Attention (Predictive Networks)

Question 1

1. What are predictive network models and how are they constructed?

Answer 1

= Models that predict a behavioural measure from pattern of functional brain connectivity

Connectome-based predictive modelling approach:

1. Compute each participant’s unique connectivity pattern, by correlating each node in a whole-brain atlas with each other
2. Identify behaviourally relevant connections, by correlating every connection with the behaviour across subjects
3. Build models relating relevant connections to behaviour, keeping the strong connections in either direction from step 2
4. Apply model to novel individual’s connectivity pattern to predict behaviour (leave-one-out cross-validation). Compare predicted and observed behaviour to assess predictive power

Question 2

2. What are the pros and cons of predictive network models comparing to standard approach?

Answer 2

Pros
• Directly linking brain variables to behaviour ones  give us insight into how the brain gives rise to attention
• Provide empirical evidence that attention is actually a network property
• Allow individualised exploration
• Generalisable  allow predictions
o Potentially applicable in clinical or translational settings

Cons
• Resulting models are difficult to interpret
o What does e.g. an attentional profile mean?
• Problems related to resting state data (what influence do confounds in fMRI have)
• What about connectivity over time? Stable within people?
• Legal aspects? (Responsibility etc.)

Question 3

3. What are clinical applications of predictive network models?

Answer 3

Provide a new way of measuring cognitive functions (not only attention)
• Instead of actively doing a task, people can just lie in the scanner – less confounding variables
• Can be applied retroactively to existing data
• No limit to number of tests, because they can all be performed on the same data

Support treatment
• Identify individuals at risk
• Match people to the most appropriate training or treatment
• Track changes over time / treatment course

Question 4

4. How do PNM relate to known attentional and resting-state networks?

Answer 4

Sustained attention CPM does not neatly overlap with existing neuroanatomical models of attention
• Is more widely distributed
• Might be based on involving certain neurotransmitter systems
o High attention network overlaps with methylphenidate network, low-attention with unmedicated network
o  CPM may be related to dopamine and noradrenalin expression

Connections of canonical networks in the high and low attention networks
• Researchers selectively removed connections known from the canonical networks from the high and low attention networks
• Generally robust to this “lesioning”, but especially removing connections from Frontoparietal and default mode network lead to insignificant trends  these networks seem to be especially important

Both approaches to network unravelling seem to be able to predict behavioural and clinical outcomes

Question 5

5. How should PNM be validated?

Answer 5

Internal validation
• Leave-one-subject-out cross validation

External validation
• Test the model on a completely independent group of individuals

Ecological validation
• Test if model generalises to a real-world measure of attention
• E.g. test attention model on group of children, of which some were diagnosed with ADHD