/VSMood

Primary LanguagePythonApache License 2.0Apache-2.0

VSmood

Visual scanning data and analysis code for patients with bipolar and unipolar study.

The technical description of the algorithm was presented in the paper "Learning Differences between Visual Scanning Patterns can Disambiguate Bipolar and Unipolar Patients", AAAI 2018.

Getting started

The analysis code requires the following dependencies (the installation instructions are linked below).

Note: using python from Anaconda can obtain many of the above dependencies.

Usage

Run to following code to obtain the analysis results.

python main_JAD.py

The proceduces are as follows:

  1. Reading, parsing, and processing the data. The following line will be printed on the screen

    Reading, parsing, and processing visual scanning data

  2. Training the neural network for the leave one out set of subjects (note that this takes about 30 minutes on a laptop computer). Training is complete when the second message is printed.

    1) Training leave one out classification model...
Training leave one out classification model complete

  3. The AUC and classification of the leave one out set are then provided and the results should be (tp= true positive, fp=false positive, fn=false negative, tp=true negative):

    Results 1
AUC score of the leave one out set=0.874762808349
Classification of leave one out set
tp=15, fp=2, fn=7, tn=24

  4. This is followed by training and evaluting the held-out set. The messages are similar to steps 2 and 3. (Note that training for the held-out set should take a couple of minutes).

  5. The results of both the leave one out and held out sets, remitted bipolar, remitted unipolar and healthy controls are provided.

    5.1) The AUC of classifying bipolar and unipolar (both in a depressed state) from the both the leave one out and held out set are provided as follows:

    AUC score of group 1 + 2 = 0.867430441899

    5.2) The AUC of classifying bipolar (depressed + remitted) vs unipolar (depressed + remitted) and controls from the held out set are provided as follows:

    AUC for Bipolar (depressed + remitted) vs unipolar (depressed + remitted) and controls = 0.839037927845
Classification of Bipolar (depressed + remitted) vs unipolar (depressed + remitted) and controls
tp=33, fp=13, fn=14, tn=80

    5.3) The similarity index (values in Figure 2) and the corresponding statistic tests are then provided:

    Similarity index
    Depressed bipolar disorder -0.468464264873+1.27849608688
    Depressed unipolar disorder 1.04657521002+0.881033207938
    Remitted bipolar disorder -0.119614326249+1.14093650413
    Remitted unipolar 1.34760658937+0.899898773382
    Healthy control 1.16913592799+1.07028219018
Statistic tests (t-tests)
	Bipolar depressed vs unipolar depressed t=-3.34788360318, p=0.00278888654443
	Unipolar depressed vs remitted t=-0.989352985765, p=0.329281182329
	Unipolar depressed vs healthy controls t=-0.375520481766, p=0.709257514171
	Unipolar depressed vs Bipolar remitted t=3.26934124438, p=0.0024714469487
	Bipolar depressed vs unipolar remitted t=-4.28287147359, p=0.000195991579003
	Bipolar depressed vs healthy controls t=-3.64659619235, p=0.00090619356333
	Bipolar depressed vs bipolar remitted t=-0.70748451032, p=0.485327890579
	Healthy controls vs bipolar remitted t=3.84708163655, p=0.000382196980871
	Healthy controls vs unipolar remitted t=-0.596550638954, p=0.553796871336

    5.4) Finally the plots are generated (please view the img folder for the desired image):

    ROCs saved in img/rocs.png..
Similarity index box plots saved in img/mle_bar.png...

    mle_bar.png

Note:

  • In keras.json, the "image_data_format" should be set to "channels_last" and "backend" should be set to "tensorflow".

  • For the technical description use main_AAAI.py. This implementation provides an exploratory interface to test multiple parameters:

    • Data type: RNN: "fix", "glance", LRCN: "fix-sequence"
    • Multiple instance: how to solve the multiple instance problem options: mean, 2d-mean, max-likelihood, similar, log-prob
    • Model investigations:
      • Number of CNN layers
      • Region models (semantic5, grid9, semantic8, grid16)
    • Additional features:
      • scan_path, glance duration (for glance the data type)
      • image type and image position

    To run the 3 conditions presented in the AAAI paper:

    	python main_AAAI.py fix 128 mean --region_model_type semantic8 (or semantic5)
 	python main_AAAI.py fix 128 mean --region_model_type grid16 (or grid9)
 	python main_AAAI.py fix-sequence 256 mean --print_sub