A simple and well-designed structure is essential for Machine Learning / Deep Learning projects. Template is from: https://github.com/MrGemy95/Tensorflow-Project-Template.
link: https://www.kaggle.com/c/trends-assessment-prediction
Deadline:
Merger Deadline: June 22, 2020
Entry Deadline: June 22, 2020
End Date (Final Submission Deadline): June 29, 2020 11:59 PM UTC
https://www.kaggle.com/aerdem4/rapids-svm-on-trends-neuroimaging
score: 0.159
interesting cuML python API runs ML on GPU
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0066572
FE: feature extraction
EDA: exploratory data analysis
Neuroimaging
Goals:
Prediction of target variables, generalization to different sites/scanners
Data Interpretation:
Target data: [age, domain1_var1, domain1_var2, domain2_var1, domain2_var2], continuous variable
How features are obtained: scICA, GIG-ICA
sMRI: high spatial resolution, providing structural/anatomical info in the brain. fMRI: low spatial resolution, record brain activity (Blood Oxygen Level Dependent signals) along time.
features: SBM loadings — subject-level weights, gray matter concentration maps from sMRI
loading.csv.
gray matter volume may be affected by person’s age or hobbies (e.g., play tetris)
197 regions in the brain?
FNC matrices — subject-level cross correlation values among 53 component timecourses (time series signals)
in fnc.csv file: 53 * 52 / 2 = 1378 features
component SM — subject-level 3D images of 53 spatial networks
SM_features — 53 3D spatial maps for train samples (possibly the key features)fMRI_mask — Variations among participants, movement artifacts
ICN_numbers — names for different networks, (each network corresponds to a index in the connectivity matrix)
Possible Bias in Data:
Different sites: 1. Quality of measured data due to different devices
2. Different sample population
3. Different conditions when taking the measurements, e.g., time, weather, etc. Missing values, especially for domain 1 — require data imputation techniques to be applied here
Pre-analysis of the data
Correlation among target variables — domain 1 variable have high correlations
Stratification technique? — to represent the original dataset, to deal with the missing values?
Statistical Analysis:
FNC correlations:
across the participants, the correlation between FNC matrices and target values
study of univariate/multi-variate effects
elbow points for explained variance and PCA technique
Structural (grey matter volume) correlations
data_load