Predicting cognitive decline: Deep-learning reveals subtle brain changes in pre-MCI stage

Ling Yue, Yongsheng Pan, Wei Li, J. R. Mao, Bo Hong, Zhen Gu, Mingxia Liu, Dinggang Shen, Shifu XiaoThe Journal of Prevention of Alzheimer s Disease 2025 · DOI

TL;DR

This paper introduces a novel deep-learning framework that can predict the progression from normal cognition to mild cognitive impairment (MCI) years before clinical diagnosis, using only structural MRI data. By identifying subtle changes in specific brain regions, this approach offers a promising tool for early detection and intervention in Alzheimer's disease.

What Problem Does This Paper Address?

Mild cognitive impairment (MCI) is a transitional stage between normal aging and dementia, and individuals with MCI are at a significantly higher risk of progressing to Alzheimer's Disease (AD). However, not all individuals with MCI will progress, making early and accurate prediction crucial for timely intervention. Current diagnostic methods often rely on expensive or invasive biomarkers. This paper addresses the challenge of identifying individuals at risk of cognitive decline in the "pre-MCI" stage, even before they show clear signs of MCI, using readily available structural MRI data. The goal is to develop a non-invasive, cost-effective method for early risk stratification.

Methods

The researchers developed a two-step deep-learning framework:

  1. Single-Region-of-Interest Network (SRNet): This network was used to identify individual brain regions that are most discriminative in distinguishing between individuals who will progress to MCI/AD and those who will remain stable. It analyzes patches of MRI data centered on specific regions.
  2. Multi-Region-of-Interest Network (MRNet): This network takes the top 10 most informative regions identified by SRNet and uses them to build a predictive model for MCI conversion. It processes information from these regions in parallel.

The output of the MRNet is a "Progressive Index" (PI), a continuous score between 0 and 1 that quantifies an individual's propensity to progress to MCI. This PI is designed to capture the subtle, gradual nature of cognitive decline, moving beyond binary classifications. The model was trained using cross-entropy loss, aiming to make the predicted PI align with the ground-truth diagnosis labels.

Datasets

  • ADNI-1: Used for training the deep-learning model (n=845).
  • ADNI-2: Used for validation (n=321).
  • China Longitudinal Aging Study (CLAS): A private, community-based cohort study in Shanghai, also used for validation (n=109).

The ADNI datasets provided longitudinal data with 36-month follow-ups, while the CLAS dataset offered a longer follow-up period of approximately 7 years.

Key Insights and Ideas

  • Subtle Brain Changes Precede MCI: The study demonstrates that even in individuals with normal cognition, subtle structural changes in specific brain regions can predict future progression to MCI.
  • Importance of Specific Brain Regions: The deep-learning model identified a core set of 10 brain regions crucial for predicting MCI progression. These include the hippocampus, amygdala, temporal lobe, insula, and parts of the cerebellum. This highlights the interconnectedness of these regions in cognitive health.
  • Progressive Index (PI) as a Novel Metric: The development of the PI score is a key innovation. It provides a continuous, intuitive measure of an individual's risk, moving beyond simple classification and offering a more nuanced understanding of cognitive decline.
  • Cross-Validation Across Datasets: Validating the model on independent datasets (ADNI-2 and CLAS) strengthens the findings and suggests the generalizability of the approach.

Key Results

  • The deep-learning model successfully identified the top 10 discriminative brain regions for predicting MCI progression.
  • In the ADNI-2 dataset, the PI score showed good predictive capability for MCI conversion, with an AUC of 0.7048.
  • In the CLAS dataset, the PI score also demonstrated strong predictive performance, with an AUC of 0.711.
  • When the PI score was combined with clinical measures (age, gender, education, and associative learning scores) in the CLAS dataset, the predictive accuracy for MCI progression increased significantly, achieving an AUC of 0.812 with a 7-year follow-up.

Limitations and Future Work

  • Limited Cohorts: The study primarily used two cohorts and suggests that collaboration with more research institutions could further validate the findings.
  • Absence of AD Biomarkers: The study did not incorporate classical AD biomarkers like APOE E4, amyloid PET, or tau protein. While acknowledging the practical challenges in community settings, the authors suggest that future work could integrate these for enhanced prediction.
  • Annual Follow-ups: The study lacked annual follow-ups, which could have provided more granular insights into the transition from normal cognition to MCI.
  • Single Modality: The current model relies solely on T1-weighted structural MRI. Future work could explore incorporating other neuroimaging modalities (e.g., T2-weighted MRI, DTI, resting-state fMRI) to potentially improve predictive performance.
  • Further Exploration of Cerebellum: The specific role of the identified cerebellar subregions in AD pathology warrants further investigation.

Summary updated

Abstract

This study effectively predicted the progression to MCI among order individuals at normal cognition state by deep learning algorithm with MRI scans. Exploring the key brain alterations during the very early stages, specifically the transition from NC to MCI, based on deep learning methods holds significant potential for further research and contributes to a deeper understanding of disease mechanisms.
BibTeX
@article{yue2025predicting,
  author = {Ling Yue and Yongsheng Pan and Wei Li and J. R. Mao and Bo Hong and Zhen Gu and Mingxia Liu and Dinggang Shen and Shifu Xiao},
  title = {{Predicting cognitive decline: Deep-learning reveals subtle brain changes in pre-MCI stage}},
  journal = {The Journal of Prevention of Alzheimer s Disease},
  year = {2025},
  doi = {10.1016/j.tjpad.2025.100079},
}
Figure