Training recurrent neural networks robust to incomplete data: Application to Alzheimer’s disease progression modeling

• Published in: Medical image analysis 2019-04, Vol.53, p.39-46
• Main Authors: Mehdipour Ghazi, Mostafa, Nielsen, Mads, Pai, Akshay, Cardoso, M. Jorge, Modat, Marc, Ourselin, Sébastien, Sørensen, Lauge
• Format: Article
• Language: English
• Subjects: Aged; Algorithms; Alzheimer Disease - diagnostic imaging; Alzheimer Disease - pathology; Alzheimer's disease; Biomarkers; Biomarkers - analysis; Brain; Cases (containers); Cortex (entorhinal); Cortex (temporal); Discriminant analysis; Disease Progression; Disease progression modeling; Female; Humans; Inspection; Learning algorithms; Linear discriminant analysis; Long short-term memory; Machine learning; Magnetic Resonance Imaging; Male; Mathematical models; Missing data; Modelling; Neural networks; Neural Networks, Computer; Neurodegenerative diseases; Neuroimaging; Neurological diseases; NMR; Nuclear magnetic resonance; Rank tests; Recurrent neural networks; Regression analysis; Temporal gyrus; Training; Trajectories
• ISSN: 1361-8415; 1361-8423
• DOI: 10.1016/j.media.2019.01.004

•We propose a generalized algorithm to train LSTM networks robust to incomplete data.•We introduce an end-to-end approach for biomarker modeling and clinical status prediction.•It is applied to model Alzheimer’s disease progression using volumetric MRI biomarkers.•Our proposed algorithm predicts biomarker measurements with the lowest MAE.•This is the first time RNNs are applied for neurodegenerative disease progression modeling. Disease progression modeling (DPM) using longitudinal data is a challenging machine learning task. Existing DPM algorithms neglect temporal dependencies among measurements, make parametric assumptions about biomarker trajectories, do not model multiple biomarkers jointly, and need an alignment of subjects’ trajectories. In this paper, recurrent neural networks (RNNs) are utilized to address these issues. However, in many cases, longitudinal cohorts contain incomplete data, which hinders the application of standard RNNs and requires a pre-processing step such as imputation of the missing values. Instead, we propose a generalized training rule for the most widely used RNN architecture, long short-term memory (LSTM) networks, that can handle both missing predictor and target values. The proposed LSTM algorithm is applied to model the progression of Alzheimer’s disease (AD) using six volumetric magnetic resonance imaging (MRI) biomarkers, i.e., volumes of ventricles, hippocampus, whole brain, fusiform, middle temporal gyrus, and entorhinal cortex, and it is compared to standard LSTM networks with data imputation and a parametric, regression-based DPM method. The results show that the proposed algorithm achieves a significantly lower mean absolute error (MAE) than the alternatives with p