Appendicular skeletal muscle mass, or ASMM, refers to the muscle contained in the arms and legs and is a central criterion in the diagnosis of sarcopenia, the age-related loss of muscle mass and function. Sarcopenia is linked to frailty, disability, falls, poorer surgical outcomes and increased mortality, and consensus definitions from European, Asian, Australasian and international working groups all rely on quantifying low muscle mass as a core diagnostic element. Dual-energy X-ray absorptiometry, or DXA, is widely regarded as a high-quality reference method for measuring ASMM, but it is expensive, immobile and exposes patients to a small dose of ionizing radiation. BIA, by contrast, is inexpensive, quick, non-invasive and genuinely bedside-accessible, which explains its popularity in geriatric practice. The catch is that BIA does not measure muscle directly; it estimates body composition from the electrical resistance and reactance of body tissues, using prediction equations that were typically developed in relatively healthy populations.

That reliance on equations derived from healthy subjects is precisely where the problem arises. In people with multimorbidity, altered hydration, edema, inflammation or other disturbances of body water distribution, the assumptions underlying BIA equations break down, and the resulting estimates of muscle mass can be substantially wrong. Clinicians have long known this in general terms, but they have had no practical way of knowing, for a specific patient sitting in front of them, whether the number on the BIA printout is close to the truth or off by several kilograms. The Padua study set out to address exactly this gap, asking whether machine learning algorithms could serve as a decision-support layer that tells the clinician when a BIA-derived ASMM estimate is sufficiently reliable to use in clinical reasoning.

The study enrolled 701 participants aged 65 years or older, comprising 550 healthy subjects and 151 outpatients recruited through geriatric services. All participants underwent both BIA and DXA on the same assessment protocol, allowing the researchers to compare the BIA-estimated ASMM against the DXA-measured reference value for each individual. An estimate was classified as accurate if its absolute error was 1.14 kilograms or less, a threshold chosen because it corresponds to the standard error of the original BIA prediction equation used. In other words, if the BIA equation was performing as well as it was designed to, the error should fall within that band; anything larger represents a genuine failure of the estimate rather than ordinary measurement noise.

The first key finding confirmed the scale of the problem. In healthy subjects, the median absolute difference between BIA-estimated and DXA-measured ASMM was 0.81 kilograms, and 62.4 percent of estimates fell within the acceptable error threshold. In outpatients, the median absolute difference more than doubled to 2.37 kilograms, a statistically significant difference, and the proportion of accurate estimates collapsed to just 25.2 percent. Expressed as a percentage of the true value, the mean absolute percentage error was 1.74 percent in healthy subjects versus 4.12 percent in outpatients. The researchers also found that comorbidity burden, assessed with the Cumulative Illness Rating Scale, was independently associated with estimation inaccuracy, meaning that the sicker the patient, the less the BIA equation could be trusted, even after accounting for other characteristics.

To build the decision-support layer, the team trained five machine learning algorithms: Extreme Gradient Boosting, Support Vector Machine (SVM), Random Forest, logistic regression, and neural networks. The task given to each algorithm was classification: given a set of patient characteristics, predict whether that patient’s BIA-derived ASMM estimate would be accurate by the 1.14-kilogram criterion. The algorithms were trained using three hierarchical sets of predictors. Set 1 contained anthropometric and bioimpedance variables, the basic information available from a standard BIA assessment. Set 2 added handgrip strength, a simple bedside measure of muscle function. Set 3 further added anthropometric circumferences of the arm, waist and calf, along with knee height, measurements that require nothing more than a tape measure and a caliper and are easily obtained in any clinic.

The results showed a clear progression in performance as predictors were added. Discriminative performance, measured by the area under the receiver operating characteristic curve (AUC), improved with each successive predictor set, indicating that functional and anthropometric information carries real signal about when BIA equations fail. In the fullest model, Set 3, the support vector machine achieved the best performance, with a cross-validation AUC of 0.813 and a test AUC of 0.867, together with an accuracy of 0.70. An AUC of 0.867 indicates good discrimination, meaning the model could separate patients with trustworthy BIA estimates from those with unreliable ones substantially better than chance. The SVM outperformed the other four algorithms in this configuration, though the study evaluated all five across all three predictor sets.

The practical implication is a workflow that geriatricians could apply at the bedside. A clinician measures a patient’s height, weight, bioimpedance, handgrip strength and a few circumferences, feeds these into the trained model, and receives a probability that the BIA-derived ASMM value is accurate. If the model indicates high reliability, the BIA estimate can be used directly for sarcopenia screening and diagnosis. If it indicates low reliability, the clinician knows the estimate should be treated with caution and, where the diagnosis would change management, confirmed with DXA or another reference method. This converts what has historically been an invisible, unquantifiable source of error into an explicit, standardized accept-or-reject decision, potentially sparing many patients unnecessary imaging while protecting those in whom the estimate would be misleading.

The study has limitations worth noting. It was cross-sectional, so the findings describe associations at a single time point rather than validating the approach longitudinally. The cohort was drawn from a single Italian center and consisted of free-living Caucasian older adults, raising questions about generalizability to other ethnic groups, other healthcare settings, hospitalized patients or populations with different body composition norms. The outpatient group, though clinically essential to the study’s purpose, was considerably smaller than the healthy group, at 151 versus 550 participants, which may affect how well the models capture the full heterogeneity of comorbid patients. The accuracy threshold of 1.14 kilograms, while principled, is itself derived from the original equation’s standard error, and different thresholds would yield different classification results. The reported test accuracy of 0.70 for the best model, while respectable, also means that roughly three in ten classifications are incorrect, so the tool should be understood as a support for clinical judgment rather than a replacement for it.

Even with these caveats, the work sits within a growing body of research applying machine learning to body composition assessment. Recent studies have used machine learning to predict body fat percentage from simple anthropometric measurements, to model lean body mass and appendicular skeletal muscle mass in large datasets such as NHANES and the Look AHEAD study, to predict low muscle mass in patients with obesity and diabetes, and to automate body composition analysis of clinically acquired computed tomography scans using neural networks. What distinguishes the Padua study is its focus not on predicting muscle mass itself, but on predicting the reliability of an existing, widely used bedside estimate, a framing that directly targets the clinical decision the clinician actually faces.

The implications for sarcopenia care could be meaningful. Because BIA is one of the few muscle mass assessment tools that can be deployed in nursing homes, outpatient clinics, emergency departments and patients’ homes, improving the trustworthiness of its output, or at least quantifying when it should be doubted, could extend structured muscle mass assessment to settings where DXA is unavailable. This is particularly relevant for older adults with multimorbidity, who are precisely the group at highest sarcopenia risk and in whom BIA is least reliable. A standardized framework for accepting or rejecting bedside estimations could also improve the consistency of sarcopenia diagnosis across centers, support better-informed nutritional and pharmacological interventions, and contribute to more accurate dosing decisions in contexts where lean body mass influences drug distribution.

The authors, a collaboration between the University of Padua’s Department of Medicine and Geriatric Unit and the university’s Department of Management and Engineering, suggest that combining BIA with machine learning provides a practical, bedside-applicable framework for determining whether an individual ASMM estimate can be trusted. Before such models enter routine practice, external validation in independent and more diverse cohorts will be needed, along with studies demonstrating that using the model actually improves diagnostic accuracy and patient outcomes compared with current practice. But the study offers a concrete demonstration that the accuracy problem of bioimpedance in comorbid older adults is not merely an unavoidable limitation; it is, at least in part, a predictable one, and predictability is something machine learning is well placed to exploit.