AD diagnosis model based on fusion of heterogeneous brain imaging and genomic data

TL;DR

This study improves early Alzheimer's detection by combining brain imaging and genetic data with a new matching strategy.

Contribution

A gender-corrected random matching strategy for non-paired multi-modal data fusion in early AD screening.

Findings

Multi-modal data outperformed single-modal data in predictive performance.

Ensemble learning models showed stronger fitting capabilities on paired datasets.

16 genetic and 6 brain region volume features were identified as highly important.

Abstract

Alzheimer's disease (AD) is a common neurodegenerative disorder in the elderly population, and early screening can effectively delay the progression of the disease. Mild cognitive impairment (MCI) occurs prior to the onset of AD; however, the accuracy of existing MCI-to-AD prediction methods remains relatively low. Additionally, small sample sizes and high feature dimensions often lead to model overfitting, highlighting the need for effective early screening approaches. To address the aforementioned issues, this study integrated non-paired multi-modal features—including clinical indicators from the ADNI database, blood biomarkers, brain region volume features extracted from MRI, and genetic biomarkers from the GEO database—and proposed a gender-corrected random matching strategy. The Random Forest algorithm was adopted to evaluate this strategy, analyze feature importance, and compare the performance of 9 machine learning algorithms based on the top 40 ranked features. The predictive performance of multi-modal data was superior to that of single-modal data, and the proposed strategy achieved favorable results in early AD screening. 16 specific genetic features (e.g., IFI27, EDF1, RAP2A, KIF5C, SERPINA3, FBXW7, IFITM1, ISG15, PSMB3, APOE4, KCNB1, PSPH, HMGN2, S100A13, IFIT3, and CALM1) and 6 brain region volume features ranked high in terms of importance. When validated using paired datasets from ADNI across the 9 algorithms, ensemble learning models demonstrated significantly stronger fitting capabilities. The non-paired multi-modal fusion approach not only expands the sample size but also enhances the generalization ability and robustness of the model. This provides a theoretical basis for the application of this strategy in the field of small-sample medical research.

Figures