Learning Differences Between Visual Scanning Patterns Can Disambiguate Bipolar and Unipolar Patients
Authors: Jonathan Chung, Moshe Eizenman, Uros Rakita, Roger McIntyre, Peter Giacobbe
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we present novel methods to differentiate between BD and MDD patients. The methods use deep learning techniques to quantify differences between visual scanning patterns of BD and MDD patients. ... Using data from 47 patients with MDD and 26 patients with BD we showed that using semantic Ro Is, the RNN improved the performance of a baseline classifier from an AUC of 0.603 to an AUC of 0.878. |
| Researcher Affiliation | Academia | Jonathan Chung,1 Moshe Eizenman,1,2 Uros Rakita,3 Roger Mc Intyre,3,4,5 Peter Giacobbe3,5 1 Department of Electrical and Computer Engineering 2 Ophthalmology and Vision Sciences & Institute of Biomaterials and Biomedical Engineering 3 Department of Psychiatry, 4 Department of Pharmacology and Toxicology University of Toronto, Toronto, ON 5 Department of Psychiatry, University Health Networks, Toronto, ON |
| Pseudocode | No | The paper describes network architectures and provides mathematical equations for LSTM cells but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statements about the release of source code or links to a code repository for the described methodology. |
| Open Datasets | Yes | Fifteen slides contained images of emotional faces with happy and sad expressions that were selected from the Karolinska Directed Emotional Faces (KDEF) database (Lundqvist, Flykt, and Ohman 1998). |
| Dataset Splits | Yes | Because of the limited number of subjects, we employed a leave one out 3-fold cross validation scheme to evaluate the deep learning methods (LSTM with user defined Ro Is and LRCN). That is, at each step, all the sequences from one individual were removed and the remaining sequences were fed into a 3-fold cross validation to randomly split the data into training and validation data. The additional 3-fold cross validation was performed to reduce the chances that the parameters of the model were carefully tuned for the small dataset. The models were trained with 2/3 of the patients in each of the 3-folds until convergence (validation data 1/3 of the subjects) and P(Y |C = BD) (Equation 7) for the left out subject was calculated. |
| Hardware Specification | Yes | The Titan Xp used for this research was donated by the NVIDIA Corporation. |
| Software Dependencies | No | The networks were implemented with Keras (Chollet 2015) (that was built on top of Tensorflow (Abadi et al. 2016)). Specific version numbers for Keras and TensorFlow are not provided. |
| Experiment Setup | Yes | For the CNN within the LRCN, the weights of the convolutional layer were initialized uniformly (Glorot and Bengio 2010). After the max out and fully connected layer, dropout layers with a dropout probability of 50% were included. The weights of the LSTM were initialised orthogonally. Dropout layers with a dropout probability of 50% were applied to non-recurrent layers (Lipton et al. 2015). The LSTM was optimised with the Adam algorithm (Kingma and Ba 2014) with a learning rate of 0.001 and mini batch sizes of 40. Early stopping with a patience of 30 was applied. |