Skip to content

Latest commit

 

History

History
66 lines (54 loc) · 6.13 KB

README.md

File metadata and controls

66 lines (54 loc) · 6.13 KB
Raw

BIOMAG2022 competition "Ketamine in Depression"

This repository contains the code for our stab at the BIOMAG2022 data analysis competition "Ketamine in Depression". The dataset includes MEG resting-state DATA from 35 participants (22 with MDD, 14 healthy). Each participant had two sessions: one under ketamine and one sober (after taking placebo). The recording are unlabelled, we only know which sessions belong to the same participant. Our task is to either classify subjects as depressed/healthy or to classify sessions as ketamine/placebo.

Here is the description from the BIOMAG 2020 Data Analysis Competions page (as of Apr, 23 2021):

Organisers: Section on the Neurobiology and Treatment of Mood Disorders, National Institute of Mental Health, National Institutes of Health, Bethesda MD USA

Contact: Jessica Gilbert, PhD, [email protected]

Data: Resting state MEG data from a clinical trial of ketamine’s mechanisms of action in major depression

The data include 250 second eyes-closed resting-state data from a randomized, double-blind, placebo-controlled study of ketamine’s mechanisms of action. Data for the competition come from a total of 36 male and female participants aged 18–65 years old, 22 of which had major depression (MDD) and 14 of which were healthy control subjects. Subjects with MDD had been diagnosed with recurrent MDD without psychotic features using the Structured Clinical Interview for Axis I DSM-IV Disorders (SCID)-Patient Version. Subjects with MDD were required to have a score greater than 20 on the Montgomery-Åsberg Depression Rating Scale (MADRS) at screening. MDD participants were considered treatment-resistant and had to have not responded to at least one adequate antidepressant trial during their current episode, as assessed using the Antidepressant Treatment History Form and the current episode had to have lasted at least four weeks. Subjects were free from psychotropic medications in the two weeks before randomization (five weeks for fluoxetine, three weeks for aripiprazole). Healthy control subjects had no Axis I disorder as determined by SCID-NP, and no family history of Axis I disorders in first degree relatives. Healthy control subjects were also free of medications affecting neuronal function or cerebral blood flow or metabolism. Subjects in both groups were in good physical health as determined by medical history, physical exam, blood labs, electrocardiogram, chest x-ray, urinalysis, and toxicology.

The competition data include two resting-state scans per participant, one occurring 6-to-9 hours following ketamine administration and one occurring 6-to-9 hours following placebo saline administration. These data were collected using a CTF Omega 275-channel system. The data have been de-identified, so each participant has been given a randomized 8-letter code at the beginning of the filename. You will also find the date of the scan in the filename (i.e., YYYYMMDD). Each scan occurred approximately 14-days apart, and participants were randomized to receive either ketamine or placebo-saline during their first infusion. You can approach the data analysis competition in one of two ways. First, you can attempt to classify participants with MDD from healthy control subjects (i.e., a between-subjects factor). Second, you can attempt to classify the scan session (i.e., a within-subjects factor: ketamine versus placebo). In either case, you should submit a short, written report detailing how you approached the data analysis (i.e., describing whether you classified based on the between-subjects or within-subjects factor), the methods used, and defining which scan(s) correspond to your chosen grouping of interest. You will receive one point for each scan that is correctly classified at either the within or between-subjects level.

Deadline: Monday 2nd May 2022

Limitations: You are not allowed to use/distribute the data for any purposes irrelevant to the competition, without the organizer’s permission.

Team (in the order of the increasing first name length):

  • Egor Levchenko
  • Kirill Stepanovskikh
  • Evgenii Kalenkovich

Preprocessing

Raw files was downloaded from Globus system and save to BIOMAG2020_comp_data. By now, we have 3 problematic subjects and info about them saved in subject_problems_description.txt.

At first raw files were converted to BIDS compatible format and stored to data_bids folder using 01_switch-to-BIDS.ipynb. All further steps were done on BIDS valid data and saved to derivatives subfolder in data_bids.

Preprocessing is implemented in Snakefile The preprocessing steps:

  1. Downsample to 300 Hz and linear filtering: Notch at [60, 120] and high-pass at 0.3 Hz.
  2. ICA to remove eye and heart artifacts. Importantly, bad components are detected iteratively.

All preprocessing info is saved to html report. Some notebooks and code are in 01_preprocessing folder. Preprocessing html reports should be manually inspected to assess the quality of the preprocessing.

Clustering

Calculate eigenvalues in create_eigenvalues rule. After that we make clustering using KMeans with a Wasserstein distance and Wasserstein barycenters. All necessary info about clustering algorithm is stored in 02_clustering/clustering.ipynb

How reproduce the results?

  1. Git clone the repository
  2. Create conda environment using conda-environment.yml file
  3. Add path to original data and bids data to system environment variable
  4. Run 01_switch-to-BIDS (can skip)
  5. Run snakefile: snakemake (if you are in the same folder with Snakefile) or snakemake -some_flag -PATH/Snakefile
  6. Run clustering.ipynb

Things to improve

  1. Update switch-to_BIDS to make true path (use env system variable)
  2. Nice to have some info about the files to be sure that they are the same between developers (for example, switch-to-BIDS returns the same result)
  3. Push subject_problems_description.txt file to GitHub repository.
  4. Add references (for example, an article about Laplacian and Wasserstein distance)