Sleep scoring model

Prerequisite

A CUDA-enabled GPU machine.

Setup

1. Run setup.sh to install neccessary python packages.
2. Run cudamat_example.ipynb to verify cudamat installation, cudamat provides an interface to perform matrix calculations on CUDA-enabled GPUs.

Preprocess data

Add feature file to sample_data/ dir. The feature file should be of the following format.

Bands

NxM data array, where N is the number of epochs, and columns refer to Delta PFC, Theta HPC etc.

EpochsLinked

Nx4 data array, where N is the number of epochs, and columns are described as follows:

• column 1: epoch ID
• column 2: epoch index (currently not used)
• column 3: ground truth sleep stage ID, where
  • 1 is associated with wakefulness,
  • 2 is associated with NREM sleep,
  • 3 is associated with REM sleep
• column 4: the subject ID (used in multi-subject analysis only)

EpochTime

Nx3 data array, where N is the number of epochs, and columns are described as follows:

• column 1: epoch ID
• column 2: recording mode (i.e. baseline or recovery), where
  • 1 is associated with baseline,
  • 2 is associated with recovery (after sleep deprivation)
• column 3: the epoch date-time

If your feature file is not in the above format, use mcRBM_input_features.ipynb in sample_data/ dir to generate the feature file (.npz) needed as input for the model.

Configuration

Once you have the feature file (.npz) ready, update the experiment details in configuration_files/exp_details file.

1. Set dsetDir to the absolute path of sample_data dir.
2. Set dSetName to the name of the feature file.
3. Set expsDir to the absolute path of output or analysis dir where model weights, inference analysis and plots will be stored.
4. Set expID to the unique name of your experiment

There are some parameters and flags that you can set that are useful for training the model and they are described in the configuration_files/exp_details file.

You can also tune parameters described in the configuration_files/input_configuration file that are also useful for training the model.

Training and inference

Now, the model can be trained following which we can do inference analysis to get latent states and classification of the states.

Step 1: Train the model

Run train_model.ipynb

Step 2: Latent state inference

Run infer_states.ipynb

Step 3: Latent state analysis

Run latent_states_analysis.ipynb