Can artificial intelligence simplify the screening of muscle mass loss?

Background: Sarcopenia is a risk factor for morbidity and preventable mortality in old age, with consequent high costs for the national health system. Its diagnosis requires costly radiological examinations, such as the DEXA, which complicate screening in medical centers with a high prevalence of sarcopenia. Objectives: Developing a nearly zero-cost screening tool to emulate the performance of DEXA in identifying patients with muscle mass loss. This can crucially help the early diagnosis of sarco-penia at large-scale, contributing to reduce its prevalence and related complications with timely treatments. Methods: We exploit cross-sectional data for about 14,500 patients and 38 non-laboratory vari-ables from successive NHANES over 7 years (1999-2006). Data are analyzed through a state-of-the-art artificial intelligence approach based on decision trees. Results: A reduced number of anthropometric parameters allows to predict the outcome of DEXA with AUC between 0.92 and 0.94. The most complex model derived in this paper exploits 6 variables, related to the circumference of key corporal segments and to the evaluation of body fat. It achieves an optimal trade-off sensitivity of 0.89 and a specificity of 0.82. Restricting exclusively to variables related to lower limb, we obtain an even simpler tool with only slightly lower ac-curacy (AUC 0.88-0.90). Conclusions: Anthropometric data seem to contain the entire informative content of a more complex set of non-laboratory variables, including anamnestic and/or morbidity factors. Compared to previously published screening tools for muscle mass loss, the newly developed models are less complex and achieve a better accuracy. The new results might suggest a possible inversion of the standard diagnostic algorithm of sarcopenia. We conjecture a new diagnostic scheme, which requires a dedicated clinical validation that goes beyond the scope of the present study