This study was approved by the institutional ethics committees of the participating centers (Shanghai Tenth People’s Hospital, West China Hospital of Sichuan University, Zhongshan Hospital of Fudan University; SHSY-IEC-5.0/23K56/P01). All participants provided written informed consent before enrollment. Participants were informed of the study’s purpose, procedures, potential risks, and benefits. Privacy and confidentiality were strictly protected throughout the study; all data were anonymized before analysis and stored securely. No monetary or material compensation was provided to participants for their involvement in the study. The research protocol was registered at the Chinese Clinical Trial Registry (ChiCTR2300073651).

This was a prospective multicenter study. This study was designed and reported in accordance with the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) guidelines [17]. In this study, we developed three monomodal models, two bimodal models, and one multimodal model based on ultrasound images, clinical information, and laboratory information.

First, three candidate monomodal models were developed and evaluated. The monomodal model with the best performance was selected as the basis. Second, we combined the best monomodal model with one of the other two data types to develop two candidate bimodal models. Finally, we combined the three data types to develop one multimodal model. After 5-fold cross-internal validation, the diagnostic performance of the candidate monomodal, bimodal, and multimodal models were compared. Moreover, the relative contributions of each data type to the overall classification results were also analyzed. Finally, the optimal candidate model was determined as the SARCO model by comprehensive assessment. The SARCO model was performed with external validation and proof-of-concept validation. Figure 1 illustrates a comprehensive design of the study.

The population characteristics of the different cohorts are shown in Figure 1A. From October 2021 to December 2022, a total of 2735 ultrasound images and the associated clinical data from 357 participants were collected for training and internal validation in Shanghai Tenth People’s Hospital. From May 2023 to December 2023, a total of 352 ultrasound images from 145 participants were prospectively collected for external validation of the model in the West China Hospital, Sichuan University. The inclusion criteria were as follows: (1) aged 60 years or older and (2) participants with suspected sarcopenia, for example, functional decline or limitation, unintentional weight loss, and diabetes mellitus [1]. The exclusion criteria were as follows: (1) inability to complete muscle strength and physical performance assessments and (2) not agreeing to participate in this study.

From October 2023 to January 2024, a total of 817 ultrasound images from 82 participants were prospectively collected for proof-of-concept validation of the model in Zhongshan Hospital, Fudan University. The inclusion criteria were as follows: (1) aged 60 years and older; (2) participants with decompensated cirrhosis [18]; (3) inability to complete muscle strength and physical performance assessments due to weakness, bedridden, or other bodily dysfunction; and (4) available abdominal computed tomography (CT) scan from which an adequate image at the L3 level could be obtained. The exclusion criteria were as follows: (1) without abdominal CT scan or poor image quality and (2) not agreeing to participate in this study. The flow diagram of the inclusion and exclusion is shown in Multimedia Appendix 1.

The assessment of participants in Shanghai Tenth People’s Hospital and West China Hospital was based on the consensus of the Asian Working Group for Sarcopenia, where the reference criteria for the diagnosis of sarcopenia include low muscle mass, low muscle strength, and/or low physical performance [1]. Extremity muscle mass was determined using the logic Discovery Wi dual-energy x-ray bone densitometer (Hologic Discovery QDR Series) or the Inbody 770 body composition analyzer. The Appendicular Skeletal Muscle Mass Index was calculated using the following formula:

Appendicular Skeletal Muscle Mass Index=extremity muscle mass (kg)/height squared (m²)

Handgrip strength (kg) was measured by the Camry electronic grip strength meter (EH101; Camry Scale). Physical performance was assessed by 6-meter walking speed (m/s) and 5-time chair stand test (s). The participants at Zhongshan Hospital were assessed for sarcopenia using the L3-skeletal muscle index by CT. The cutoff values for diagnosing sarcopenia were 44.77 cm²/m² for men and 32.50 cm²/m² for women [18].

The clinical information included the presence of diabetes, the presence of osteoporosis, age (years), gender, height (m), weight (kg), BMI (kg/m²), and calf circumference (cm). The laboratory information was obtained before muscle strength and physical function assessments. The laboratory information included total cholesterol (mmol/L); triglycerides (mmol/L); high-density lipoprotein (mmol/L); low-density lipoprotein (mmol/L); glycosylated hemoglobin (%); albumin (g/L); hemoglobin (g/L); 1,25(OH)2D (ng/mL); osteocalcin (ng/mL); free triiodothyronine (pmol/L); free tetraiodothyronine (pmol/L); total triiodothyronine (nmol/L); total tetraiodothyronine (nmol/L); and thyroid-stimulating hormone (mIU/L).

The ultrasound equipment and transducer of the centers are shown in Multimedia Appendix 2. The ultrasound equipment in the training and internal validation cohort was the Aixplorer Ultrasound system (SuperSonic Imagine), with an SL 10-2 multifrequency linear transducer. In the external validation cohort, the ultrasound equipment was the Aixplorer Ultrasound system (SuperSonic Imagine) with an SL 10-2 multifrequency linear transducer, the Mindray M9 Portable Ultrasound Machine (Mindray) with an SL 10-3 multifrequency linear transducer, and the Mindray MX7 Portable Ultrasound Machine (Mindray) with an SL 13-3 multifrequency linear transducer. In the proof-of-concept cohort, the ultrasound equipment was the R10 Prestige ultrasound system (Samsung Medison Co. Ltd) with an SL 12-3 multifrequency linear transducer. All participants underwent ultrasound examination with radiologists with 5 years or more of experience in musculoskeletal ultrasound. The radiologists reached a consensus on the examination procedure.

The participants were placed in the supine position. The rectus femoris (RF) was chosen for ultrasound examination, with the transducer positioned at the middle position between the anterior superior iliac spine and the lower edge of the patella. The musculoskeletal mode was preset, and the gain, depth, and focus were adjusted to visualize the muscle properly. The transducer was oriented perpendicularly to the long axis of the RF to capture three to five cross-sectional images of RF. The ultrasound images were saved in DICOM or JPG format for further processing and analysis.

All experiments were conducted using Python 3.8.3 and Pytorch 1.7.0 on a 64-bit Ubuntu 18.04.4 LTS operating system. We used two NVIDIA TITAN graphics processing units with 24 GB of memory for training. The batch size used was 48, with a 3×10^–4 learning rate. We used a warm-up strategy at the beginning of the training and the CosineAnnealingLR learning rate decay strategy. The maximum number of epochs was set as 150. AdamW was used as an optimizer, and the loss function we chose is binary cross-entropy.

In order to minimize experimental errors arising from discrepancies in data distribution, we adopted a 5-fold cross-validation approach to evaluate the performance of CNN models. The data were partitioned at the participant level into five subsets. Subsequently, four subsets were designated as the training set, with the remaining one as the validation set. Recognizing the imbalance between positive and negative images, we introduced random down-sampling of negative images within the training data, ensuring parity with the number of positive images. This balanced dataset was then used to train the model. Finally, we applied a model ensemble technique to merge these three sets of trained weights, creating an integrated model. This integrated model was then used to evaluate performance on the validation set.

To better interpret the CNN prediction results, the gradient-weighted class activation mapping method was used to generate heat maps. Heat maps can visualize the image’s most indicative areas to interpret the CNN models’ predictive mechanism, which reflects the contribution of each pixel in the ultrasound images to the prediction of sarcopenias. All heat maps were produced by applying the packages Matplotlib v3.8.0 and TorchCAM v0.3.2.

All statistical analyses were conducted using Python (version 3.6.13; Python Software Foundation), sci-kit-learn (version 0.20.0), and SPSS (version 26.0; developed by IBM Corp). Data were presented as mean (SD), median (IQR), or number (%) as appropriate. Comparisons of differences in clinical factors between groups with or without sarcopenia were analyzed using the independent 2-sample t test, Wilcoxon rank-sum test, or chi-square test. The receiver operating characteristic (ROC) curves were constructed to assess the prediction performance of models. The area under the ROC curve (AUC) with 95% CIs, sensitivity, specificity, and accuracy were investigated. The F₁-score was also calculated:

The Delong test was performed to compare the AUCs. Statistical significance was set at 2-sided P<.05. To comprehensively assess the impact of multimodal data on classification outcomes, we used the PyTorch Shapley additive explanations (SHAP) library for a detailed statistical analysis. Following model training, SHAP values were computed using the PyTorch SHAP library. The SHAP values were analyzed to understand the relative contributions of each modality to the overall classification results. By comparing the magnitudes of SHAP values across modalities, we quantified the influence of ultrasound images, clinical information, and laboratory data on the model’s decision-making process.

We enrolled a total of 357 participants (median age 67.00, IQR, 63.00-72.00 years), including 101 participants with sarcopenia (608 ultrasound images) and 256 participants without sarcopenia (2127 ultrasound images), in the training and internal validation cohort. In the external validation cohort, we enrolled a total of 145 participants (median age 71.00, IQR 67.00-77.00 years), including 68 participants with sarcopenia (70 ultrasound images) and 77 participants without sarcopenia (252 ultrasound images). In the proof-of-concept group, we enrolled a total of 82 participants (median age 66.00, IQR 63.00-71.00 years), including 19 participants with sarcopenia (185 ultrasound images) and 63 participants without sarcopenia (632 ultrasound images).

The clinical and laboratory information of the training and internal validation cohort is shown in Table 1. The age of participants with sarcopenia was older than that of those without sarcopenia (median age 72.00, IQR 66.00-80.50 vs 66.00, IQR 62.00-69.75 years; P<.001). Participants with sarcopenia had lower BMI, reduced calf circumference, and a lower prevalence of diabetes (all P<.05). Regarding laboratory information, patients with sarcopenia demonstrated lower levels of albumin, hemoglobin, and free triiodothyronine but higher levels of thyroid-stimulating hormone (all P<.05).

The sensitivity analysis was conducted to evaluate the potential impact of imputing missing values. A total of 86 cases with missing data were excluded, and model analysis was performed on the remaining 271 cases with complete data. The number of cases with missing features is summarized in Multimedia Appendix 7. The diagnostic performance of three monomodal models based on complete data is shown in Table 3. The clinical information model achieved an AUC of 0.787 and an F₁-score of 0.630, and the laboratory information model achieved an AUC of 0.666 and an F₁-score of 0.480. The ultrasound images model achieved the highest AUC of 0.851 and F₁-score of 0.698. The diagnostic performance of the monomodal, bimodal, and multimodal models based on complete data is shown in Table 4. Although the multimodal CNN model achieved the best diagnostic performance statistically (AUC=0.883; F₁-score=0.728), SHAP analysis revealed that ultrasound images still contributed the most to diagnosing sarcopenia (0.810) and nonsarcopenia (0.795). Finally, we selected the monomodal model based on only ultrasound images as the optimal one for diagnosing sarcopenia and named it the SARCO model.

The SARCO model achieved excellent prediction performance in external validation, with sensitivity=0.705, specificity=0.794, accuracy=0.753, AUC=0.801, and F₁-score=0.727 (Figure 5A). The SARCO model also achieved satisfactory prediction performance in proof-of-concept validation, with sensitivity=0.789, specificity=0.825, accuracy=0.817, AUC=0.757, and F₁-score=0.666 (Figure 5B).

The corresponding heat maps of the ultrasound images of the SARCO model are presented in Figure 6. The different color distributions in the heat maps indicate the models’ emphasis on the most predictive regions and image features in the context of sarcopenia. The red parts of the heat map indicate that those parts provide more informant features during the network’s predictive process. In contrast to nonsarcopenia (Figure 6A and B), the hyperechoic areas due to muscle fascia or fibrosis involvement are characterized by high activation (Figure 6C and D).

In this study, we successfully developed several models using ultrasound images, clinical information, and laboratory information to diagnose sarcopenia. As expected, our findings showed that the multimodal model statistically demonstrated the highest diagnostic performance. Additionally, we noted that ultrasound images played a dominant role compared to clinical and laboratory information in the diagnostic decision-making process. Accordingly, the absolute difference in performance between the monomodal model based on ultrasound images and the multimodal model was minimal. Therefore, considering the accessibility of ultrasound examination, the monomodal model based on ultrasound images was considered the most practiced solution for screening sarcopenia, despite it resulting in some accuracy losses.

The pathological changes of sarcopenia are the foundation for the macroscopic changes in muscle [7,19]. Macroscopic changes encompass alterations in muscle thickness, cross-sectional area, and hardness, all of which can be further reflected in the ultrasound images [19,20]. Muscle thickness and cross-sectional area not only reflect muscle mass but also show a positive correlation with muscle strength and physical function, facilitating the diagnosis of sarcopenia [21,22]. For example, Chen et al [23] developed a diagnostic model for sarcopenia based on age, cross-sectional area changes, and shear wave elasticity value changes, yielding an AUC of 0.883. Tang et al [24] developed an ultrasound-derived muscle assessment system based on muscle thickness, handgrip strength, and gait speed. The system had an overall diagnostic sensitivity of 92.7% and a specificity of 91% for sarcopenia. These studies have proved the effectiveness of conventional ultrasound methods; however, they are time-consuming and performed in diverse protocols. Moreover, it is challenging to detect minor muscle texture abnormalities in the early phase of sarcopenia through apparent macroscopic muscle morphology by conventional imaging methods.

The minor abnormalities in muscle may be identified using advanced image processing and artificial intelligence (AI) techniques. Early screening of individuals with suspected muscle texture abnormalities enables timely referral to hospitals for further evaluation and treatment, thereby optimizing the diagnostic procedure. Recently, some efforts have been made to use AI methods in ultrasound images for sarcopenia assessment. For example, Yang et al [25] demonstrated the feasibility of evaluating muscle mass through quantitative muscle texture features using a gray-level co-occurrence matrix. Wilkinson et al [26] discovered that gray-level co-occurrence matrix parameters extracted from ultrasound images were correlated with muscle function in patients having chronic kidney disease. These methods involving manual or semiautomatic feature extraction of muscle texture features could introduce subjectivity and variability, limiting their generalizability across different datasets [27,28]. CNNs happen to be well suited to address these shortcomings [29,30]. CNNs can automatically handle images from large datasets and various imaging conditions, enhancing their suitability for sarcopenia screening.

Exploration of CNN models for diagnosing sarcopenia based on ultrasound images is currently insufficient. Yi et al [31] used three CNNs (VGG19, ResNet50, and DenseNet121) for predicting sarcopenia, achieving accuracies of 0.65-0.75 (based on grayscale ultrasound images) and 0.70-0.80 (based on shear-wave elastography images). Their study primarily focused on assessing the applicability of different CNNs for predicting sarcopenia based on ultrasound images. In this study, we compared three networks and selected the superior-performing ConvNeXt. ConvNeXt, inspired by ResNeXt, improves feature expression through group convolutions, enhancing network depth and width. It allows for better extracting rich structural and detailed information in ultrasound images of sarcopenia [32].

Compared to other studies, another methodological advantage of this study is integrating clinical and laboratory information with ultrasound images to improve the performance of predicting sarcopenia. As expected, the complementary of different data types typically synergistically enhances disease detection. However, based on our results, the improvement in the performance of multimodal models was slight. Clinical and laboratory information indirectly reflects the pathophysiological changes of sarcopenia and may be influenced by external factors, and thus, may not have enough sensitivity to diagnose chronic conditions. Additionally, in the real world, it is difficult to obtain complete clinical and laboratory information from bedridden or housebound older individuals.

On the contrary, morphological changes, such as reduced muscle mass and muscle fiber atrophy, are the most direct manifestations of sarcopenia and can be visualized through imaging methods. Moreover, ultrasound devices possess the advantages of accessibility, mobility, radiation-free, and cost-effectiveness compared with CT and dual-energy x-ray absorptiometry, making them preferable for community preventive services or chronic health care. Though the performance of the SARCO model in a small-scale, proof-of-concept validation was not excellent enough, considering the participants with decompensated cirrhosis were all in weak, bedridden, or nearly disabled conditions, this study preliminarily confirmed the benefits of the SARCO model, which provides theoretical and practical support for large-scale data verification in the future.

This study represented an initial exploration of the use of ultrasound images to diagnose sarcopenia and an important step toward clinical application. However, there were still some limitations, and the model’s accuracy has room for further improvement. First, the sample size for model training and validation was relatively small, although this study applied cross-validation to mitigate biases and errors caused by the sample distribution. Second, our data were obtained from Chinese single-race cohorts. Third, in order to reduce the influence of other tissues on the feature extraction, we use manual segmentation to separate muscle regions. Fourth, some cases lacked partial clinical or laboratory data. Although our sensitivity analysis demonstrated consistent performance trends, future work should address this issue by collecting a more comprehensive dataset and exploring advanced imputation techniques, such as multiple imputation or machine learning-based methods. Finally, it was better to evaluate our SARCO model on public datasets. Unfortunately, we did not find any database containing muscle ultrasound images of sarcopenia.

With an aging population, sarcopenia diagnosis models hold significant potential for clinical application. Moving forward, we will focus on improving the model’s accuracy and clinical applicability through several steps. First, enhancing data quality, optimizing algorithms, optimizing image processing techniques, and developing automatic segmentation techniques will help improve the model’s accuracy and reliability. Second, incorporating more imaging and clinical data, such as CT data and laboratory biomarkers, will enhance the model’s multidimensional predictive capabilities. Third, expanding the sample size to include diverse populations in terms of age, gender, ethnicity, and health status will ensure broader applicability and generalizability. Fourth, external validation across multiple centers and regions is necessary to assess the model’s performance in different populations and improve its clinical adaptability. Finally, beyond the CNNs applied in this study, advanced AI techniques hold promise for improving sarcopenia assessments. For instance, self-supervised learning can mitigate the limitations of labeled data by leveraging large-scale unlabeled ultrasound images for pretraining, enhancing model generalizability. Federated learning enables multicenter collaboration without compromising data privacy, promoting robust and diverse model training [33]. Additionally, the combination of medical imaging with large language models for contextual understanding and predictive analysis could further enhance personalized sarcopenia risk assessment [34]. These advancements, combined with ongoing improvements in data quality and standardization, and the collaboration of physicians and engineers, have the potential to enhance the accuracy and clinical utility of sarcopenia assessments.

In conclusion, ultrasound images, clinical information, and laboratory information all contributed to diagnosing sarcopenia; however, ultrasound images played a dominant role in model decision-making. Both external and proof-of-concept evaluations have demonstrated that the SARCO model can be a feasible and convenient solution for diagnosing sarcopenia. This advancement holds promise for large-scale sarcopenia screening and health care management of older individuals in clinical or community settings.