Validation of the canonical hemodynamic response function model used in fMRI studies

Introduction There is need to validate the fMRI analysis methods using large-scale datasets [1]. It is known that the hemodynamic response function (HRF) varies across brain regions and subjects [2]. Different HRF models are currently available in fMRI analysis packages [3]. The most popular HRF model is the canonical HRF: sum of two fixed gamma curves. We compare some of the implemented HRF models in terms of sensitivity-specificity efficiency. Aims We aim to identify an optimal HRF model for healthy subjects and to check if this model is flexible enough for unusual subject populations like impaired consciousness patients or the elderly. For such patient populations the hemodynamic response might differ and the canonical model might be too fixed to detect experimentally induced neural activation. Methods We compare HRF models in the most popular fMRI analysis softwares: AFNI, FSL and SPM, regarding: (1) positive rates (at least one significant voxel in the entire brain) (2) spatial distribution of the significant voxels and (3) numbers of significant voxels. For our analyses we employ three different baselines: (1) simulated data, for which we developed an R package: https://github.com/wiktorolszowy/HRF_simulation_for_fMRI_experiments, (2) resting state data with assumed dummy design and (3) task based data with assumed wrong design. Results The following table shows mean numbers of significant voxels for 70 impaired consciousness subjects performing a visual task. The visual stimulus was presented for 16s after 16s of rest. This design was repeated 10 times. The results refer to different assumed experimental designs and different HRF models. Conclusions The FIR model as implemented in FSL returned suspiciously high numbers of significant voxels which suggests that the degrees of freedom might be mis-specified. FIR behaved in a similar way when using simulated data and resting state data and for different subject populations (not shown in the above table). The effect of adding the first derivative to the gamma models was negligible. This suggests that the canonical HRF model is inflexible. Interestingly, significance was detected more often for assumed low experimental frequencies. This might suggest that the employed degrees of freedom do not model the correlations between periods when the stimulus was shown and the rest periods.