This is a repository consisting of the scripts working as analysis tools based on Machine Learning for neuroimaging data. It's for academic purpose only.
- Change the path in the instance of Logger class to output results;
- Check the datasets.py for filename and label name or add new ones;
- Change the FILENAME and the self.X attribute in the instance of Data class in main.py to use different datasets, or even concatenate ones;
- Change the model class to instantiate to use different model.
- A new datasets module for data and label manipulation.
- Used fancy indexing to select data;
- Output indices as list for reuse.
- Added regression models;
- Plotted correlations between regression predictions and original y;
- Classification can now output feature weight vectors.
- Defined a Data class, its attributes include data name, feature matrix, label vector and list of features, its methods consists of preprocessing and split;
- Used ROC curve as one of the performance metrics;
- Defined model classes, their attributes include model instance, model name and grid search parameter settings.
- Feature selection methods, e.g. f-score, are not completely done;
- ROC curves of different models are not plotted on a single figure.
- Separated data_acquisition and model_optimization from the original file;
- Used the label vectors from info.csv, avoiding creating too many datasets;
- Used plt.savefig to save figures generated.
- Still no new preprocessing method used;
- A file-to-file IO model remains to be figured out.
- Completed the integrated_rgs_model, with neg_mean_absolute_error as performance metric. Relevance vector regression (RVR) remains to be configured;
- Run on different datasets with 345 subjects;
- Run on concatenated features to obtain higher accuracy.
- Choose the right column to predict when using the regression models;
- It's not yet fully automated, thus it'd be better to modify the datasets manually;
- The accuracy is far from satisfactory, consider using different preprocessing methods.
Refactoring - add clf_models.py and rgs_models.py to reduce the amount of code in integrated.py; each *_models.py contains the model initialization and parameter settings for grid search.
In the last version there was a mistake in line 205, causing index-out-of-range error,
list_selected_features.append(list_features[2+i])the '2+i' was non-sense and is now 'i'.
Updates also include adding LogisticRegression, LinearDiscriminantAnalysis, and KNeighborsClassifier models along with their grid search parameter settings. Moreover, f-score feature selection is also included.
Committing the first version of scripts, including a integrated classification model and a integrated regression model (not completely done yet).
Since an external validation loop will result in different parameter settings, doing grid search followed by feature selection will produce different model after each iteration. Therefore, it would be a wiser choice to do a grid search first outside the validation loop inside of which feature selection will be performed, hence can we select the optimal model and obtain the corresponding selected features with larger weights, along with the mean accuracy. Should mention that most of the models are not yet configured.
It's not done so far while the main idea is to use seaborn to generate a bunch of figures showing correlations as well as the regression of PANSS scores.
Running the script integrated.py for the first time, an error like below occurred,
ModuleNotFoundError: No module named 'tkinter'refer to this site for solution, it seems that tkinter is not installed correctly.