AI in RT Flashcards
machine learning + different types
Statistical methods of learning from data
- Supervised learning
- Unsupervised Learning
- Semi-supervised learning
- Reinforcement learning
- Transfer learning
supervised learning - 2 categories
- classification
- predicts a class outcome –> dosimetry goal (met or not met)
- regression
- predicts a numerical outcome (OAR dose volume)
How do we ensure safety of our patients is maintained?
- Relies on an understanding of how AI systems are designed and applied
- Education of undergraduate and qualified staff on AI basics
- Ability to understand the assumptions and limitations of an AI application to
ensure accountability in decision making - Communication with patients about how these tools are used
In practice, MRPs need to…
- have an understanding of the technology, data science, including privacy and security
- Should use and deploy these systems, maintain and overview them, direct their application in their practice
- Should have control of AI supported systems to ensure that patient safety is not compromised.
Both undergraduate & qualified medical radiation practitioners are expected to engage in:
- Understanding the language of AI
- Understanding the principles of AI, the development of data algorithms and the evaluation and validation
of AI supported healthcare - Awareness of the benefits and risks, including supervised and unsupervised machine learning,
- Informing and explaining to patients about the role of AI in their healthcare
- Understanding and applying health informatics to ensure safe AI-supported healthcare
- Data curation, governance and stewardship that assures patient safety, privacy and data protection
- Integrating research driven, evidence-informed approaches to safe AI-supported healthcare, and
- Partnering and leading AI-supported point-of-care image analysis and diagnosis, treatment planning and
adaptive technologies
causes for inappropriate quality in data sampling
- selection bias
- class imbalance
- insufficient quantity
- prejudice bias
selection bias
collecting data that is not representative of the target population
class imbalance
Underrepresentation of a class
in the outcome variable
insufficient quantity
Low sample size insufficient to train model relative to the complexity of the outcome
prejudice bias
training data includes (human) biases containing implicit racial, gender or ideological prejudices
causes for inappropriate quality in data management
- noise
- uninformative or less informative
- exclusion bias
- measurement bias
- survey bias
- observer bias (confirmation bias)
data quality issue - noise
Particularly related to medical
images where artefacts or blurry
images may reduce image quality
uninformative or less informative
Records do not convey all the
necessary information to solve
the problem at hand
exclusion bias
Deleting valuable data that was
thought to be unimportant
measurement bias
systematically favoring a
particular result when observing
or measuring variables in data
survey bias
Substantial missing, incomplete,
and inconsistent responses to
surveys, questionnaires, or
interviews used to collect data
observer bias
Favoring information that does
not contradict the researcher’s
desire or previous beliefs
causes for inappropriate quality in data annotation
- recall bias
- annotation bias
- lateral variability (interobserver variability)
- longitudinal variability (clinical drift)
recall bias
Labelling similar types of data
inconsistently
annotation bias
Inaccuracies induced by
human errors when labelling
lateral variability (interobserver variability)
Difference in practice
between physicians or
institutions
longitudinal variability (clinical drift)
Difference in practice over
time due to change in
techniques or guidelines
missing data + different types
Missing data can be a common issue in healthcare research, particularly when relying on electronic health records
Missing data can be classified as:
- Missing completely at random (MCAR)
- Missing at random (MAR)
- Missing not at random (MNAR)
how to fix missing data
- One option is to remove the rows with missing data
- Another option is to use multiple imputation, a statistical method for estimating the missing values
- Data augmentation is an additional method for increasing the dataset size when limited data is available
Aleatoric uncertainty
- Aleator is the Latin word for someone who rolls the dice
- this uncertainty is based on randomness in the data
- it cannot be reduced by collecting more data
- It can only be measured and quantified
- It is seen when an experiment is repeated multiple times and there is variation in the outcome due to
randomness
Epistemic uncertainty
- From the Greek word relating to knowledge
- Often described as model uncertainty
- Can be reduced by gaining greater “knowledge” of the question being asked of the model.
- Gaining greater “knowledge” may be in the form of more data, or more input variables (higher dimensionality) to give the model a greater ability to fit to the data
AI ethics principle
- human, societal, and environmental wellbeing
- human-centred values
- fairness
- privacy protection and security
- reliability and safety
- transparency and explainability
- contestability
- accountability
supervised vs unsupervised learning
In supervised learning, the algorithm learns from labeled data, meaning the input data is paired with corresponding correct outputs, allowing the model to learn patterns and make predictions.
Unsupervised learning, on the other hand, deals with unlabelled data, aiming to discover underlying structures and relationships within the data without explicit guidance
What QA should be completed in the implementation stage of AI?
- identify workflow position
- identify test data set
- end to end testing of AI tool
- set clinical limits for use
- risk analysis
What QA should be completed in the case-specific stage of AI?
- manual checks of outcomes
- recording instances of manual adjustment
- send feedback to model developers
What routine QA should be completed for AI?
- use standard datasets from implementation to check any model updates
- review logbook for trends
- review patient cohort for shifts in background
- send feedback to model developers
What is involved in an AI contingency plan?
Set a threshold for action, identify the logistics of changing to a manual process, determine what the AI tool is failing because of (1 pt, group, or all)
What risks need to be analysed before the use of AI?
- how much autonomy is the system given
- do we trust the results
- what are the potential consequences if something goes wrong
- what is the downstream impact
- what areas could the error impact
What is distributed learning?
Data collected across multiple hospitals where data is sent to a central terminal
Why is model interpretation important?
1- ethical considerations
2 - if something goes wrong understand how AI works
3- clinical reasoning
model agonistic interpretation
can be applied to any models and are more flexible than model-specific interpretations
global interpretations
provide information on what variables were most influential from the training dataset
local interpretation
provide information on what variables from an individual’s data were most influential on model output ]
