multimodal-arousal-detector

This repository contains scripts for preprocessing .edf files to .txt files in the folder 'matlab'. The folder 'python' contains scripts for training and using the model for predicting arousals and wake.

Getting started

Copy the repository to a local directory.
Create a folder to store preprocessed .edf files.
Create a folder to store output arousal and wake predictions.

Requirements

Matlab >= 2017b
Python >= 3.6
Python module numpy >= 1.14.0
Python module tensorflow >= 1.5.0

Making new predictions

Locate path to folder containing .edf files of interest (p_edf) and select an output path (p_output)
The matlab script LoadEDF.m is currently able to handle .edf files from Wisconsin Sleep Cohort, Stanford Sleep Cohort, MrOS Sleep Study, and Cleveland Family Sleep Study. The script needs to be changed to handle differently structured .edf files if necessary.
Run "matlab" function PreprocessNewData.m as preprocess.PreprocessNewData(p_edf,p_output,ftype,Overwrite), where ftype can currently be one of {'mros','cfs','wsc'}, and if Overwrite = 1 files are overwritten if existing in p_output, otherwise put Overwrite = 0.
Run ar_predict.py with flags 'pathname', 'output_dir', and 'overwrite' set to input .txt file directory, output prediction directory, and binary decision to overwrite predictions, respectively. 'pathname' should be the same as 'p_output' when PreprocessNewData.m was run.
In matlab, run getPred(fname,T1,T2,L) to postprocess predictions (fname: file path, T1: arousal prediction threshold = 0.225, T2: wake prediction threshold = 0.45, L: desired output length (default = prediction length)).

Data access

Polysomnography data from the MrOS Sleep Study and Cleveland Family Sleep Study is available upon request from the National Sleep Research Resource (NSRR).

Training the model

In matlab/resources folder, a file called data_paths.m specifies directories for input data. These should be changed to list the target data.
The file PreprocessData.m in matlab/+preprocess and LoadEDF.m is set up to preprocess data from Wisconsin Sleep Cohort, Stanford Sleep Cohort, MrOS Sleep Study, and Cleveland Family Sleep Study.
PreprocessData.m can now be run to preprocess data and save it in the location assigned in paths.m.
In python/ardetector ar_config.py the data_dir variable should be set to the same location as in paths.m.
Run ar_train.py with hyper-parameters specified in flags or with random hyper-parameters by setting randomgen=True in the main function. Training can be continued from pre-trained weights by setting the flag "proceed" different from 0.