Identifying Key Predictors of Sarcopenic Obesity in Italian Severely Obese Older Adults: Deep Learning Approach

Background/Objectives: Sarcopenic obesity (SO), the coexistence of sarcopenia and obesity, poses serious health risks, such as increased mortality. Despite its clinical significance, key predictors of SO remain unclear, especially in severe obesity. This study aimed to identify independent predictors of SO in Italian older adults with obesity using a deep learning neural network. Methods: A cross-sectional study was conducted with hospitalized older adults diagnosed with severe obesity. SO was defined according to the 2022 ESPEN/EASO Statement Criteria, based on skeletal muscle function assessed by the five-repetition sit-to-stand test (5-SST) and body composition parameters evaluated using Dual X-ray Absorptiometry. A total of 42 independent variables were analyzed. Data normalization was performed using MinMaxScaler, and an optimal neural network architecture was selected via grid search with stratified 5-fold cross-validation. Model performance was assessed using accuracy, precision, recall, F1-score, AUC-ROC, and AUPRC metrics. Results: The correlation analysis revealed strong negative associations between SO and handgrip strength (HGS) (r = -0.785) and appendicular lean mass (ALM) (r = -0.745), as well as moderate correlations with 5-SST (r = 0.603), 30-second chair stand test (r = -0.474), 6-minute walking test (6m-WT) (r = 0.289), and waist circumference (WC) (r = 0.127). The deep learning model achieved an average classification accuracy of 72%, with a precision of 83% and an AUC of 0.9333. Conclusions: The main key predictors of SO were HGS, ALM, 5-SST, 30s-SST, 6m-WT, and WC in the early detection of this condition. The findings highlight deep learning's potential to improve SO diagnosis, risk assessment, clinical decision-making, and prevention in severely obese older adults.