Identifying Explainable and Generalizable Features for MEG Decoding

Sensory neural coding, the brains process of transforming inputs into informative patterns of neural activity, generates complex and multiplexed neural codes which are hard to interpret. Although decoding methods have facilitated the interpretation of these codes, the specific features of neural activity that generalize across individuals, along with their precise timing, largely remain elusive. To address this gap, we investigated the potential of interpretable time-series features in magnetoencephalography (MEG) for decoding visual stimulus attributes (spatial frequency and orientation) and their generalizability across individuals. We extracted a comprehensive set of highly informative features from 18 subjects engaged in a visual task and performed a time-resolved decoding analysis. Our findings revealed that particular features, especially those capturing rapid changes in neural activity within the first 200 milliseconds of stimulus presentation, yielded high decoding accuracy. Notably, these early, transient features exhibited robust cross-subject generalizability, suggesting a shared neural coding mechanism during the initial processing of visual inputs. Furthermore, these features outperformed previously successful electroencephalography (EEG) wavelet features when applied to MEG data. While within-subject decoding demonstrated sustained above-chance performance, cross-subject generalization diminished after the initial 200 milliseconds, indicating more individualized processing at later stages. Our results underscore the importance of systematic, data-driven evaluation of neural signals for elucidating neural codes and for developing more transparent and generalizable Brain-Computer Interface (BCI) systems that capitalize on these reliable neural signatures.