For a more detailed breakdown of the competition rationale, please visit the original repository's README file or the brain_anatomy_schizophrenia_starting_kit.ipynb notebook
This is our solution submitted to a competition hosted by RAMP-Studio. The relevant files are:
-
brain_schizophrenia_f1.ipynb:
- contains the main code for benchmarking a plethora of models with hyperparameters tuned via Bayesian optimization.
- We also propose a more suitable scoring method using F1 for binary classification in addition to balanced accuracy, to replace ROC-AUC, which is flawed for not taking class imbalance into account
- We also run experiments comparing standard K-Fold cross-validation, group-stratified K-Fold cross-validation
-
log reg is all you need.pdf:
- a brief presentation sharing my approach which achieved 3rd place before the RAMP-Studio competition deadline achieving a cross-validation score of 0.77 balanced accuracy.
-
Schizophrenia classification from brain anatomical data using machine learning.pdf:
- a report written in the style of a paper documenting our experiments
-
estimator.py:
- contains several models we've submitted, with the best model (after the competition deadline) being an precomputed RBF-kernel SVM achieving a .82 cross-validation balanced accuracy
-
PS: for you to rerun the brain_Schizophrenia_f1.ipynb notebook, you may have to use the skopt_edited library I uploaded. There is some deprecation warning issues when trying to use
BayesSearchCV, please see this StackOverflow question.