Data science in medicine: machine learning

Question 1

What is machine learning and what is it based on?

Answer 1

• The construction of approximate, generalizing (predictive) models by learning from examples, for problems for which no full physical model is known (yet).
• Machine learning algorithms build a mathematical model based on sample data, known as ‘training data’, in order to make predictions or decisions without being explicitly programmed to perform the task.

Question 2

Machine learning will find structure in data by:
- clustering
- outlier/anomaly detection
- dimensionality reduction, selecting useful features
- regression
- classification

Where are all these processes aimed at?

Answer 2

They are all aimed at generalisation → making a prediction for data you have not yet seen.

Question 3

What is clustering in regard to machine learning?

Answer 3

Identification of ‘natural groups’ within a population, e.g. among a population of apples → identifying red and green apples.

Question 4

What is outlier detection in regard to machine learning?

Answer 4

Identifying strange outliers/object within a population, e.g. among a population of apples → identifying one pear.

Question 5

What is dimensionality reduction in regard to machine learning?

Dimensionality reduction is also called feature selection

Answer 5

Finding predictive measurements to identify supgroups within a population. For example, identifying specific characteristics for apples that are red and identifying different specific characteristics for apples that are green.

Question 6

What is regression in regard to machine learning?

Answer 6

Identifying real-valued outputs, e.g. predicting prices of the different types of apples given the various characteristics of these different types of apples.

Question 7

What is classification in regard to machine learning?

Answer 7

Distinguishing different groups within a population, e.g. distinguishing apples from pears.

Question 8

Example of machine learning in gene expression diagnostics

In this example, the genetics of different patients are collected for the diagnosis/relapse in childhood leukemia.
- Clustering is used to identify similar subtypes of the disease in the group of patients.
- Clustering is also used to identify ‘clusters’ of genes with similar ‘disruptive’ processes.
- Outlier detection is used to identify technical errors and rare patient-rare genetic backgrounds.
- Dimensionality reduction is used to identify potential biomarkers that can predict the disease.
- Regression is used for predicting the survival time of the patients
- Classification is used for the prediction of e.g. metastasis

Question 9

What is important to take into consideration when wanting to apply machine learning to e.g. be able to diagnose patients with a certain disease with the use of their genetics?

Answer 9

That the algorithm/machine learning is based on a mathematical representation of all the objects.

Question 10

Thus: to implement machine learning, we have to find a mathematical representation of objects. Objects are usually represented by features (i.e. srts of useful measurements obtained from some sensors).
Imagine a dataset that contains information about the object (apple or pear), weight and colour.
- How would a dataset look like that is clustered by machine learning?
- How would a dataset look like where regression is applied by machine learning?
- How would a dataset look like where classification is applied by machine learning?

Answer 10

• Clustering → Based on the characteristics weight and colour, machine learning idenifies apples and pears within one population.
• Regression → Based on the characteristics object, weight and colour, machine learning identifies the price of the apples and pears.
• Classification → Based on the given characteristics, the apples and pears are labeled separately.

Question 11

Look at the picture: what dataset (left or right) can more easily be applied for machine learning and why?

Answer 11

The left dataset, because:
- simple
- knowledge is present
- a few good features
- almost seperable classes (classification) or a linear relation (regression)

Not the right dataset, because:
- complex
- lack of knowledge
- many poor features
- overlapping classes (classification) or highly non-linear relation (regression)

Question 12

Explain what a vector and vector space is.

Answer 12

• Vector → mathematical object characterized by size and direction (e.g. speed and power, both characterized by size and direction)
• Vector space → a set vectors added together and multiplied (i.e. scaled) by numbers called scalars.

Question 13

Questions about vectors and model development are still coming xx