Journal paper available at Frontiers Neurology
-
Update (2/3/18): Modified by adding new comparisons.
-
Update (6/1/18): Updated to work with the newer version of TensorFlow.
-
Update (8/3/18): Jupyter Notebook Converted to py file with cells.
-
Update (6/5/19): Fixed backward compatibility issues.
This paper contains code for comparing different structures of neural networks with a different number of hidden layers.
This project requires scikit-learn, TensorFlow, and Keras, data are analyzed using NumPy, pandas, and visualize with Matplotlib. Most of these packages are included in the Anaconda Environment.
To Install Anaconda, please visit the official website for instructions.
To install [Keras]
conda install -c keras
The original code was written using Jupyter Notebooks. The current form is employed
The data contain information about patients' productions that may be employed to identify the patients. Therefore, they cannot be released in this project, based on the Ethics Agreement. Nevertheless, you may use the code using your own data.
Ethic approvals for the study were obtained by the local ethical committee review board (reference number: L091-99, 1999; T479-11, 2011); while the currently described study was approved by the local ethical committee decision 206-16, 2016. For more information see Riksbankens Jubileumsfond – The Swedish Foundation for Humanities and Social Sciences, through the grant agreement no: NHS 14-1761:1.
This research has been funded by Riksbankens Jubileumsfond – The Swedish Foundation for Humanities and Social Sciences, through the grant agreement no: NHS 14-1761:1.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
This study has showed that a Deep Neural Network architecture can identify MCI speakers and can enable the development of valid tools for identifying cognitive changes early and enable multidomain life style interventions and/or pharmacological treatments at the MCI stage, which can potentially delay or even prevent the development of AD and other types of dementia.
Future research is required (i) to evaluate multivariable acoustic predictors, e.g., predictors from consonants and non-acoustic predictors, i.e., linguistic features, such as parts of speech, syntactic and semantic predictors, sociolinguistic predictors like the education of the speaker; (ii) to establish whether these acoustic variables could be useful in predicting conversion from MCI to dementia; and (iii) to create an automated differential diagnostic tools, which will enable the classification of unknown MCI individuals from conditions with similar symptoms (cf., 57). A system of this form, will require more data from a larger population, yet our current findings do provide a promising step toward this purpose.