Urine Metabolomic Profiling and Machine Learning in Autism Spectrum Disorder Diagnosis: Toward Precision Treatment.

Abstract

Background: Autism spectrum disorder (ASD) diagnosis traditionally relies on behavioral assessments, which can be subjective and often lead to delayed identification. Recent advances in metabolomics and machine learning offer promising alternatives for more objective and precise diagnostic approaches.

Methods: First-morning urine samples were collected from 52 children (32 with ASD and 20 neurotypical controls), aged 5.04 ± 1.87 and 5.50 ± 1.74 years, respectively. Using liquid chromatography-mass spectrometry (LC-MS), 293 metabolites were identified and categorized into 189 endogenous and 104 exogenous metabolites. Various machine learning classifiers (random forest, logistic regression, random tree, and naïve Bayes) were applied to differentiate ASD and control groups through 10-fold cross-validation.

Results: The random forest classifier achieved 85% accuracy and an area under the curve (AUC) of 0.9 using all 293 metabolites. Classification based solely on endogenous metabolites yielded 85% accuracy and an AUC of 0.86, whereas using exogenous metabolites alone resulted in lower performance (71% accuracy and an AUC of 0.72).

Conclusion: This study demonstrates the potential of urine metabolomic profiling, particularly endogenous metabolites, as a complementary diagnostic tool for ASD. The high classification accuracy highlights the feasibility of developing assistive diagnostic methods based on metabolite profiles, although further research is needed to link these profiles to specific behavioral characteristics and ASD subtypes.