Deep-learning analysis suggests a facial image signal associated with low muscle mass in a clinical cohort enriched for chronic liver disease: a proof-of-concept study

Source abstract

Background Sarcopenia is a clinically relevant complication of chronic liver disease. Because changes in muscle mass, adiposity, and nutritional status may affect facial morphology, this exploratory study examined whether standardized frontal facial photographs contain an algorithmically detectable signal associated with low measured muscle mass. Methods Prospectively collected clinical data and facial photographs from 255 participants were retrospectively analyzed using the 2010 European Working Group on Sarcopenia in Older People (EWGSOP1) framework. The binary target was low versus normal measured muscle mass. Stage 1 developed ImageNet-pretrained VGG16, MobileNetV2, and DenseNet121 models in 204 participants and evaluated them in a 51-participant internal hold-out set. Stage 2 independently redeveloped the task using four region-specific EfficientNet-B0 and DenseNet121 components integrated into a single decision pipeline and evaluated it separately within the same source cohort. A clinical-variable-only comparator, human image assessment, Grad-CAM, and controlled perturbation analyses were also performed. Results Overall, 173 participants had low and 82 had normal muscle mass. Stage-1 balanced accuracy was 0.71 for VGG16, 0.70 for DenseNet121, and 0.63 for MobileNetV2 The integrated stage-2 system showed high sensitivity (94.3%) but limited specificity (50.0%), resulting in a balanced accuracy of 0.72. Human image assessment showed low interrater agreement. Grad-CAM and perturbation analyses indicated contributions from facial regions but also substantial dependence on peripheral and non-face image content. The clinical-variable comparator outperformed the stage-1 facial models. Conclusion Standardized facial photographs contained a image-derived information associated with low measured muscle mass across two independently implemented modelling stages. The integrated stage-2 pipeline achieved high sensitivity, supporting further investigation of facial image analysis as a low-burden pre-screening approach. However, limited specificity, potential clinical and diagnostic-pathway confounding, dependence on non-face image content, and the absence of external validation indicate that these findings remain exploratory. Prospective external validation and assessment of incremental value beyond routine clinical variables are required before clinical implementation.