Abstract
Organic acidemias (OAs) are inherited metabolic disorders requiring accurate profiling of urinary organic acids for diagnosis. Although gas chromatography-mass spectrometry (GC-MS) remains a clinical standard, interpretation is challenging because urinary metabolic profiles are complex and heterogeneous, reflecting disease-related changes and biological variability arising from physiological, dietary, and genetic factors. Together with limited authentic derivatized reference materials, these factors make interpretation labor-intensive and dependent on expert judgment. We developed a feature-based machine-learning framework for OA interpretation under limited-sample conditions. GC-MS chromatographic patterns and metabolite response ratios from 727 urine samples (475 controls and 252 OA cases across 18 categories) were evaluated using convolutional neural networks (CNN), support vector machines (SVM), random forests (RF), and deep neural networks (DNN). All models achieved F1 scores >0.80 for OA discrimination, with DNN showing the highest performance (F1 = 0.84, AUC = 0.85). For the 13-class task, DNN achieved an accuracy of 0.89, F1 score of 0.90, and AUC of 0.98, followed by RF (F1 = 0.88, AUC = 0.97) and SVM (F1 = 0.85, AUC = 0.95). Notably, our framework used only 34 pathway-informed features from 727 samples, whereas previous GC-MS machine-learning studies used 132 metabolites from 531 samples and 135 metabolites from 355,197 samples, reporting accuracies of >0.981 and 0.969, respectively, under different classification settings. These findings demonstrate that a compact, biochemically interpretable feature set can capture disease-associated patterns under data-limited conditions and facilitate consistent GC-MS interpretation, providing complementary support for expert OA diagnosis.