P4‐210: Neural‐Net Model for Predicting Clinical Symptom Scores in Alzheimer's Disease

Development of diagnostic and prognostic tools for Alzheimer's disease (AD) is complicated by the substantial clinical heterogeneity in prodromal stages [Lam 2013, Querbes 2009]. Many neuroimaging studies have focussed on case-control classification and AD-conversion prediction. However, prediction of clinical scales, such as ADAS-cog, has received less attention. Such scales may be more capable of capturing not only disease severity, but also subdomains of clinical presentation. Therefore prediction of these scores may be integral in providing nuanced prognosis [Coupé 2015, Stonnington 2010, Zhang 2012]. Machine-learning approaches have shown great promise in using complex, high-dimensional data, such as is found in neuroimaging, towards prediction tasks [Plis 2014, Suk 2013]. In this work, we present a novel neural-network (NN) model that combines input from two structural imaging modalities: hippocampal (HC) segmentations and cortical thicknesses (CT), which are highly relevant to neuroanatomical degeneration patterns observed in AD and predicts the ADAS-13 score for a given individual. The ADNI-1 baseline cohort comprising 819 subjects was processed through MAGeT Brain [Pipitone 2014], and CIVET [Lerch 2005] pipelines to extract HC segmentations and vertex-wise CT values. Using a custom data-driven cortical parcellation, CT vertices were grouped into 688 regions of interest with roughly equal numbers of vertices (Fig.1). HC segmentations and CT measures were combined using a hierarchical NN model (Fig. 2) for ADAS-13 score prediction. The NN model performance was compared against linear-regression (LR), support-vector-machine (SVM), and random-forest (RF) models under 10-fold cross-validation paradigm. Based on Pearson’s r correlation between actual and predicted ADAS-13 scores, performance statistics (mean, median, std. error, see Fig. 3) indicate that NN [.58,.61,.06] outperforms the other three models: LR [.53,.51,.07], SVM: [.54,.55,.08], and RF:[.52,.54,.10]. Significant improvement is seen with HC modality with NN model compared to baseline models. All four model classes show boost in performance when HC and CT modalities are combined. The correlation scores validate the utility of NN as prediction model. The proposed modular architecture of the NN provides a framework for multimodal analysis, and can be extended to incorporate other data modalities and/or predicting other clinical scales. A custom cortical surface panellation (atlas) comprising 688 regions of interest (ROI), with each comprising roughly equal number of vertices. The parcellations were obtained using a triangular surface mesh obtained from a CIVET model. The vertices of the mesh were grouped together based on spatial proximity using spectral clustering method*. Bilateral symmetry within the vertices of the hemispheres was preserved. The atlas was propagated to each subject to obtain thickness samples per ROI. ∗http://scikit-learn.org/stable/modules/generated/sklearn.cluster.SpectralClustering.html A) Structure of a generic artificial neural network (NN) model. A neural net may comprise multiple hidden layers that encode hierarchical set of features from input, informative of the prediction/classification task at hand. B) The proposed partitioned, multi-layer NN model comprises separate computing branches for left and right HC segmentations, as well as, for CT modality. Layer A comprises input nodes with total dimensionality equaling to 22634. Layer B comprises hidden nodes that extract individual features from each modality, which are subsequently combined into layer C. Finally, layer D comprises a single output node which provides the prediction value of the ADAS-13 score for a given subject. The partitioned architecture reduces the number of model parameters which helps mitigate overfitting issues, as well as, allows structure specific feature analysis. Top) Performance of four model classes subject to individual and combined input modalities. The numbers and the black lines indicate the mean and standard error, respectively, of correlation values of each model over 10-folds. The baseline models (LR, SVM, and RF) use left and right HC volumes + mean CT values based on Automated Anatomical Labeling (AAL) atlas (www.cyceron.fr/index.php/en/plateforme-en/freeware) as input. Whereas the artificial NN models use HC segmentation masks and CT values based on custom atlas (see Fig. 2) with total dimensionality > 22500 for the combined input case. All models were trained with a nested-inner loop that searched for optimal hyperparameters. *implies that input data was normalized to center to the mean and component wise scale to unit variance. Bottom) Correlation plots for 4 models with HC+CT input for all subjects from concatenated list of heldout subsets from each fold.