Long Short-Term Memory Neural Networks for Identifying Type 1 Diabetes Patients with Functional Magnetic Resonance Imaging

The neuronal activation in the human brain is one of the most complex systems known nowadays that can be measured through functional magnetic resonance imaging (fMRI). Modeling this phenomenon could help in better understanding diseases with an impact on brain. Type 1 diabetes is a disease associated with the metabolism of energy, that has been associated to cognitive disorders. Here, we propose to classify Type 1 diabetic fMRIs during a working memory test using Long Short-Term Memory (LSTM) recurrent artificial neural networks due to its ability to model complex time series. We compared 20 different LSTM architectures on our database using mean and standard deviations of accuracy, specificity and F1 score. Our best result was obtained with a bidirectional LSTM obtaining a mean accuracy of 0.87, mean specificity of 0.89 and mean F1 score of 0.86. Our results have paved the way for doing similar models for other diseases and larger databases.

I. Introduction

Artificial Neural Networks (ANNs) are machine learning algorithms that are very efficient finding complex patterns and classification markers without making any prior assumptions. This makes ANNs an excellent approach to model the high complexity of brain activation patterns. Functional Magnetic Resonance Imaging (fMRI) is a technique that indirectly quantifies neuronal activation through a Blood Oxygen Level Dependant (BOLD) measurement with a high spatial resolution in expense of a low time resolution. Proposed ANN architectures for analyzing fMRIs have emerged, including Convolutional NNs [1] , [2] , autoencoders [3] , and other deep learning approaches [4] , [5] . Most of these approaches use precomputed metrics of functional connectivity, transforming the time series to linear correlation features, which might underrepresent the complexity of the brain circuitry and their emergent features through time. One limitation of fMRI-based data analysis with ANN is that it is difficult to obtain a large cohort of data to train and test the ANN.