Machine Learning and Reasoning for Health - Slides Summary
Lecture 1: Introduction
AI disciplines
Machine learning
● Definition:
○ “Machine learning is to automatically identify patterns from data”
○ “A computer program is said to learn from experience E with respect to some
class of tasks T and performance P, if its performance at tasks in T improves with
E.” (Mitchell)
● Why did it become so popular?
○ More data
○ More computational power
○ Better algorithms
Knowledge represention/reasoning
● Representing information about the world in a form that a computer can use
● Representing knowledge about solving complex tasks
● Tasks like: diagnosing, planning, having a dialog with human...
● Several forms of knowledge representation
○ (logic, rules, semantic nets, ontologies etc.), and thus of reasoning.
● KRR: symbols & manipulation of symbols
● data => information => knowledge
● ...- - -... => SOS => emergency alert -> start rescue operation
AI for Health
● In health care there is of course a lot of data
● And we have an increase in devices that collect usable data
● Mobile health: over 100,000 mobile apps in the iTunes store alone
● What can AI contribute to health?
1
,Machine Learning and Reasoning for Health - Slides Summary
● How good is AI? → AI vs. Doctors
● Early detection/diagnosis
● Personalization of treatments
○ Provide advice to doctors on best treatments per patient
Prevention
● How can AI aid in prevention? Some examples:
○ Personalized coaching
○ Signaling unhealthy behavior
○ Personalized health programs
Patterns
Characteristics of medical data
● We focus on structured data (similar to ICU data)
○ Features are assumed to be numerical, categorical, ....
○ Think of EMR data (example from GP), see above.
● What are the characteristics?
○ Sparsity
○ Missing values
○ Lot of assumptions needed to process the data
2
,Machine Learning and Reasoning for Health - Slides Summary
Sparse data
● Imagine we consider diagnostic codes per patient
● Medical data is often highly sparse
○ Sparse data contains a lot of zeros/NAs
● Example from medical dataset of GP data
○ ICPC coding
Missing values
● Think about data on heart rate or lab results
○ Often not continuously measured
○ Often measured for a reason
○ Registration itself can be poor (and have different causes), see previous blue box
● Three types of missing data
○ Missing completely at random (MCAR). When data are MCAR, the fact that the
data are missing is independent of the observed and unobserved data. In other
words, no systematic differences exist between participants with missing data
and those with complete data.
○ Missing at random (MAR). When data are MAR, the fact that the data are missing
is systematically related to the observed but not the unobserved data.
○ Missing not at random (MNAR). When data are MNAR, the fact that the data are
missing is systematically related to the unobserved data, that is, the missingness
is related to events or factors which are not measured by the researcher.
● Can lead to biases
Assumptions needed
● To be able to apply AI successfully,
assumptions need to be made
● Can you think of examples of assumptions
we need to make to apply machine learning
to this data?
3
, Machine Learning and Reasoning for Health - Slides Summary
● Examples
○ Diagnostic code not provided means the patient did not have the disease
○ Diagnoses are entered when the disease starts
Machine learning
● Machine learning thrives when a lot of data is available
● Sparse data requires a lot more instances
● Lot of missing values result in more noise, potential biases in models
○ A Review of Challenges and Opportunities in Machine Learning for Health
■ Sources of missingness should be carefully examined before deploying a
learning algorithm.
■ Including missingness indicators provides the most information for making
predictions. However, learning models without an appropriate model of
missingness leads to issues such as inaccurate assessment of feature
importance and models that are brittle to changes in measurement
practices.
■ We note that data may also be missing because of differences in access,
practice, or recording that reflects societal biases
● When assumptions we make do not hold the learned models can be inaccurate
Injecting knowledge into machine learning
● Knowledge can help to overcome some of these issues
● We will focus on four approaches to inject knowledge:
○ Feature selection using medical knowledge
○ Feature abstraction using medical knowledge
○ Missing value handling
○ Machine learning techniques that embed knowledge (ExpertRuleFit)
● Let us start with some basic ingredients, namely ontologies like SNOMED-CT
Coding schemes
● International Classification of Primary Care (ICPC) Anatomical Therapeutic Chemical
(ATC) Classification system
4
Lecture 1: Introduction
AI disciplines
Machine learning
● Definition:
○ “Machine learning is to automatically identify patterns from data”
○ “A computer program is said to learn from experience E with respect to some
class of tasks T and performance P, if its performance at tasks in T improves with
E.” (Mitchell)
● Why did it become so popular?
○ More data
○ More computational power
○ Better algorithms
Knowledge represention/reasoning
● Representing information about the world in a form that a computer can use
● Representing knowledge about solving complex tasks
● Tasks like: diagnosing, planning, having a dialog with human...
● Several forms of knowledge representation
○ (logic, rules, semantic nets, ontologies etc.), and thus of reasoning.
● KRR: symbols & manipulation of symbols
● data => information => knowledge
● ...- - -... => SOS => emergency alert -> start rescue operation
AI for Health
● In health care there is of course a lot of data
● And we have an increase in devices that collect usable data
● Mobile health: over 100,000 mobile apps in the iTunes store alone
● What can AI contribute to health?
1
,Machine Learning and Reasoning for Health - Slides Summary
● How good is AI? → AI vs. Doctors
● Early detection/diagnosis
● Personalization of treatments
○ Provide advice to doctors on best treatments per patient
Prevention
● How can AI aid in prevention? Some examples:
○ Personalized coaching
○ Signaling unhealthy behavior
○ Personalized health programs
Patterns
Characteristics of medical data
● We focus on structured data (similar to ICU data)
○ Features are assumed to be numerical, categorical, ....
○ Think of EMR data (example from GP), see above.
● What are the characteristics?
○ Sparsity
○ Missing values
○ Lot of assumptions needed to process the data
2
,Machine Learning and Reasoning for Health - Slides Summary
Sparse data
● Imagine we consider diagnostic codes per patient
● Medical data is often highly sparse
○ Sparse data contains a lot of zeros/NAs
● Example from medical dataset of GP data
○ ICPC coding
Missing values
● Think about data on heart rate or lab results
○ Often not continuously measured
○ Often measured for a reason
○ Registration itself can be poor (and have different causes), see previous blue box
● Three types of missing data
○ Missing completely at random (MCAR). When data are MCAR, the fact that the
data are missing is independent of the observed and unobserved data. In other
words, no systematic differences exist between participants with missing data
and those with complete data.
○ Missing at random (MAR). When data are MAR, the fact that the data are missing
is systematically related to the observed but not the unobserved data.
○ Missing not at random (MNAR). When data are MNAR, the fact that the data are
missing is systematically related to the unobserved data, that is, the missingness
is related to events or factors which are not measured by the researcher.
● Can lead to biases
Assumptions needed
● To be able to apply AI successfully,
assumptions need to be made
● Can you think of examples of assumptions
we need to make to apply machine learning
to this data?
3
, Machine Learning and Reasoning for Health - Slides Summary
● Examples
○ Diagnostic code not provided means the patient did not have the disease
○ Diagnoses are entered when the disease starts
Machine learning
● Machine learning thrives when a lot of data is available
● Sparse data requires a lot more instances
● Lot of missing values result in more noise, potential biases in models
○ A Review of Challenges and Opportunities in Machine Learning for Health
■ Sources of missingness should be carefully examined before deploying a
learning algorithm.
■ Including missingness indicators provides the most information for making
predictions. However, learning models without an appropriate model of
missingness leads to issues such as inaccurate assessment of feature
importance and models that are brittle to changes in measurement
practices.
■ We note that data may also be missing because of differences in access,
practice, or recording that reflects societal biases
● When assumptions we make do not hold the learned models can be inaccurate
Injecting knowledge into machine learning
● Knowledge can help to overcome some of these issues
● We will focus on four approaches to inject knowledge:
○ Feature selection using medical knowledge
○ Feature abstraction using medical knowledge
○ Missing value handling
○ Machine learning techniques that embed knowledge (ExpertRuleFit)
● Let us start with some basic ingredients, namely ontologies like SNOMED-CT
Coding schemes
● International Classification of Primary Care (ICPC) Anatomical Therapeutic Chemical
(ATC) Classification system
4
