Reduced parameter space fitting: shift and tilt without ECAP profile

In a study by Smoorenburg et al. (2002) the T and C and ECAP data was analysed using a principal component analysis. This analysis showed that more than 96% of the total variance in the data could be described by 2 or 3 components: one relating to the overall level of the profile (shift), the second related to the slope or tilt of the profile and the last one relating to the curvature [1] . It must be noted that although we use the terms shift and tilt, these are not linear straight shifts or tilts: they are profiles that come from the data of the group. For instance, a tilt will result in less tilt at the complete extremes of the array than what would be expected from a straight tilt.
 
The next step was not to look at the correlation between ECAP and T and C levels but to look at the correlation between shift and tilt in the ECAP and T and C level profile. As expected the correlation between ECAP and T and C shift was very low (0.64/0/39), however, the correlation between the T-level tilt and the ECAP tilt was significant and in the 0.82 range. This means that the ECAP could ‘predict’ the shape (albeit not the level) of the T-profile, which supports the ECAP based methods described here . However, a follow-up analysis on a larger dataset (Cafarelli Dees et al., 2005) did not show a significant correlation.
 
From these results Smoorenburg proposed to use a method based on the average profile across clients (assuming the ECAP profile does not give any additional information) and using the shift/tilt:
  1. T-level is set to the average profile.
  2. C-level set just above T.
  3. Whole profile (T and C) is dropped down until subthreshold.
  4. In live mode the complete profile is increased (shifted up) until just audible.
  5. T level is fixed; C level is increased (applying shift) until sound is comfortable.
  6. To adjust for optimal sound quality a tilt to the C and or T levels can be applied.
Effectively it means we are describing the map as a linear combination of the two principal components.
There is no scientific approach to setting the tilt values yet, but anecdotal results show that a change in the tilt is clearly connected to a sense of ‘timbre’. Alternatively the recently introduced bass and treble components can be used.
Note that if better NRT becomes available and we DO find a significant predicting value in T-NRTs, the above method can be used also with the T-NRT profile instead of the average profile.
 
Advantage: Fast, no need for ECAP. Takes temporal integration into account.
Disadvantage: Effect of very low T-levels may be suboptimal speech scores at very low levels, although data does not support this (Smoorenburg et al., 2002). May not be accurate enough for some recipients.
 
See the related item on shift and tilt fitting based on ECAP data
Back to recent developments in fitting methods.
Back to Cochlear Implant Fitting.


[1] Since the concepts of curvature and shift may not have a clear relation with auditory perception ratings such as sharpness, Smoorenburg recently proposed to use a slightly different decomposition based on volume, bass, and treble (Smoorenburg 2007).