPFN-based multimodal model yielded a MAE of 1.41 kg (95% CI: 1.38\u0026ndash;1.45) in external validation. These findings suggest that our model offers a viable alternative to BIA, a traditional method recommended in guidelines, while offering the practical advantage of using routinely available clinical data.\u003c/p\u003e \u003cp\u003eOur Bland-Altman analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) further confirmed that the multimodal model is reliable for regression tasks, showing negligible fixed bias even in the external cohort. While the 95% limits of agreement in external validation (approximately\u0026thinsp;\u0026plusmn;\u0026thinsp;3.5 kg) were wider than in internal validation, this range is comparable to the reported error margins between BIA\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Although some outliers exceeded these limits\u0026mdash;likely due to image artifacts or extreme body compositions inherent in real-world clinical data\u0026mdash;the majority of predictions fell within a clinically acceptable range. These findings support the model's robustness and its potential utility as a screening tool, where minimizing systematic bias is crucial for identifying individuals at risk.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003ePerformance Differences by Sex\u003c/h2\u003e \u003cp\u003eIn our external validation, we observed a divergence in performance metrics between sexes: the model achieved a higher AUROC for detecting low muscle mass in males (0.825) compared to females (0.730), whereas the estimation error (RMSE) for absolute ALM values was larger in males (2.37 kg) than in females (1.64 kg) (Tables\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and \u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This discrepancy in RMSE is likely attributable to the physiological difference in baseline muscle mass; males typically possess greater total muscle mass, resulting in a wider range of values and consequently larger absolute errors, a phenomenon consistent with scaling effects. Conversely, the superior discriminative ability (AUROC) in males aligns with findings from previous AI studies in sarcopenia\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. This may be because the age-related decline in muscle mass is often more pronounced and distinct in males, making the differentiation between normal and low muscle mass relatively clearer for the model. Importantly, our SHAP analysis identified sex as the second most important feature, suggesting that the multimodal model successfully learned and adjusted for these physiological sex-based differences within a single architecture.\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eModel Interpretability and Technical Advantages\u003c/h2\u003e \u003cp\u003eAnalysis of SHAP values revealed that the top predictors were six EHR-derived features\u0026mdash;weight, sex, height, age, creatinine, and LDH, in that order\u0026mdash;followed by the 'Atelectasis' feature from frontal chest X-rays (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). SHAP analysis demonstrated a positive correlation between creatinine levels and predicted ALM (higher creatinine associated with higher predicted ALM). This finding is consistent with the fact that creatinine is a byproduct of muscle metabolism, and its serum levels are directly influenced by an individual's skeletal muscle mass\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. The prominence of demographics (weight, sex, height, and age) as top predictors underscores their fundamental importance in determining muscle mass, while the contribution of chest X-ray features suggests that imaging provides complementary information about thoracic anatomy that relates to overall body composition. Although the Grad-CAM visualization was derived from the unimodal VGG16 model rather than the final transformer-based architecture, the consistent focus on anatomically relevant regions\u0026mdash;specifically the proximal humerus and thoracic spine\u0026mdash;validates that the deep learning algorithms are leveraging biologically plausible signals from the chest radiographs. This supports the clinical interpretability of the imaging features integrated into our multimodal framework.\u003c/p\u003e \u003cp\u003eThe TabPFN-based multimodal model consistently outperformed traditional machine learning approaches across all evaluation metrics (Table S3). While several conventional models showed comparable performance in the internal validation, TabPFN maintained its superior predictive capability in the external validation dataset where traditional methods exhibited substantial performance degradation. This finding aligns with research demonstrating that transformer-based models excel on tabular datasets, particularly for smaller datasets with fewer than 10,000 samples, compared to conventional approaches\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Notably, our sensitivity analysis revealed that the TabPFN-multimodal model maintained robust performance even with missing data modalities. The model demonstrated stable prediction errors even when handling patients with missing EHR parameters (RMSE: 2.00 kg), missing frontal chest X-rays (RMSE: 1.99 kg), or missing lateral chest X-rays (RMSE: 2.08 kg). This excellent performance with incomplete data supports recent findings that transformer-based architectures effectively manage missing healthcare data through their self-attention mechanisms\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Such robustness to missing modalities is particularly valuable for clinical implementation, where data completeness varies significantly across patients and institutions, potentially eliminating the need for complex imputation strategies while maintaining high predictive accuracy.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eClinical Significance and Practicality\u003c/h2\u003e \u003cp\u003eThe multimodal AI model developed in this study demonstrated the potential to accurately estimate ALM and detect low muscle mass from readily available clinical data, including chest X-ray images, blood test results, and basic anthropometric information without requiring special examinations. This could be a powerful solution to the current challenges where DXA, the gold standard for measuring ALM, is expensive and has limited availability. Specifically, by utilizing data from regular health check-ups or routine inpatient examinations, this model is expected to serve as a gatekeeper, efficiently screening patients at potential risk of sarcopenia and facilitating early intervention or more detailed examinations. For example, integrating this model into EHR systems to automatically present estimated ALM values and the likelihood of low muscle mass, when relevant test data are available would enable physicians to quickly recognize sarcopenia risk and promote appropriate nutritional guidance, exercise therapy, or referrals to specialists. Particularly in regions or facilities where access to DXA scans is difficult, this model could contribute to improving the quality of sarcopenia care and reducing disparities in medical access. Furthermore, the model's robust performance even with missing data is a crucial factor enhancing its practicality, as stable operation can be expected under diverse data acquisition conditions in real-world clinical practice. However, from a practical standpoint, the trade-off between maximal accuracy and implementation feasibility must be considered. While the multimodal model achieved the highest predictive accuracy, the performance gain over the EHR-only model was incremental. This suggests that the EHR-only model may serve as a sufficient and scalable screening tool in resource-limited settings. Nevertheless, the multimodal approach offers additional value in terms of robustness against missing biochemical data and provides complementary anatomical information, serving as a comprehensive gatekeeper for sarcopenia risk assessment. As a tangible implementation of this approach, the web application we have developed (available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://alm-application.onrender.com\u003c/span\u003e\u003cspan address=\"https://alm-application.onrender.com\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) validates the value of the EHR-only model as a tool that clinicians can use immediately without requiring special infrastructure.\u003c/p\u003e \u003cp\u003eA critical aspect for optimizing this model as a screening tool is the adjustment of its diagnostic threshold. In the external validation, the model demonstrated high specificity (0.910) but a moderate sensitivity of 0.558. This sensitivity was observed at the fixed diagnostic cut-off values defined by the AWGS criteria, which are designed for confirmatory diagnosis rather than screening. Given the model's acceptable discriminative performance (AUROC\u0026thinsp;\u0026gt;\u0026thinsp;0.80 for males and \u0026gt;\u0026thinsp;0.73 for females), the sensitivity can be readily enhanced for screening purposes by adjusting the decision threshold (e.g., using the Youden Index) to prioritize sensitivity over specificity. This flexibility allows the model to be tailored to specific clinical needs, where identifying potential at-risk individuals is prioritized over minimizing false positives.\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study has several limitations. First, the study population may have a selection bias. Although the derivation cohort consisted of participants from a single health screening center and the external validation cohort used data from a different university hospital, these populations may not necessarily represent the general population or patient groups with diverse disease backgrounds. However, the model's demonstrated accuracy in external validation across two potentially distinct cohorts suggests a degree of generalizability. Further multicenter collaborative research is desirable to validate the model's performance in broader medical settings and among diverse racial and ethnic groups.\u003c/p\u003e \u003cp\u003eSecond, a consideration for the future application of our model concerns its scalability. In this study, we employed TabPFN, which demonstrated excellent performance on our dataset sizes (derivation: 3,295; external validation: 3,771). This aligns with reports that TabPFN performs particularly well on tabular data with fewer than 10,000 samples\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. However, the scalability of this specific approach could be a limitation when applying it to much larger datasets, such as nationwide registries or biobanks with tens of thousands of observations. In such scenarios, the computational efficiency and performance might not be maintained. Therefore, future work to scale our multimodal framework would necessitate exploring alternative or supplementary methods, such as ensembling TabPFN with gradient boosting models like LightGBM, or adopting other transformer-based architectures designed for large-scale data.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis study developed and validated a novel multimodal AI model using transformer-based TabPFN that accurately estimates ALM and detects low muscle mass from routinely collected clinical data (chest X-rays, blood tests, demographics), achieving superior performance over unimodal approaches. The model demonstrated robust performance even with missing data modalities, supporting its potential utility as a screening tool in clinical settings where data completeness varies. These findings underscore the potential of this multimodal AI approach as an accessible screening tool for early low muscle mass identification, particularly in settings where DXA availability is limited, facilitating timely interventions and improved patient outcomes.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eALM: Appendicular Lean Mass; DXA: Dual-energy X-ray Absorptiometry; EHR: Electronic Health Record; AI: Artificial Intelligence; RMSE: Root Mean Square Error; MAE: Mean Absolute Error; AUROC: Area Under the Receiver Operating Characteristic Curve; SMI: Skeletal Muscle Index; CNN: Convolutional Neural Network; Grad-CAM: Gradient-weighted Class Activation Mapping; SHAP: SHapley Additive exPlanations.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Ethics Committee of the University of Osaka Graduate School of Medicine (approval number: 22099). Informed consent was waived due to the retrospective nature of the study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e The datasets generated and/or analysed during the current study are not publicly available due to patient privacy protocols but are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e The authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e This work was supported by Japan Osteoporosis Foundation, and JSPS KAKENHI Grant Numbers JP21K20966, JP24KJ1559.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e KK and TF conceived and designed the study. KK, YS, and TF performed the data analysis. KK wrote the manuscript. TF supervised the project. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCruz-Jentoft AJ, Sayer AA, Sarcopenia. Lancet. 2019;393:2636\u0026ndash;46.\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. Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing. 2019;48:16\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen L-K, Woo J, Assantachai P, Auyeung T-W, Chou M-Y, Iijima K, et al. Asian Working Group for Sarcopenia: 2019 Consensus Update on Sarcopenia Diagnosis and Treatment. J Am Med Dir Assoc. 2020;21:300\u0026ndash;e3072.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee SY, Ahn S, Kim YJ, Ji MJ, Kim KM, Choi SH, et al. Comparison between Dual-Energy X-ray Absorptiometry and Bioelectrical Impedance Analyses for Accuracy in Measuring Whole Body Muscle Mass and Appendicular Skeletal Muscle Mass. Nutrients. 2018;10:738.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRyu J, Eom S, Kim HC, Kim CO, Rhee Y, You SC, et al. Chest X-ray-based opportunistic screening of sarcopenia using deep learning. J Cachexia Sarcopenia Muscle. 2023;14:418\u0026ndash;28.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuo X, Ding H, Broyles A, Warden SJ, Moorthi RN, Imel EA. Using machine learning to detect sarcopenia from electronic health records. Digit HEALTH. 2023;9:20552076231197098.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhassemi M, Naumann T, Schulam P, Beam AL, Chen IY, Ranganath R. Practical guidance on artificial intelligence for health-care data. Lancet Digit Health. 2019;1:e157\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaki S, Furuya T, Katsumi K, Nakajima H, Honjoh K, Watanabe S et al. Multimodal Deep Learning-based Radiomics Approach for Predicting Surgical Outcomes in Patients with Cervical Ossification of the Posterior Longitudinal Ligament. Spine. 2024;49:1561\u0026ndash;1569.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKita K, Fujimori T, Suzuki Y, Okada S, Kido S. Bi-modal network combining convolutional neural network and TabNet, differentiating spinal tumors based on images and clinical risk factors. In: Iftekharuddin KM, Chen W, editors. Medical Imaging 2023: Computer-Aided Diagnosis. SPIE: San Diego, United States; 2023. p. 4.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKita K, Fujimori T, Suzuki Y, Kanie Y, Takenaka S, Kaito T, et al. Bimodal artificial intelligence using TabNet for differentiating spinal cord tumors-Integration of patient background information and images. iScience. 2023;26:107900.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLynch DH, Spangler HB, Franz JR, Krupenevich RL, Kim H, Nissman D, et al. Multimodal Diagnostic Approaches to Advance Precision Medicine in Sarcopenia and Frailty. Nutrients. 2022;14:1384.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKline A, Wang H, Li Y, Dennis S, Hutch M, Xu Z et al. Multimodal machine learning in precision health: A scoping review. npj Digit Med 2022;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang K, Wang Y, Ren R, Zou H, Shao Z. Transformer-Based Feature Aggregation and Stitching Network for Crowd Counting. IEEE Access. 2023;11:124833\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B et al. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024;e078378.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen JP, Viviano JD, Bertin P, Morrison P, Torabian P, Guarrera M et al. TorchXRayVision: A library of chest X-ray datasets and models. 2021 doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2111.00595\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2111.00595\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHollmann N, M\u0026uuml;ller S, Eggensperger K, Hutter F. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. 2023 doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.2207.01848\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.2207.01848\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSelvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In: 2017 IEEE International Conference on Computer Vision (ICCV). IEEE: Venice; 2017. pp. 618\u0026ndash;626.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoo HJ, Kim YJ, Hong H, Hong SH, Chae HD, Choi J-Y. Deep learning\u0026ndash;based fully automated body composition analysis of thigh CT: comparison with DXA measurement. Eur Radiol. 2022;32:7601\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNickerson BS, McLester CN, McLester JR, Kliszczewicz BM. Agreement Between 2 Segmental Bioimpedance Devices, BOD POD, and DXA in Obese Adults. J Clin Densitometry. 2020;23:138\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYin L, Zhao J. An Artificial Intelligence Approach for Test-Free Identification of Sarcopenia. J Cachexia Sarcopenia Muscle. 2024;15:2765\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYim J, Son N-H, Kyong T, Park Y, Kim J-H. Muscle mass has a greater impact on serum creatinine levels in older males than in females. Heliyon. 2023;9:e21866.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThongprayoon C, Cheungpasitporn W, Kashani K. Serum creatinine level, a surrogate of muscle mass, predicts mortality in critically ill patients. J Thorac Dis. 2016;8:E305\u0026ndash;311.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTran VQ, Byeon H. Predicting dementia in Parkinson\u0026rsquo;s disease on a small tabular dataset using hybrid LightGBM\u0026ndash;TabPFN and SHAP. Digit HEALTH 2024;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCasella M, Milano N, Dolce P, Marocco D. Transformers deep learning models for missing data imputation: an application of the ReMasker model on a psychometric scale. Front Psychol. 2024;15:1449272.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHollmann N, M\u0026uuml;ller S, Purucker L, Krishnakumar A, K\u0026ouml;rfer M, Hoo SB, et al. Accurate predictions on small data with a tabular foundation model. Nature. 2025;637:319\u0026ndash;26.\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":"journal-of-translational-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jtrm","sideBox":"Learn more about [Journal of Translational Medicine](http://translational-medicine.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/jtrm/default.aspx","title":"Journal of Translational Medicine","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Sarcopenia, Multimodal artificial intelligence, Transformer, Missing modality, Chest radiography","lastPublishedDoi":"10.21203/rs.3.rs-8556884/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8556884/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eSarcopenia is a muscle disease that increases the risk of falls, fractures, and mortality. Appendicular lean mass (ALM) assessment is central to its diagnosis, but standard methods like dual-energy X-ray absorptiometry (DXA) have accessibility and cost issues. Previous artificial intelligence (AI) studies for sarcopenia assessment have been limited to single modalities and have not adequately addressed missing data modalities. This study aimed to develop and validate a multimodal AI model using frontal and lateral chest radiographs and electronic health record (EHR) data to estimate ALM and detect low muscle mass, and to investigate the robustness of a transformer-based algorithm to missing modalities.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003e This model development and validation study adhered to TRIPOD guidelines. The derivation cohort included 3,295 observations from 1,524 participants (mean age 61.9 years, 30% female). External validation was performed on an independent cohort of 3,771 observations from 1,976 participants (mean age 58.4 years, 59% female). Our multimodal model uses a transformer-based TabPFN to predict ALM from 75 features, which integrate 18 features extracted from each of the frontal and lateral chest radiographs by TorchXray with 39 features from EHR data (demographics and blood tests). Model performance for ALM estimation was evaluated using the root mean square error (RMSE), and mean absolute error (MAE), and Pearson correlation coefficient. Estimation of appendicular lean mass performance was assessed using the area under the receiver operating characteristic curve (AUROC).\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe prevalence of low muscle mass was 30% in the derivation set and 29% in the external validation set. The multimodal model achieved high accuracy in ALM estimation, with an RMSE of 1.23 kg, MAE of 0.96 kg, and correlation coefficient of 0.958 in the internal test set. In external validation, the model yielded an RMSE of 1.87 kg, and MAE of 1.41 kg, and correlation of 0.930. For estimation of appendicular lean mass in the external validation set, the model yielded an AUROC of 0.825 for males and 0.730 for females. Sensitivity analyses demonstrated the model's robustness to missing modalities, maintaining stable prediction errors in participants missing part of the EHR (RMSE: 2.00 kg), frontal (RMSE: 1.99 kg), or lateral (RMSE: 2.08 kg) chest radiographs.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eOur transformer-based multimodal AI model accurately estimates ALM and detects low muscle mass from routinely collected clinical data, outperforming unimodal approaches. The model demonstrated robustness even with missing data modalities, supporting its potential utility as a screening tool in clinical settings where data completeness varies. This approach has the potential to serve as an accessible screening tool for low muscle mass, especially in settings where DXA is not readily available.\u003c/p\u003e","manuscriptTitle":"Transformer-based multimodal model for estimation of appendicular lean mass using incomplete chest radiographs and electronic health record","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-22 07:15:35","doi":"10.21203/rs.3.rs-8556884/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2026-01-19T15:15:45+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-19T14:50:50+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-16T05:08:12+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Translational Medicine","date":"2026-01-13T22:49:46+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-translational-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jtrm","sideBox":"Learn more about [Journal of Translational Medicine](http://translational-medicine.biomedcentral.com)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/jtrm/default.aspx","title":"Journal of Translational Medicine","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b580989a-f09b-478f-9a20-4faa05fa0baf","owner":[],"postedDate":"January 22nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-07T16:08:32+00:00","versionOfRecord":{"articleIdentity":"rs-8556884","link":"https://doi.org/10.1186/s12967-026-08079-0","journal":{"identity":"journal-of-translational-medicine","isVorOnly":false,"title":"Journal of Translational Medicine"},"publishedOn":"2026-04-02 15:59:30","publishedOnDateReadable":"April 2nd, 2026"},"versionCreatedAt":"2026-01-22 07:15:35","video":"","vorDoi":"10.1186/s12967-026-08079-0","vorDoiUrl":"https://doi.org/10.1186/s12967-026-08079-0","workflowStages":[]},"version":"v1","identity":"rs-8556884","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8556884","identity":"rs-8556884","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}