Description

Advanced Machine Learning
Task 1: Predict a person’s age from brain image data

MRI image
MRI
● Magnetic Resonance Imaging (MRI) is a key technology in medical imaging
● Non-invasive + Non-radiative investigation of sensitive organs (brain)
https://data.oecd.org/healtheqt/magnetic-resonance-imaging-mri-units.htm http://www.panafrican-med-journal.com/content/article/30/240/full/
MRI processing
Raw brain scans are difficult to handle
• 3D brain scans are ~ 200x200x200 ~ 107 features/voxels (3D pixels ~ 1mm3)
• 3D structure + individual brain shapes → difficult to recognize disease patterns
MRI feature extraction
We are using ~200 anatomical features for this project
● Informative features derived from image data (with Freesurfer)
● No need to process big images (6 GB) → csv sheet (3 MB)
● No need for image analysis (feature extraction)
● Information loss
Task 1: Age prediction
● We have modified the derived input data in three ways:
1. Irrelevant features
2. Outliers
3. Perturbations (e.g. missing values, etc.)
Description about the dataset
File description
We provide the following 3 files:
• train.csv: the training set, including the features and labels
• test.csv: the test set (make predictions based on this file)
• sample.csv: a sample submission file in the correct format
Task description
Subtask 0: Filling missing values

Background
There are some missing values in the data, originally they are set to NaN values. Most of the methods can not handle then automatically. There are different strategies how to impute them: mean, median, most frequent etc.
Task requirement
We require that students fill missing values in the training and the test set.
Subtask 1: Outlier Detection
Background
Task requirement
We require that students build an outlier detection model to make classification for samples in the training set i.e. whether they are outliers.
Subtask 2: Feature selection
Background
To make the task a bit more challenging, we add some manual features to the FreeSurfer processed dataset.
Feature selection approach is thus used for the following reasons:
• Simplification of models to make them easier to interpret
• Shorter training times and avoiding the curse of dimensionality
• Enhanced generalization by reducing overfitting
Task requirement
We require that students use feature selection methods to label the features as selected features and unselected features.
Here, unselected features includes irrelevant features and redundant features.
Main task: Age Prediction
Background
After primary preprocessing and dimensionality reduction, now we finally arrive at the regression task. Tuning hype-parameter and parameter to achieve better regression performance.
Task requirement
We require that students use suitable regression methods to predict the age of a person from his/her brain data.
Evaluation metric
Coefficient of Determination R2 is the proportion of the variance in the dependent variable that is predictable from the independent variable(s). The best score is 1 and it can be negative. Prediction of always the same constant, equals to the expected value of y, will give score of 0.
How to compute it in Python: from sklearn.metrics import r2_score score = r2_score(y_true, y_pred)

Q&A