Resolving heterogeneity in the diagnosis of Alzheimer’s disease and its progression using multimodal data

Alzheimer’s disease (AD) is a complex and diverse illness that makes early detection extremely difficult. Most existing research utilizes data to identify biomarkers and more homogeneous subgroups to improve the detection, prediction of progression, and prognosis of AD. However, AD still suffers from a lack of appropriate biomarkers for early symptom detection and blurred boundaries between different subgroups. Here, an unsupervised clustering method known as similarity network fusion (SNF) was employed to analyze multimodal data from 972 subjects, including 370 with cognitively normal (CN), 565 with mild cognitive impairment (MCI), and 37 patients with AD. First, we constructed a similarity network for subjects using cognitive scores, genetics, and magnetic resonance imaging (MRI) related data, respectively. Then the SNF fusion method was employed to integrate the data, and spectral clustering was used to find subgroups sharing similarities across modalities. Our results indicated that the approach accurately diagnosed both current and prospective AD (~ 90%). Notably, we successfully identified two MCI subtypes with biological and clinical significance, validated by longitudinal studies of cognitive, clinical, fluid biomarkers and MRI-related features, dementia diagnosis, and pseudo-trajectory analysis. We also observed many dysregulated processes and signaling pathways between MCI subtypes, such as the GnRH signaling pathway, VEGF signaling pathway, and insulin signaling pathway. Overall, our research offers a distinctive viewpoint on the diversity of AD, and the more specific subtypes of MCI help create customized treatment plans.