A classification model for predicting corticosteroid and cyclosporin A responsiveness in pediatric idiopathic uveitis

To identify serum metabolic biomarkers that distinguish corticosteroid and cyclosporin A (CS & CsA) resistant pediatric idiopathic uveitis (PIU) patients from sensitive counterparts.

Serum samples were collected from 32 CS & CsA-sensitive PIU patients and 24 CS & CsA-resistant PIU patients, respectively. UHPLC-OE-MS was employed for comprehensive metabolic profiling of the serum samples. Bioinformatic analyses were performed to identify differentially expressed metabolites (DEMs) between the two patient groups. A machine learning-based classification model was constructed using the identified DEMs as predictive features. For validation purposes, an independent internal cohort of 16 CS & CsA-sensitive and 10 CS & CsA-resistant patients was recruited to evaluate the model's stability.

Compared with the CS & CsA-sensitive PIU patients, serum samples from CS & CsA-resistant PIU patients displayed significant metabolic reprogramming. Among the identified differential metabolites, lipids were the most prominently dysregulated class, accounting for 72.47% of all differential metabolites. A machine learning based multivariate feature selection approach including NNET, LASSO, and XGBoost identified 4 candidate metabolite biomarkers. ROC analysis showed that three of these biomarkers (MG 15:0, PI-Cer 28:0;3 O, and SPB 20:0;2 O) exhibited AUC values of 0.934, 0.953, and 0.904, respectively, and were all upregulated in CS & CsA resistant patients. In contrast, N-acetylaspartic acid showed an AUC of 0.934 and was downregulated in CS & CsA resistant patients. The combined classification model incorporating these 4 metabolites achieved an AUC of 1.0. Validation in an independent internal cohort confirmed the model's excellent performance, with AUC values of 0.971 for NNET, 0.971 for LASSO, and 0.957 for XGBoost.

We have established a classification model capable of effectively discriminating CS & CsA-resistant from -sensitive PIU patients. The machine learning model leveraging metabolic biomarkers demonstrates exceptional classification accuracy and generalizability, offering potential for clinical subtype classification.