Sarcopenia is associated with increased morbidity and mortality in Crohn's disease. The present study is aimed at investigating the different diagnostic performance of different machine learning models in identifying sarcopenia in Crohn's disease. Patients diagnosed with Crohn's disease at our center provided clinical, anthropometric, and radiological data. The cross-sectional CT slice at L3 was used for segmentation and the calculation of body composition. The prevalence of sarcopenia was calculated, and the clinical parameters were compared. A total of 167 patients were included in the present study, of which 127 (76.0%) were male and 40 (24.0%) were female, with an average age of 36.1 ± 14.3 years old. Based on the previously defined cut-off value of sarcopenia, 118 (70.7%) patients had sarcopenia. Seven machine learning models were trained with the randomly allocated training cohort (80%) then evaluated on the validation cohort (20%). A comprehensive comparison showed that LightGBM was the most ideal diagnostic model, with an AUC of 0.933, AUCPR of 0.970, sensitivity of 72.7%, and specificity of 87.0%. The LightGBM model may facilitate a population management strategy with early identification of sarcopenia in Crohn's disease, while providing guidance for nutritional support and an alternative surveillance modality for long-term patient follow-up.
Keywords: Crohn’s disease; machine learning; sarcopenia.