Detection of developmental deficits in epileptic children using multimodal tensor decomposition techniques

Abstract

Early childhood epilepsy can affect the child’s development and lead to developmental deficits. Early detection and intervention are key to enabling the child to develop normally. Resting state electroencephalogram (EEG) and Magnetic resonance imaging (MRI) are the main tools clinicians use to diagnose children with epilepsy. This motivates us to take advantage of these available data and jointly analyse them to explore the features related to developmental deficits and predict the developmental scores of newly-onset patients. In particular, our work considers EEG information, sMRI volumetric data, and psychometric evaluation scores. We use matrix-tensor decompositions to analyse the shared features between each modality all at once. This allows us to investigate the occurrence of shared profiles in EEG and sMRI related to developmental impairment. Hence, this thesis develops data fusion methods based on well-established tensor decomposition methods (canonical polyadic decomposition, CPD; block term decomposition, BTD; and Tucker decomposition, TD). The methods are validated in a publicly available dataset with healthy children (Child Mind Institute: CMI) and, more importantly, in a local dataset of preschool children with epilepsy (NEUROPROFILE: Neu).

First, the thesis focuses on a CPD data fusion model, which decomposes the multi-way data into a sum of rank-one factor matrices with the subject factor shared across three modalities. The model is optimised via grid search. The CPD model reveals distinct features associated with developmental deficits that agree with prior clinical knowledge. Then, we expand the model through direct projection to predict the developmental scores from EEG and sMRI data. A support vector machine (SVM) is used as a benchmark to compare the predicted score performance. The result reveals CPD model is better at estimating the developmental scores than the SVM. The CPD shows the feasibility of score prediction but still lacks the ability to correctly identify the deficits, which highlights the need for a more flexible data fusion model.

Next, the thesis adopts block term decomposition (BTD) to bring in additional flexibility in the modelling of the EEG tensor data. In BTD (Lᵣ,Lᵣ, 1), one mode of interest is fixed to rank one while the others vary together to rank L. Subjects with missing scores and more sMRI regions and sub-scores are included in this analysis. Bayesian optimisation is applied to reduce the hyperparameter optimisation time. The results show that BTD (Lᵣ,Lᵣ, 1) can extract additional features related to the deficits that the CPD model does not pick up. Then, we built a model to predict the developmental scores. Overall, the prediction from BTD is generally better than the CPD. However, the result shows both models may not be fully compatible with EEG tensors and suggests the need for a better-fit model.

Therefore, we adopt TD as a flexible model for the EEG data. TD can decompose tensors into factor matrices with different ranks interacting through a core tensor. However, TD without constraints is not unique. Thus, we promote the sparseness in the TD core tensor in our joint decomposition. In addition, we use structural connectivity information in the form of diffusion tensor imaging (DTI) as a graph regularisation to the data fusion model to promote interpretability. The effects of each constraint are investigated, and the most stable result is extended to predict the scores. Since not all the patients have DTI data, the score prediction is executed for both patients with and without DTI. Implementing the DTI graph regularisation is found to result in predicted scores in a more plausible range. The sparse core TD with graph regularisation performs best with the Neu dataset. However, some deficit patients are estimated to score within the normal range, which does not fulfil the aim of identifying deficits accurately.

In addition, and given that the BTD (Lᵣ,Lᵣ, 1) tensor decomposition is closely related to CPD, we investigate and expand the existing principle of CPD core consistency diagnosis (CORCONDIA) to BTD (Lᵣ,Lᵣ, 1). BTDCORCONDIA is built to assist in determining the number of components and the data compatibility to the model. The model is tested with simulated and real EEG tensor data. We show that data generated with a unique core compatible with BTD (Lᵣ,Lᵣ, 1) results in BTDCORCONDIA values of ∼ 100%. In contrast, incompatible data will lead to low values. The result confirms that it is possible to perform a core consistency diagnosis to check the compatibility between the model and data in BTD.

In summary, multimodal data fusion of paediatric brain data through matrix-tensor decomposition offers a new approach to studying the shared underlying profiles and developmental status of children with neurological diseases such as epilepsy. This could be a stepping stone for future research seeking to integrate and adopt data fusion approaches as additional tools for clinicians to prioritise children for an exhaustive assessment of their development.