In a recent study published in eClinical Medicine, researchers analyzed disease progression to identify distinct patterns in the Alzheimer's disease (AD) trajectory using Mendelian randomization (MR) and deep learning (DL).
Study: Identifying underlying patterns in Alzheimer's disease trajectory: a deep learning approach and Mendelian randomization analysis. Image Credit: Andrey Suslov/Shutterstock.com
Alzheimer's disease is a neurodegenerative condition with varied clinical presentation at the inter-individual and intra-individual levels.
The heterogeneous nature of the condition warrants the development of effective tools to facilitate early diagnosis, determine the risk of disease progression, and initiate prompt treatment to improve the overall standard of care for AD patients.
About the study
In the present study, researchers developed and validated a novel DL-based model to assess the risk of progression through different stages of cognitive decline.
The researchers developed a DL-based model to assess patterns of progression from the cognitively normal (CN) stage to the mild cognitive impairment (MCI) stage and further to Alzheimer's disease (AD) development by estimating time-to-conversion and survival clustering of distinct subgroups defined by comprehensive variables and varying progression rates.
The model was trained using clinical and T1 relaxation-weighted MRI data from 1,370 individuals in the AD Neuroimaging Initiative (ADNI) group and externally validated using data from the Australian Imaging Biomarkers and Lifestyle Study of Aging (AIBL) group (233 individuals).
The team evaluated the model's ability to identify clinically and physiologically significant trends in AD trajectories. Furthermore, time-to-conversion prediction was performed to assess the prognostic value of the found patterns.
A Mendelian randomization study was also conducted to evaluate the causal associations between the identified patterns and Alzheimer's disease development.
The researchers assessed the transitions between consecutive stages of Alzheimer's disease progression: patients identified as cognitively normal (CN) at study initiation who developed mild cognitive impairment (CN to MCI) and those with MCI at study commencement who developed AD (MCI to AD).
There were 587 individuals in the CN to MCI dataset [315 (54%) were female, with a mean age of 75 years]. 119 (20%) who developed mild cognitive impairment. The MCI to AD dataset included 783 individuals [306 (39%)] females, with a mean age of 75 years, of which 44% developed AD dementia.
The study only included those who had two or more T1-structured MRIs. The team chose MRI and clinical data at the start of the trial and at the moment of transition for individuals who converted from cognitively normal to the MCI stage or from the MCI stage to AD-related dementia. The researchers chose data from the trial's beginning and the study's end for censored individuals.
Only clinical characteristics with less than 30% missing data in every dataset were included to guarantee that useful clinical features were chosen. The chosen ADNI participants had 1.5 and 3.0 T magnetic resonance imaging data retrieved.
The team recovered mean cortical thickness, gray matter volume, and cortical surface area values as neuroimaging parameters from the right and left hemispheres of the brain. The hazard ratios (HRs) were calculated using Cox proportional hazard modeling. The concordance index (C-index) values were also determined.
The model identified patterns distinguished by considerably diverse biomarkers and varying advancement rates. HR values (CN to the mild cognitive impairment stage, hazard ratio, 3.5; MCI to Alzheimer's disease, hazard ratio, 8.1), concordance index (cognitively normal to MCI, 0.6; MCI to AD, 0.7), and area under the curve (cognitively normal to MCI, three years 0.8, five years 0.9; MCIAD, three years 0.9, five years 0.96) showed a significant prediction capacity.
The model performed well in conversion time estimation in the external validation cohort (CN to the mild cognitive impairment stage, concordance index, 0.7; MCI to Alzheimer's disease, concordance index, 0.8).
The model not only detected simulated patterns with varying atrophy rates, but it also outperformed other state-of-the-art models in terms of time-to-disease conversion estimation in the AD trajectory. Notably, consistent findings were observed after controlling for variables such as field strength and manufacturer, indicating the model's dependability and robustness.
Individuals with positive amyloid beta (A+) or phosphorylated tau (T+) status had a higher chance of developing MCI or AD. The MCI to AD conversion patterns indicated discrete underlying subgroups of neurodegeneration, which differed not only in biomarker composition, cognitive scores, and imaging signals but also in genetic origins.
A causal relationship was observed between MCI to AD patterns and time-to-disease conversion in the initial three years. Three genetic variants on chromosome 18 in DESL-AS1 (rs176004, rs393881, and rs281552) were related to converting MCI to AD dementia.
In both CN to MCI and MCI to AD dementia cases, female sex, lower brain volumes (in regions such as the hippocampus, entorhinal cortex, and middle temporal gyrus), and poorer cognitive assessment scores [such as the Mini-Mental State Examination (MMSE) and AD Assessment Scale-Cognitive Subscale (ADA)] were associated with faster clinical deterioration of cognitive function [Clinical Dementia Rating Scale Sum of Boxes (CDRSB) and Functional Assessment Questionnaire (FAQ)].
Furthermore, APOE4, a significant risk factor for developing AD, was strongly expressed in the MCI to AD sample. The whole brain, entorhinal cortex, and fusiform gyrus played a critical role in distinct patterns of CN to MCI and MCI to AD progression, indicating that these regions emerge as a pivotal neural signature for distinguishing the patterns of AD progression.
Overall, the study findings highlighted a model to predict AD progression using real-world data. The model identified clinical and biological patterns, improving our understanding of AD progression. It could aid in clinical trial design and decision-making.
Survival clustering enabled the modeling of clinical biomarkers and neuroimaging features to provide a realistic representation of individual patient trajectories.