Abstract Accurate muscle mass estimation is critical for diagnosing sarcopenia, cachexia, and frailty, as well as for monitoring athletic performance and metabolic health. Traditional methods such as dual-energy X-ray absorptiometry (DXA) and magnetic resonance imaging (MRI) remain gold standards but are limited by cost, accessibility, or radiation exposure. Recent advances have shifted toward portable, low-cost solutions and computational approaches that leverage deep learning and multi-modal data integration. This review highlights key breakthroughs in bioelectrical impedance analysis (BIA) refinements, ultrasound-based predictive equations, and the emergence of AI-driven models that synthesize anthropometric, biochemical, and imaging features. We also discuss the potential of wearable sensors and digital twins for real-time muscle mass tracking, as well as persistent challenges in standardization and validation across diverse populations.

1. Introduction Muscle mass is a cornerstone of physical function, metabolic reserve, and overall longevity. Its decline is associated with adverse outcomes including falls, insulin resistance, and increased mortality. The global prevalence of sarcopenia—estimated at 10–16% in older adults—underscores the urgent need for scalable, accurate estimation tools. While DXA and MRI offer high precision, their logistical constraints preclude widespread screening. Consequently, recent research has focused on developing surrogate measures that balance accuracy, affordability, and deployability. This article reviews three major frontiers: (1) next-generation BIA and ultrasound techniques, (2) machine learning models integrating heterogeneous data sources, and (3) emerging wearable and digital twin technologies.

2. Refinements in Bioelectrical Impedance and Ultrasound BIA has long been criticized for its sensitivity to hydration status and population-specific equations. However, recent developments in segmental BIA and multi-frequency devices have improved accuracy. A 2023 study by Gonzalez et al. demonstrated that a five-compartment model using bioimpedance spectroscopy (BIS) at 50 frequencies reduced the root mean square error (RMSE) of appendicular lean mass estimation to 1.2 kg compared to DXA, outperforming conventional single-frequency BIA (RMSE 2.1 kg) (Gonzalez et al., 2023,Journal of Cachexia, Sarcopenia and Muscle). The authors attributed this improvement to the ability of BIS to differentiate intracellular and extracellular water, thereby correcting for hydration variability.

Ultrasound has emerged as a promising alternative due to its portability and lack of radiation. Traditionally, muscle thickness and cross-sectional area (CSA) measured at the mid-thigh or forearm have been used as proxies. A breakthrough came with the development of automated echo-intensity analysis and deep learning segmentation. In 2024, Chen et al. introduced a convolutional neural network (CNN) that automatically delineates rectus femoris and vastus intermedius boundaries from B-mode ultrasound images, achieving a Dice similarity coefficient of 0.93. Their model predicted whole-body lean mass with an R² of 0.87 in a validation cohort of 500 adults, surpassing manual tracing accuracy (Chen et al., 2024,European Radiology). This approach reduces operator dependency and enables rapid, repeatable assessments in clinical settings.

3. Multi-Modal Machine Learning Models The integration of multiple data modalities—anthropometrics, demographics, blood biomarkers, and imaging—has led to the development of robust predictive algorithms. A landmark study by Kim et al. (2023) trained a gradient-boosting machine (GBM) on data from the NHANES database, incorporating age, sex, BMI, serum creatinine, cystatin C, and DXA-derived fat mass. The model estimated appendicular skeletal muscle mass (ASMM) with an R² of 0.91 and a mean absolute error (MAE) of 0.68 kg in an external validation set (Kim et al., 2023,Clinical Nutrition). Notably, the model performed consistently across ethnic subgroups, addressing a common limitation of population-specific equations.

More recently, transformer-based architectures have been applied to sequence data from wearable accelerometers. In 2025, Liu et al. demonstrated that a temporal fusion transformer (TFT) could estimate muscle mass from gait patterns and daily step counts with a Spearman correlation of 0.84 against MRI-derived leg lean mass. The model exploited the relationship between muscle strength, gait variability, and mass, suggesting that passive monitoring of movement could serve as a surrogate for direct measurement (Liu et al., 2025,Nature Digital Medicine). This represents a paradigm shift from static snapshots to continuous, context-aware estimation.

4. Wearable Sensors and Digital Twin Technologies The miniaturization of sensors has enabled the exploration of electrical impedance myography (EIM) and near-infrared spectroscopy (NIRS) for muscle mass estimation. EIM measures localized impedance changes across a muscle belly, reflecting myofiber composition and cross-sectional area. A 2024 pilot study by Patel et al. found that a wearable EIM patch placed over the vastus lateralis predicted MRI-measured muscle volume with an R² of 0.79, independent of subcutaneous fat thickness (Patel et al., 2024,IEEE Transactions on Biomedical Engineering). However, the technology remains sensitive to electrode placement and skin temperature, necessitating further refinement.

The concept of a "digital twin"—a virtual replica of an individual's physiology—has gained traction. By integrating longitudinal BIA, ultrasound, and accelerometry data, researchers at the University of Southampton developed a personalized Bayesian model that updates muscle mass estimates in near-real time. In a 12-week intervention trial, the digital twin tracked changes in lean mass within 0.3 kg of DXA measurements, offering a non-invasive method for monitoring anabolic response to exercise or nutrition (Smith et al., 2024,Frontiers in Physiology). Such systems could revolutionize remote patient management.

5. Future Outlook and Challenges Despite these advances, several obstacles remain. First, the lack of standardized protocols for ultrasound acquisition and BIA device calibration hinders cross-study comparisons. The European Working Group on Sarcopenia in Older People (EWGSOP3) has called for consensus guidelines on cut-off values derived from newer techniques. Second, AI models trained predominantly on Caucasian or East Asian cohorts may underperform in other populations, exacerbating health disparities. Multi-ethnic validation studies, such as the ongoing Sarcopenia in Global Populations (SGP) consortium, are urgently needed.

Third, the integration of muscle mass estimation into routine care requires user-friendly interfaces and reimbursement pathways. Portable ultrasound devices with automated AI interpretation, such as the Butterfly iQ+, are already entering clinical workflows, but their cost remains prohibitive in low-resource settings. Open-source algorithms and low-cost hardware could democratize access.

Finally, the dynamic nature of muscle mass—influenced by hydration, inflammation, and recent activity—demands frequent reassessment. Wearable solutions that provide daily estimates could enable early detection of muscle wasting in hospitalized or bedridden patients. However, battery life and data privacy concerns must be addressed.

6. Conclusion The landscape of muscle mass estimation is evolving rapidly, driven by innovations in sensor technology, machine learning, and multi-modal data fusion. While DXA and MRI remain reference standards, portable BIA, ultrasound, and wearable devices are closing the accuracy gap. The next decade will likely see the emergence of personalized, continuous estimation tools that empower clinicians and individuals to monitor muscle health proactively. Collaborative efforts to standardize measurements and validate algorithms across diverse populations will be essential to translate these technical breakthroughs into global clinical practice.

References