Predicting Alzheimer's disease progression using deep recurrent neural networks.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Data Science

Background:

• Early identification of Alzheimer's disease (AD) dementia is crucial for developing effective disease-modifying therapies.
• Predicting AD progression accurately requires handling complex longitudinal data with missing values.
• Existing models often struggle with missing data, a common issue in Alzheimer's Disease Neuroimaging Initiative (ADNI) studies.

Purpose of the Study:

• To develop and apply a minimal recurrent neural network (minimalRNN) model for predicting clinical diagnosis, cognition, and ventricular volume in individuals with Alzheimer's disease.
• To evaluate the minimalRNN model's performance against baseline algorithms using multimodal AD markers and longitudinal data.
• To investigate the efficacy of different missing data imputation strategies within the minimalRNN framework.

Main Methods:

• Utilized longitudinal data from 1677 participants in The Alzheimer's Disease Prediction Of Longitudinal Evolution (TADPOLE) challenge.
• Proposed and applied a minimal recurrent neural network (minimalRNN) model, comparing it with support vector machine/regression, linear state space (LSS), and long short-term memory (LSTM) models.
• Explored three missing data handling strategies: forward filling, linear filling, and 'model filling' using the minimalRNN itself during training and testing.

Main Results:

• The minimalRNN model with 'model filling' demonstrated superior performance compared to baseline algorithms in predicting AD progression up to 6 years.
• The model achieved high rankings in the TADPOLE challenge, securing 5th place in 2019 and 2nd place as of June 2020.
• The trained minimalRNN model showed robust performance even when using only one or four input timepoints, suggesting potential utility with cross-sectional data.

Conclusions:

• The minimalRNN model with 'model filling' offers a powerful and accurate approach for predicting Alzheimer's disease progression.
• This method effectively addresses the challenge of missing data in longitudinal Alzheimer's studies.
• The model's ability to perform well with limited data suggests its adaptability for various clinical data scenarios.