Open Dementia Reports (ODR) - theory

This is my master's thesis where I delve in depth in the psycho and bio markers of Dementia, checking whether different types of Dementia can be diagnosed using Machine Learning. Then on top of these, I add some explainability using the Shapley method for Machine Learning. This, in principle, is a good starting ground for standardizing the Dementia Reporting for different patients, and can later become a valuable tool for clinicians.

• Thesis
• Thesis, Chapter 2 - Theory
• Thesis, Chapter 3 - Methodology
• Thesis, Chapter 4 - Results
• Thesis, Appendix (mainly on preprocessing)
• Notes

Note: This is previous to last version, and for some reason I dont have the last version. So, some references are missing. Keep your history folks.
Note: the following are not updated

Processing Units

Subprocess	Possible options
Imbalance
Over-sampling	No over-sampler, RandomOverSampler(), SMOTENC()
Under-sampling	No under-sampler, CustomHandler(), RandomUnderSampler(), NearMiss()
categorical
Missing data	No imputation, SimpleImputer(*)
Transforming data	No encoding, OneHotEncoding()
Feature selection	No selection, SelectPercentile(Chi2), SelectFromModel(*)
numerical
Missing data	No imputation, SimpleImputer(*), KNNImputer(), IterativeImputer(*)
Transforming data	No scaling, StandardScaler(), RobustScaler()
Feature selection	No selection, VarianceThreshold, SelectPercentile(Pearson) , SelectFromModel(*)
method
Classifier	RandomForest(), XGBoost(), LightGBM(), HistGradientBoostingClassifier()
Cross-Validation	GridSearchCV(scorer='f1_micro', cv=StratifiedGroupKFold)

Table: Transformers and Estimators used in the pipelines.

Possible Pipeline	Sample selected Pipeline for ADNI

Note: In case you are using RandomForest, need to use a specific combination of preprocessing steps. To be updated.

Results

Best predictors for each class

For each class, the importance of the features are averaged, and then shared with the docs.

Main predictors for big classes	Main predictors for smaller classes
(a) Healthy Controls	(b) Fronto-Temporal Lobe Degeneration
(c) Alzheimer's Disease	(d) Dementia with Movement Disorders

Single case for each class

The following are the importance of predictors for a single case. As you can see, its different from the predictors mentioned above, as the ones above are averaged among all the cases for a single class.

Single case	Single case
(a) Healthy Control	(b) Fronto-Temporal Lobe Degeneration
(c) Alzheimer's Disease	(d) Dementia with Movement Disorders