Un An Interpretable Multi-Omics ElasticNet Pipeline for Alzheimer's Disease Classification
DOI:
https://doi.org/10.5935/jetia.v12i61.4555Resumen
Alzheimer’s disease (AD) classification from heterogeneous biomedical data remains challenging because omics feature spaces are high-dimensional, sample sizes are limited, and data integration must remain interpretable. This study developed an interpretable machine learning framwork that integrated biological interpretation tools to distinguish cognitively normal (CN) from AD participants using Alzheimer’s Disease Neuroimaging Initiative (ADNI) data. Peripheral-blood gene expression, genome-wide SNPs, and clinical/MRI variables were analyzed in two cohorts: a complete-data common cohort (n = 290; 150 CN, 140 AD) and a broader union cohort (n = 578; 284 CN, 294 AD). After upstream quality control and preprocessing, transcriptomic and genetic matrices were reduced in a separate cross-validated stability-selection stage which yielded fixed consensus top-5000 features per modality for downstream benchmarking. Five predictive configurations were then evaluated under stratified 5-fold cross-validation: expression-only, SNP-only, clinical-only, omics-only, and multi-omics. In the common cohort, SNP-only achieved the highest discrimination (AUC = 0.936; AP = 0.923), whereas multi-omics provided the best overall balance of performance (AUC = 0.932; accuracy = 0.890; Brier score = 0.099). In the union cohort, multi-omics performed best overall (AUC = 0.960; AP = 0.966; accuracy = 0.913). Biological interpretation identified recurrent AD-related candidate signals, including IQCK, APH1B, PTGS2, and APOE/APOC1.
Descargas
Descargas
Publicado
Número
Sección
Licencia
Derechos de autor 2026 ITEGAM-JETIA

Esta obra está bajo una licencia internacional Creative Commons Atribución 4.0.








