Problems with the Canonical HRF

Let's dig a little deeper into the problems with the standard GLM analysis. First, the standard GLM analysis looks for a single Canonical HRF. This is convenient because it is computationally easier to look for one predefined pattern, and when fMRI research started in the 1990s, this was probably all you could reasonably expect your computer to do.

But, if you convolve one pattern with all your tasks, you are going to be sorely disappointed and miss a lot, because the brain is more complicated than that!

Let's look at the literature on how the canonical HRF fails us.

Bonakdarpour et al, 2007 demonstrated that the HRF varied between healthy controls and stroke patients. In fact, they also demonstrated that HRF varied between brain regions, even for the same participant!
Miezen et al, 2000 reported that the amplitude and timing of the HRF in one region failed to predict the HRF in another region, even within-subject.
Handwerker (2004) demonstrated that using temporal derivatives AND subject-specific HRFs helped, but did NOT solve the problem of HRF variability: Even a homogenous group of young healthy male controls differed in both the timing and the amplitude of their response functions. Such variation could result in the GLM missing activations!
If Handwerker's ideal data (homogeneous group, simple button press to visual stimulus) shows even a few false negatives, then misestimates of HRFs will cause more false negatives and worse magnitude estimates in more complex data.
Standard GLM analysis fails because it tries to match a single pattern everywhere in the brain. That oversimplification results in loss of sensitivity.
The best we can do with a model-driven mass univariate approach (i.e., the standard GLM approach) is this: If I know how Joe's motor cortex reacts during a button press, then that BUTTON PRESS pattern should be pretty reliable IN JOE'S MOTOR CORTEX if I scan JOE again.

Watch this 8.5-minute video clip

fMRI Analysis 1: GLM vs ICA