Predicting Alzheimer's disease progression using multi-modal deep learning approach

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative condition marked by a decline in cognitive functions with no validated disease modifying treatment. It is critical for timely treatment to detect AD in its earlier stage before clinical manifestation. Mild cognitive impairment (MCI) is an intermediate stage between cognitively normal older adults and AD. To predict conversion from MCI to probable AD, we applied a deep learning approach, multimodal recurrent neural network. We developed an integrative framework that combines not only cross-sectional neuroimaging biomarkers at baseline but also longitudinal cerebrospinal fluid (CSF) and cognitive performance biomarkers obtained from the Alzheimer's Disease Neuroimaging Initiative cohort (ADNI). The proposed framework integrated longitudinal multi-domain data. Our results showed that 1) our prediction model for MCI conversion to AD yielded up to 75% accuracy (area under the curve (AUC) = 0.83) when using only single modality of data separately; and 2) our prediction model achieved the best performance with 81% accuracy (AUC = 0.86) when incorporating longitudinal multi-domain data. A multi-modal deep learning approach has potential to identify persons at risk of developing AD who might benefit most from a clinical trial or as a stratification approach within clinical trials.

Figure 1

An example using longitudinal data for MCI conversion prediction. Contrary to the experiment with baseline visit data, longitudinal data of individuals in all stages (CN, MCI, and AD) was used for training a classifier. Then, portion of longitudinal data was taken to the classifier to predict AD progression after Δt.

Figure 2

The number of subjects available in demographic data, neuroimaging data, cognitive performance, and CSF biomarkers over ∆t.

Figure 3

Predictive performances with “proposed”, “baseline”, and “single modal”. Abbreviations: COG: cognition performance biomarkers; CSF: cerebrospinal fluid biomarkers; NeuroImg: MRI biomarkers.

Figure 4

ROC curves from the “proposed”, “baseline”, and each “single modal” method. (a) 6 month prediction. (b) 12 month prediction. (c) 18 month prediction. (d) 24 month prediction.

Figure 5

Illustration of recurrent neural network. RNN is composed of input, memory state, and output, each of which has a weight parameter to be learned for a given task. The memory state (blue box) takes the input and computes the output based on the memory state from the previous step and the current input (left). Since the RNN has a feedback loop, variable-length input and output sequence can be represented as an “unfolded” sequence (right).

Figure 6

Overview of the proposed method. Our proposed method contains multiple GRU components that accept each modality of the dataset. At the first training step (blue dashed rectangle), each GRU component takes both time series or non-time series data to produce fixed-size feature vectors. And then the vectors are concatenated to form an input for the final prediction in the second training step (red dashed rectangle).