For the deep learning portion of our analysis we employed a modified multivariate LSTM-FCN (fully convolutional network). Such network architectures have been shown to perform well at temporal feature extraction and analysis of multivariate time series [1]. Our implementation was modified slightly for the context of fMRI data. The LSTM block was stacked to improve its ability to analyze the time series data. Further, dropout was replaced with LSTM dropout cells – which include recurrent dropout in the connections between the recurrent units. This was done to improve network normalization and regularization. The dimensional shuffle was not included, as this was a trick for reducing dimensionality when timesteps outnumber features, which was not always the case. Variable length sequences were padded with zeroes to make the input tensor have uniform dimension. Features were normalized to zero mean and unit variance. The model was trained using the Keras library and Tensorflow backend.
	Trained on the full range of ABIDE data the network performed relatively well, achieving ~75% validation accuracy in the best trained model. As the fMRI pipeline is somewhat dependent on the location of the scan we experimented with training and validating only on scans from a single location. This converged, but could potentially improve with careful tuning of the learning rate as validation loss is minimized. 100% accuracy was obtained in the best case, however at higher numbers of epochs the model did show signs of overfitting. Larger numbers of samples from single locations would improve this kind of analysis.
	Hyperparameter tuning variables such as convolutional kernel size and stacked LSTM depth/width could be employed to further improve performance. The data could be augmented with additional scans from single areas in order to improve performance as well, as fMRI processing pipelines are rather variable.



[1] https://arxiv.org/pdf/1801.04503.pdf