Reverse Inference in Neuroimaging

Reverse Inference in Neuroimaging

Introduction

• Functional neuroimaging is increasingly used to infer cognitive functions based on brain region activation.
• Reverse inference involves reasoning backward from brain activation to cognitive function engagement.

Inference in Neuroimaging

• Common reasoning: If cognitive process X is engaged, brain area Z is active; area Z activation suggests cognitive process X engagement.
• Reverse inference isn't deductively valid; it's affirming the consequent.
• Valid syllogism requires area Z to be active if and only if cognitive process X is engaged.

Bayesian Perspective

• Reverse inference can be restated in probabilistic terms using Bayes' theorem:
  • $P(COGX|ACTZ) = \frac{P(ACTZ|COGX)P(COGX)}{P(ACTZ)}$
  • Where:
  • $COGX$: Engagement of cognitive process X.
  • $ACTZ$: Activation in region Z.
• The belief in reverse inference depends on neural response selectivity and prior belief in cognitive process engagement $[P(COGX)]$.

Estimating Selectivity

• The strength of reverse inference relies on how selectively a brain region is activated by a cognitive process.
• Selectivity estimation can be done using neuroimaging databases like BrainMap.

Example: Broca's Area

• Examined reverse inference: activation in Broca's area implies language function engagement.
• Results:
  • Posterior probability depends on conditional probabilities and prior estimate of language processes being engaged.

Need for Cognitive Ontology

• Reverse inference identifies cognitive process engagement but requires experiments coded accordingly.
• Cognitive ontologies of existing databases are coarse compared to cognitive psychology theories.

Improving Reverse Inferences

• Increase confidence by:
  • Increasing response selectivity in the brain region.
  • Increasing the prior probability of the cognitive process.
• Experimental tasks can maximize the prior probability of a process being engaged.

Conclusions

• Exercise caution when using reverse inference, especially with low prior belief and selectivity.
• Neuroimaging databases can provide insights, but are limited by the coarseness of cognitive ontology.
• Reverse inference can suggest novel hypotheses for subsequent testing.