Deep learning to predict future cognitive decline: a multimodal approach using brain MRI and clinical data.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Predicting clinical decline in aging individuals with cognitive impairment is crucial for personalized treatment and clinical trials.
• Key metrics like the Clinical Dementia Rating scale 'sum of boxes' (sobCDR) are vital for tracking disease progression.

Purpose of the Study:

• To compare deep learning approaches for predicting 2-year changes in sobCDR scores.
• To evaluate a hybrid convolutional neural network (CNN) integrating 3D brain MRI with clinical/demographic data against an automated machine learning (AutoML) framework.

Main Methods:

• Trained a hybrid CNN using 3D T1-weighted brain MRI and tabular data (age, sex, BMI, baseline sobCDR).
• Benchmarked CNN against AutoGluon, an AutoML multimodal framework.
• Evaluated models on 2,319 participants from ADNI, OASIS-3, and NACC cohorts.

Main Results:

• Multimodal fusion of image and tabular data shows promise for dementia prognostics.
• Deep learning on MRI data may not always add significant predictive value when clinical covariates are highly predictive.
• AutoML-based multimodal fusion offers a robust baseline when tabular data are strongly predictive.

Conclusions:

• Deep learning can fuse brain imaging and clinical data for personalized dementia prognostics.
• The utility of multimodal fusion depends on the data types and the predictive power of existing clinical variables.
• Understanding the relative value of different data modalities is key for selecting appropriate prognostic strategies.