Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Preprocessing Matters: A Multiverse Analysis of EEG Decoding in Naturalistic Speech Listening
Poster Session C, Thursday, October 1, 10:45 am - 12:45 pm, Wangari Maathai
Yi Li1, Maya Inbar1, Laura Gwilliams2, Alexandra Woolgar1; 1University of Cambridge, 2Stanford University
Transforming sound into meaning involves computations spanning multiple levels of information. A powerful method for understanding the neural representation of these features is linear decoding during naturalistic listening, which uses machine learning to predict linguistic features from brain signals. To robustly decode these features, it is often necessary to preprocess the raw neuroimaging signals to improve the signal-to-noise ratio. However, it is unclear which preprocessing choices are optimal or necessary for speech decoding, since some steps may remove task-relevant neural signal along with noise. Here, we examined how seven electroencephalography (EEG) preprocessing choices affected the decoding strength of six different groups of linguistic features. In the ideal case, we would identify one preprocessing pipeline that optimised decoding strength across all features of interest. However, it was also possible that different features would benefit from different preprocessing regimes. Native English speakers (n=9) listened to The Little Prince audiobook three times, while their brain activity was recorded using portable EEG (twice) and MEG (once). We applied back-to-back ridge regression to quantify the decoding strength of six groups of features: prosodic, phonetic, sub-lexical, lexical, syntactic, and semantic. The back-to-back approach is well-suited to analysing naturalistic listening data, as it isolates feature-specific decoding from amongst the multiple collinear linguistic features. To examine the influence of preprocessing choices, we systematically varied seven steps to create a preprocessing multiverse. First, we generated nine band-pass filters by combining three high-pass filters with three low-pass filters. Second, we either applied or did not apply four steps (ICA, bad channel interpolation, linear detrending, and epoch rejection). Finally, we compared concatenating or point-by-point averaging data from the two EEG sessions to decoding based on a single EEG session. We report three main findings. First, some preprocessing choices had a similar effect on decoding across features. Lower low-pass cutoff and epoch averaging across two sessions improved decoding for most features, while bad channel interpolation had no significant effect on any feature. Second, some preprocessing steps had different effects on the decoding of different features. These feature-specific steps included epoch rejection, high-pass filtering, linear detrending, and ICA. For example, linear detrending improved decoding for most features, but reduced the strength of semantic decoding. This might suggest that detrending removed not only slow drift but also the slow neural signal that reflects meaningful semantic processing. This also meant that no single pipeline was optimal for all feature levels. For instance, syntactic decoding performed the best with quite constrained band-pass filtering, plus bad channel interpolation, epoch rejection, and linear detrending, whereas the optimal phonetic decoding used a higher high-pass filter and did not include epoch rejection. Finally, we found that preprocessing steps interacted with each other, rather than acting independently, indicating that preprocessing should be treated as a pipeline-level decision. These findings suggest that preprocessing choices substantially affect the decoding strength of linguistic features. While we make some general recommendations, systematic evaluation and transparent reporting of preprocessing pipelines will be crucial for interpretable E/MEG decoding in naturalistic speech research.
Topic Areas: Speech Perception, Methods